cs.CV [Total: 252]
cs.CL [Total: 88]
cs.AI [Total: 22]
cs.CR [Total: 2]
cs.IR [Total: 2]
cs.LG [Total: 29]
cs.NI [Total: 1]
cs.RO [Total: 23]
cs.SD [Total: 1]
cs.SE [Total: 3]
eess.AS [Total: 1]
eess.IV [Total: 2]

cs.CV [Back]

[1] How Vision Transformers Represent Position: A Mechanistic Comparison of Absolute and Rotary Positional Encodings cs.CV | cs.LGPDF

Mahmoud Mannes, Nikhil Maturi, Parva Mehta, Jiwon Jeong, Aravind Singaram Kannappan

TL;DR: 机制分析显示 APE 把行列位置存入显式特征，而 RoPE 在注意力中编码并可重建轴向位置；对 APE 位置特征置换会使分类 top-1 准确率下降 24–31 个百分点。

Details

Motivation: ViT 的位置嵌入帮助区分图像 patch 的空间布局，但不同位置编码如何塑造模型内部表征仍知之甚少。研究比较绝对位置嵌入（APE）与旋转位置嵌入（RoPE）的因果机制。

Result: 作者发现 APE 和 RoPE 采用不同机制：APE 将位置存于显式特征，RoPE 则在注意力计算中构造位置关系，并能在信息被移除时重新构建。APE 模型中，稀疏自编码器恢复出具有因果作用的行/列特征；对 RoPE 的干预显示不同旋转分量分别编码行和列。分别禁用行/列旋转可独立移除可解码的位置，再恢复旋转后部分重建，部分由轴向专化注意力头完成。两类机制都影响分类：跨 patch 置换 APE 位置特征使 top-1 准确率下降 24–31 个百分点；消融 RoPE 主要重建头的旋转，对准确率影响大于任意其他头。

Insight: 位置编码并非只加上坐标标签：它决定空间结构在特征通道与注意力运算中的组织方式，也影响模型是否能恢复被移除的位置线索。因果影响数字来自文中分类任务，摘要未列模型和数据集细节。

Abstract: Positional embeddings encode the spatial position of each image patch in a Vision Transformer (ViT), allowing the model to distinguish different spatial arrangements of visual content. However, we know little about how the choice of embedding shapes a vision transformer’s internal representation. In this work, we show that Absolute Positional Embeddings (APE) and Rotary Positional Embeddings (RoPE) encode spatial position through different internal mechanisms: APE stores position in explicit features, while RoPE builds it inside attention and rebuilds it when it is withheld. Specifically, in APE models, sparse autoencoders recover causal row- and column-specific features, whereas interventions on RoPE reveal distinct rotary components that separately encode row and column information. Additionally, causal interventions selectively disabling RoPE’s row or column rotations show that decodable row and column information can be independently removed and, once rotations are restored, partly reconstructed, in part by axis-specialized attention heads. Both mechanisms are causally important for classification: permuting APE’s positional features across patches reduces top-1 accuracy by 24–31 points, and ablating the rotations of RoPE’s primary rebuilding head reduces it more than ablating those of any other head. Our findings show that the choice of positional encoding fundamentally changes how spatial information is organized and processed inside Vision Transformers. More broadly, our work demonstrates how mechanistic analysis can reveal how models construct, reconstruct, and maintain structured representations, providing a basis for targeted interventions in vision models.

[2] Scenes as an Attack Surface: Uncovering Scene-Grounded Safety Failures in MLLMs cs.CVPDF

Tongxi Wu, Yuyang Cheng, Jiacheng Xian, Hu Huang, Zhihao Wu

TL;DR: SceneSafetyBench 揭示任务相关场景语义本身可使 MLLM 从拒绝转为配合，即使图像未加入新攻击指令；提出的 SGA 在九个模型上平均 ASR 为 86.17%。

Details

Motivation: 已有多模态越狱研究多依赖精心构造、直接携带有害意图的图像。本文考察另一类脆弱性：任务意图不变、视觉输入也不新增攻击指令时，场景语义能否单独改变模型安全行为。

Result: 作者将其定义为场景 grounding 的视觉安全漏洞，并构建 SceneSafetyBench，覆盖八类风险、九种白盒与黑盒 MLLM。分析发现，漏洞与任务相关视觉证据如何影响下游安全表征有关，单纯视觉 grounding 不足以解释安全状态变化。基于此提出 Scene-Grounded Attacks（SGA），保持原始任务不变，构造经验证的场景 grounding 查询并组合各自响应。九个目标模型上，平均攻击成功率（ASR）为 86.17%，比最强基线高 8.55 个百分点，在九个模型中的八个达到最高 ASR。

Insight: 更强的视觉依据可能与稳健安全对齐产生张力：普通场景内容而非显式攻击文本也可能改变拒绝行为。摘要报告的是安全评测中的攻击成功率，不包含防御策略或具体操作细节，结论限定于八类风险和九个模型。

Abstract: Existing jailbreak attacks have exposed serious safety weaknesses in multimodal large language models (MLLMs). Most prior attacks rely on specially constructed visual inputs to carry or support harmful intent. In this paper, we uncover a different vulnerability: task-relevant scene semantics alone can shift an MLLM from refusal to compliance, even when the visual input introduces no additional attack instructions and the original task intent remains unchanged. We formalize this failure as a scene-grounded visual safety vulnerability and introduce SceneSafetyBench to investigate it across eight risk categories and nine white- and black-box MLLMs. Our analysis reveals that this failure is associated with how task-relevant visual evidence influences downstream safety representations, suggesting that visual grounding alone cannot fully explain the observed safety shift. Building on this finding, we propose Scene-Grounded Attacks (SGA), which preserve the original task while constructing validated scene-grounded queries and combining their independent responses. Across nine target MLLMs, SGA achieves an average attack success rate (ASR) of 86.17%, outperforming the strongest baseline by 8.55 percentage points and achieving the highest ASR on eight models. Our results identify scene semantics as an overlooked attack surface in MLLMs and highlight a growing tension between stronger visual grounding and robust multimodal safety alignment.

[3] Human Pose Estimation by Probabilistic Mixtures of Gaussian and Uniform Distributions cs.CV | cs.LGPDF

Santiago Reina Alguacil, Ezequiel López Rubio

TL;DR: 该免训练后处理把人体姿态热图拟合为高斯混合加空间均匀分布，支持连续不确定性估计和概率级 TTA；在严重遮挡 OOD 检测中几何协方差优于峰值置信度启发式。

Details

Motivation: 现代人体姿态估计仍从离散热图确定性抽取坐标，在密集遮挡或分布外歧义下容易失败。离散峰值难表达多个可能位置及不确定性。

Result: 方法在不改架构、不重训的情况下拦截原始热图，将其建模为高斯混合模型并加入空间无关均匀分布，形成连续概率表示。拟合 GMM 的空间期望达到领先亚像素解码器相当的精度，无需启发式后处理。测试时增强对概率分布而非离散激活进行聚合，以减轻多尺度热图污染；多峰 GMM 还能与马尔可夫随机场等几何正则器结合。分析也揭示透视缩短和视觉退化下二维运动学先验存在脆弱性。严重遮挡 OOD 异常检测中，校准几何协方差体积优于常用峰值强度启发式；域内校准方面，Adaptive Softmax 融合几何度量后，多个基准和退化等级下 AUSE 降低或持平。该连续参数化还可扩展到多视角 3D 架构，绕开 O(N³) 体素内存瓶颈。

Insight: 热图不确定性应保留为连续且可多峰的概率密度，这既可改善增强聚合，也可提供比单峰置信更有意义的异常信号。摘要未报告各任务的具体精度收益，且2D几何先验会受视角和图像质量影响。

Abstract: Despite advancements in sub-pixel decoding, modern Human Pose Estimation (HPE) architectures remain constrained by deterministic coordinate extraction from discrete heatmaps, which systematically fails under dense occlusions and Out-of-Distribution (OoD) ambiguities. To address these limitations, we propose a post-hoc continuous probabilistic framework that intercepts raw heatmaps and models them as a Gaussian Mixture Model (GMM) augmented with a spatially agnostic uniform distribution, requiring no architectural modifications or retraining. This continuous representation allows us to comprehensively explore the distinct contributions of density modeling to HPE. We show that the spatial expectation of our fitted GMM achieves precision parity with leading sub-pixel decoders without heuristic post-processing. Leveraging this continuous space, our test-time augmentation aggregates probability distributions rather than discrete activations, mitigating multi-scale heatmap poisoning. Furthermore, because our multimodal GMM parameterization isolates competing spatial hypotheses, it naturally interfaces with geometric regularizers such as Markov Random Fields. However, our analysis exposes the inherent vulnerabilities of 2D kinematic priors under perspective foreshortening and visual degradation. Crucially, this density formulation yields significant improvements in uncertainty quantification: our calibrated geometric covariance volume outperforms standard peak-intensity heuristics in OoD anomaly detection under severe occlusion. For in-distribution calibration, dynamically fusing our geometric metric via Adaptive Softmax reduces or matches baseline Area Under the Sparsification Error (AUSE) across benchmarks and degradation levels. Finally, we establish that this continuous parameterization is extensible to multi-view 3D architectures, bypassing the $\mathcal{O}(N^3)$ voxel memory bottleneck.

[4] TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback cs.CVPDF

Jingyu Liu, Hao Jiang, Yikang Ding, Huafeng Shi, Xu Yang

TL;DR: TrajCFG 根据采样历史中各 token 的引导速度方向变化，自适应分配有界 CFG 强度，无需额外模型评估；三种文生视频主干上 VBench Total 比最强基线高 0.30–0.75 分。

Details

Motivation: 标准 CFG 对所有潜 token 使用同一引导尺度，但不同 token 的速度预测随采样演化并不相同。固定强度可能对已稳定方向过度引导、对仍需修正的位置引导不足。

Result: Trajectory-Adaptive CFG（TrajCFG）是免训练流匹配方法，根据历史引导速度方向变化为每个 token 分配尺度：方向变化较大的 token 获得更强引导，较一致的方向获较弱引导。算法将相邻步反馈与区间归一化的长跨度反馈结合，复用标准 CFG 预测而不增加模型评估次数。受控诊断和消融分析方向反馈、token 级分配及历史跨度融合。在文生视频、图生视频、文生音视频、文生图和图生 3D 多种生成任务与主干上，方法较基础采样器有改善。文生视频三种主干的 VBench Total 比各自最强评测基线高 0.30–0.75 分。

Insight: 利用去噪轨迹中方向稳定性作为局部引导强度信号，使统一 CFG 改为逐 token 分配，同时不需额外前向评估。0.30–0.75 是 VBench Total 分数差，跨任务泛化结论以摘要报告为准。

Abstract: Classifier-free guidance (CFG) typically applies a shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. We introduce Trajectory-Adaptive Classifier-Free Guidance (TrajCFG), a training-free method for flow-matching models that adapts token-wise guidance using historical changes in guided velocity direction. TrajCFG maps this feedback to bounded scales, assigning stronger guidance to larger directional changes and weaker guidance to more consistent directions. It combines adjacent-step and interval-normalized longer-span feedback, reusing standard CFG predictions without additional model evaluations. Controlled diagnostics and ablations examine the roles of directional feedback, token-wise allocation, and history-span fusion. Experiments on text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation show improvements over base samplers across multiple backbones. For text-to-video generation, TrajCFG improves VBench Total scores by 0.30-0.75 points over the strongest evaluated baseline on each of three backbones. These results support historical directional feedback as a broadly applicable signal for fine-grained guidance allocation. The project page is available at https://jyliu-98.github.io/trajcfg-project-page/.

[5] Video-Arena: A Scalable Agentic Video Benchmark for Spatiotemporal Capability cs.CVPDF

Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu

TL;DR: Video-Arena 用对战游戏记录自动生成视频与对齐标签，评测模型感知关键事件、解释状态更新并追踪后续状态；结果显示工具辅助能改善事件恢复，但长期状态理解仍弱。

Details

Motivation: 动态世界理解要求模型跟踪不断变化的状态，但真实视频中的真实状态往往部分不可观测，难以获取可靠标注。已有视频基准也较难系统测量模型对状态演变的一致理解。

Result: Video-Arena 把竞技游戏转化为可扩展、可验证的时空世界状态评测。任务分为关键事件感知、状态更新解释和结果状态追踪；管线从游戏记录大规模生成视频及对齐参考标签。为降低成本并保证复现，作者整理覆盖 9 款游戏的 180 题黄金测试集，并提供原生视觉理解与工具辅助智能体分析两种协议。摘要中两协议最佳 F1 数值缺失，无法补全。定性结果显示，原生模型难以可靠感知事件；工具辅助可改善事件恢复，但解释状态更新和跨时间一致追踪仍不稳定。

Insight: 将状态理解拆为事件、状态变化解释和后续跟踪，有助于区分“看到发生了什么”和“持续维护状态”的能力。由于摘要的 F1 数字为空，只能记录上述定性结论和评测规模。

Abstract: Understanding and tracking a dynamic world are fundamental to visual intelligence. Existing video benchmarks provide limited insight into whether models can maintain a coherent understanding of evolving world state. Building such a benchmark from real-world videos is challenging because the underlying world state is partially observable, making annotations hard to obtain. To close this gap, we introduce Video-Arena, a benchmark that turns competitive games into scalable, verifiable tests of how well models understand evolving world states from videos. Specifically, inspired by how humans understand dynamic events, we evaluate spatiotemporal world state understanding and tracking through three task categories: perceiving critical events, interpreting state updates, and tracking the resulting states. Our pipeline can generate videos and aligned reference labels from game records at scale. To keep evaluation affordable and reproducible, we curate a golden set of 180 tasks across 9 games and run evaluations under two protocols: native visual understanding and tool-assisted agentic analysis. The best-performing systems achieve F1 scores of and under the two protocols, respectively. The result highlights that native models struggle to perceive events reliably, while tool-assisted systems improve event recovery but still struggle to interpret state updates and track the resulting states consistently over time. Our project page is in https://videoarena-bench.github.io

[6] CAPTOR: a vision-language foundation model for Chest-Abdomen-Pelvis Oncologic CT cs.CV | cs.AIPDF

Séléna Jean-Mactoux, Sophie Loizillon, Ghina Jardali, Amandine Crombé, Nathalie Lassau

TL;DR: CAPTOR 是用 25,512 对训练、10,794 对测试的肿瘤富集胸腹盆 CT—报告数据训练的视觉语言模型；在目标肿瘤 CT 分类上优于基线，并在四项预后基准中的三项排名第一。

Details

Motivation: 胸、腹、盆部肿瘤 CT 需要综合多个解剖区域和单/多序列影像进行检测、预后与检索。通用 CT 基础模型未必针对肿瘤临床场景进行预训练。

Result: CAPTOR 是面向肿瘤放射学的对比式视觉语言基础模型，以北美、欧洲和亚洲多中心回顾性数据训练，主要为肿瘤病灶检查，支持单序列和多序列采集。训练/测试 CT—报告配对分别为 25,512 和 10,794；TumorImagingBench 各预后数据集含 134–421 项检查。与先进 CT 基础模型比较，胸腹盆子集疾病分类 AUC 为 0.767（95% CI 0.734–0.793），下一名 0.695（0.645–0.736），p<0.001；肿瘤子集 AUC 0.784（0.759–0.805），下一名 0.734（0.703–0.767），p<0.001。四项肿瘤预后基准中三项排名第一，在胸腹盆及肿瘤子集图像检索也表现更优。研究采用 bootstrap 置信区间和置换检验。

Insight: 结果表明与临床目标对齐、肿瘤富集的胸腹盆 CT 预训练可优于相关通用基线；AUC 提升报告于疾病分类子集，预后结果是四项中的三项第一。摘要未给出检索指标或外部临床前瞻性验证。

Abstract: Purpose : To develop and evaluate CAPTOR, a deep-learning vision-language foundation model for oncologic chest-abdomen-pelvis (CAP) CT interpretation. Materials & Methods : CAPTOR (Chest-Abdomen-Pelvis foundaTion model for Oncologic Radiology) is a contrastive vision-language model, trained using a pretraining strategy for oncologic CAP CT (in single- or multiple-series acquisitions), on a multicenter retrospective dataset from North America, Europe and Asia, predominantly from tumor lesion examinations. CAPTOR was compared with state-of-the-art CT foundation models on three downstream tasks: disease detection, survival prediction using TumorImagingBench and image retrieval. Classification and retrieval were both evaluated on held-out internal and public test sets, stratified by anatomical coverage and indication. Bootstrap resampling was used to estimate confidence intervals, and pairwise comparisons were performed using permutation tests. Results : CAPTOR was trained and tested on 25,512 and 10,794 CT-report pairs respectively. TumorImagingBench datasets comprised between 134 and 421 studies each. CAPTOR outperformed all baseline models, on the classification task, on the CAP subset (AUC 0.767, 95%CI: 0.734-0.793, vs. 0.695, 95%CI: 0.645-0.736 for the next-best, p<0.001) and oncologic subset (AUC 0.784, 95%CI: 0.759-0.805, vs. 0.734, 95%CI: 0.703-0.767 p<0.001), ranked first on three of the four oncologic prognostic benchmarks, and achieved superior retrieval performance on CAP and oncologic subsets. Conclusion : A foundation model trained on oncology-enriched CAP CT data outperformed relevant state-of-the-art models in its target clinical context, underscoring the importance of clinically aligned pretraining. We release our preprocessing, training and inference code : https://github.com/Guerbet-AI/CAPTOR and model weights : https://huggingface.co/Guerbet-AI/CAPTOR

[7] Token-Budgeted Candidate Retrieval for 3D Scene-Grounded VLMs: Category Names vs. Pairwise Geometry cs.CV | cs.LGPDF

Michał Zakrzewski, Mariia Shpir, Taras Rumezhak

TL;DR: 在 ScanRefer、Nr3D、Sr3D 固定 token 预算检索对象时，类别名称匹配在各报告 K 值上优于成对几何；在 3DGraphLLM 同预算下，MiniLM 达 0.67，优于几何检索的 0.56。

Details

Motivation: 3D 场景 grounding VLM 的上下文长度有限，需要决定优先提供哪些物体和关系。手工成对几何评分是否比直接匹配对象类别名称更有效，尚不明确。

Result: 研究以轴对齐框和 token 比例评估固定预算候选检索，比较免训练方法、两个小型文本编码器和确定性几何评分 QCRR。数据覆盖 ScanRefer gt_val 及 Nr3D、Sr3D 的自建 20% scan-ID 划分，共 35,504 个查询（不是官方 ReferIt3D val）。在所有报告 K 值上，类别名称匹配优于成对几何，召回较高时差距最大；解析式成对线索又优于单对象线索和与查询无关的 kNN 图，但 MiniLM/BGE-small 在这些切分上分数仍更高。ScanRefer R@10 中，加法式名词匹配为 0.864，QCRR 为 0.626；MiniLM/BGE-small 在 ScanRefer 为 0.911–0.913，在 Nr3D 为 0.941–0.947。修正后的 3DGraphLLM harness（完整图 native Acc@0.25=0.67，ScanRefer n=300）下，同预算 Random 0.44、QCRR 0.56、MiniLM 0.67、oracle 0.70。

Insight: 这组实验支持在固定对象预算下优先按语义类别检索，而非依赖手工几何关系；但数据切分和 harness 有明确限定，尤其 Nr3D/Sr3D 使用自建 20% scan-ID 划分。3DGraphLLM 的下游准确率比较使用 300 个 ScanRefer 查询。

Abstract: Under a fixed context budget, which objects and relations should a 3D scene-grounded VLM see first? On ScanNet indoor referring benchmarks (Dai et al., 2017), matching the query to object category names outperforms pairwise geometry scoring at every reported K. The gaps are largest at high recall. Analytic pairwise cues beat per-object cue scoring and query-agnostic kNN graphs, including on paraphrases. MiniLM and BGE-small still score higher on those splits. We evaluate token-budgeted candidate retrieval with axis-aligned boxes, a token-ratio metric, training-free baselines, two small encoders that match q to each object’s category name, and QCRR (Query-Conditioned Relation Retrieval), a deterministic pairwise-geometry scorer. Across ScanRefer gt_val plus a custom 20% scan-id split of Nr3D and Sr3D (35,504 queries; not official ReferIt3D val), additive noun matching reaches R@10 0.864 on ScanRefer against QCRR’s 0.626; MiniLM and BGE-small reach 0.911 to 0.913 on ScanRefer and 0.941 to 0.947 on Nr3D. A corrected 3DGraphLLM harness (native Acc@0.25 0.67 on the full graph, n=300 ScanRefer queries) shows that context selection changes downstream accuracy under a matched object budget (Random 0.44, QCRR 0.56, MiniLM 0.67, oracle 0.70). Semantic retrieval beats hand-designed geometry on that harness. We will release the protocol, retriever implementations, and per-query outputs so new selectors can be swapped in without rewriting the analysis.

[8] Latent World Action Model for GUI Agents cs.CV | cs.AI | cs.HCPDF

Seungjun Yu, Hojun Choi, Seojeong Park, Jaeyo Shin, Hyunjung Shim

TL;DR: 该 GUI 世界模型不重建整屏像素，而在多模态潜空间预测动作导致的未来变化，并共享表示给动作选择和转移预测；离线/在线设置成功率相对像素模型最多提高 31.9%/40.6%，推理延迟最多降低 98.9%。

Details

Motivation: 像素级 GUI 世界模型需预测完整未来屏幕，计算昂贵且重建了很多与动作选择无关的外观细节；目标屏幕生成还可能产生 identity shortcut，只保留当前状态而不捕捉动作造成的变化。

Result: 方法将预测目标从整屏重建转为潜空间中的动作诱发未来信息，以大型多模态模型作为潜在世界模型，并用共享表征的两个专用头完成动作选择和动作条件转移预测。动作头直接从潜表示生成动作，转移头捕捉动作造成的未来变化；action-aware 对比目标鼓励转移表示区分不同动作引起的状态变化，以抑制 identity shortcut。离线和在线 GUI 实验中，相对像素空间世界模型，成功率最高分别提升 31.9% 和 40.6%，推理延迟最多降低 98.9%。

Insight: 只预测与决策有关的动作诱发变化，避免生成完整屏幕造成的无关重建负担；对比目标专门防止潜空间预测退化为复制当前状态。性能数字为最高相对增益，摘要未列各任务的绝对成功率和延迟。

Abstract: Pixel-space world models have recently emerged as a promising approach for GUI agents, predicting future screens prior to execution and subsequently re-perceiving them for action selection. However, this two-stage paradigm faces two key challenges in GUI environments: (L1) the high computational cost of pixel reconstruction and (L2) the need to reconstruct large and diverse appearance-level details beyond what is required for action selection. In practice, this screen-generation objective can induce an identity shortcut problem, where future-state prediction largely preserves the current state rather than capturing action-induced changes. In this work, we reformulate GUI world modeling by shifting the prediction target from full-screen reconstruction to action-induced future information in latent space. Concretely, we leverage a large multimodal model as the latent world model, whose shared representations support two dedicated heads for both action selection and predictive transition modeling. The action head directly translates these latent representations into action selection, while the transition head captures action-conditioned future information. To mitigate the identity shortcut, we further introduce an action-aware contrastive objective that encourages the transition representations to capture discriminative action-induced changes. Extensive experiments across offline and online GUI settings show that our method achieves up to 31.9% and 40.6% relative gains in success rate over pixel-space world models, respectively, while substantially reducing inference latency by up to 98.9%. Our code is available at https://github.com/sjyu001/latent2world.

[9] Image Diffusion Models Are Good 3D Human Motion Generators cs.CVPDF

Siyuan Yang, Chong Wang, Hao Cheng, Jun Liu, Bihan Wen

TL;DR: 该方法将根中心关节轨迹编码为拓扑感知的关节—时间图像，并把全局根动态分成独立潜流；HumanML3D/KIT-ML 上专属模型达到 SOTA，共享训练在两个基准进一步提升。

Details

Motivation: 文本到动作模型通常按数据集骨架布局分别训练，难以共享不同关节定义间的动作知识。作者探索能兼容异构骨架的图像式公共表示，同时保留运动学结构，并利用图像扩散架构生成 3D 人体动作。

Result: 方法把根中心关节轨迹组织成拓扑感知的 joint-time 图像，并用互补潜流表示全局根运动。动作自编码器以根潜变量条件化局部运动解码；文本条件 Scalable Interpolant Transformer 联合生成两路潜变量，并通过解码出的 3D 关节位置监督约束运动空间生成。除针对单个数据集的模型外，作者在 HumanML3D 与 KIT-ML 上联合训练一个动作自编码器，再训练共享的文本条件 SiT，保留各自骨架定义而无需重定向到统一关节布局。实验显示，数据集专属模型在两个基准上达到 SOTA，共享训练在两者上进一步提升。

Insight: 关节—时间图像把拓扑和时序结构映射到适合图像扩散处理的表示，根运动另行建模；共享模型保留原骨架差异而不做统一重定向。摘要未给具体 FID/语义指标数值，也未说明跨数据集提升幅度。

Abstract: Text-to-motion generation typically relies on dataset-specific skeletal representations, with models trained separately for benchmarks with different joint layouts. We investigate whether image diffusion architectures can model 3D human motion through a common image-compatible representation that accommodates heterogeneous skeletal layouts while preserving motion-specific kinematic structure. To this end, we organize root-centered joint trajectories into topology-informed joint–time images and model global root dynamics through a complementary latent stream. A motion autoencoder conditions local-motion decoding on the root latent, while a text-conditioned Scalable Interpolant Transformer (SiT) jointly generates both latent streams, with supervision on decoded 3D joint positions further constraining generation in motion space. In addition to dataset-specific models, we train a single motion autoencoder on HumanML3D and KIT-ML, followed by a single text-conditioned SiT shared across both datasets, preserving their respective skeletal definitions without skeleton retargeting to a common joint layout. Experiments on HumanML3D and KIT-ML show that our dataset-specific models achieve state-of-the-art performance, while shared training yields further gains on both benchmarks, demonstrating the benefits of shared modeling across heterogeneous skeletal layouts.

[10] BigCell: Generating Gigapixel Whole-Slide Images cs.CV | cs.LGPDF

Srikar Yellapragada, Alexandros Graikos, Zilinghan Li, Kostas Triaridis, Tarak Nath Nandi

TL;DR: BigCell 用 2048² flow-matching Transformer 和病理基础模型条件生成整张 gigapixel 组织切片，推理滑窗生成最高 32768²；自适应调度快 3 倍以上，合成数据使切片级 few-shot 分类最高提升 18%。

Details

Motivation: 全切片病理图像是 gigapixel 级，而现有生成模型通常只生成不超过 1024² 的局部块；临床报告和切片诊断标签却是在整张 WSI 层面提供监督。

Result: BigCell 在超过 120,000 张、涵盖多器官的公开 WSI 上，以 2048×2048 分辨率训练 flow-matching Transformer，并使用病理基础模型嵌入作条件。推理时采用滑窗流程生成最高 32,768×32,768 的连贯图像，在 tile 级和 slide 级保真指标上均超过先前 tile 生成器。为使 gigapixel 合成可行，作者引入自适应调度优先处理重要切片区域，在质量轻微下降下将生成速度提高 3 倍以上。另训练单独模型从病理报告生成密集条件网格，实现其所称首个 8K 文本到图像合成。用合成样本增广真实数据后，切片级 few-shot 分类相较真实数据训练基线最高提升 18%。

Insight: 模型以局部窗口训练、滑窗推理拓展至整片尺度，并通过区域调度削减无效计算；文本条件生成另有专门模型。3×以上速度和18%分类收益均为摘要报告的最高值，未列不同组织/任务的具体分数。

Abstract: Whole-slide images (WSIs) are gigapixel-scale digital scans of histopathology tissue and serve as the primary basis for cancer diagnosis. Existing histopathology generative models operate at the patch scale, producing fixed-size tiles of at most 1024$\times$1024 pixels which encompass only a small fraction of the WSI. In contrast, text reports for clinical diagnoses and slide-level diagnostic labels – the supervision that exists at scale – operate at the whole-slide level. We present BigCell, the first approach to generate coherent whole-slide histopathology images at native gigapixel resolution. BigCell trains a flow-matching transformer at 2048$\times$2048 pixels on over 120,000 publicly available WSIs spanning multiple organ types. Conditioning on embeddings from pathology foundation models and using an inference-time sliding-window pipeline, BigCell generates coherent images of up to 32,768$\times$32,768 pixels, outperforming prior tile generators on both tile-level and slide-level fidelity metrics. To make gigapixel synthesis practical, we introduce an adaptive scheduling strategy that prioritizes important slide regions, speeding up image generation by over 3\ttimes with minimal quality degradation. Furthermore, we propose the first text-to-image synthesis at 8k resolution by training a separate model that produces dense conditioning grids directly from pathology text reports. We demonstrate that augmenting with synthetic samples improves slide-level few-shot classification performance by up to $18 %$ over baselines trained on real data.

[11] Thinking with Images without Coordinates cs.CVPDF

Meng Cao, Haoze Zhao, Yi Wang, Ian Reid, Xiaodan Liang

TL;DR: TiCO 将多模态视觉推理建模为潜空间语义证据发现，以 probe token 和连通区域重要性采样形成多样证据轨迹，不使用显式像素坐标；在多项高分辨率和通用推理基准上优于坐标 grounding 方法。

Details

Motivation: “用图像思考”方法常依赖明确坐标 grounding，但稀疏坐标回归难表达高维潜在语义结构，patch token 与像素坐标分辨率也不匹配。

Result: TiCO（Thinking with images without Coordinates）把视觉推理重构为潜空间证据发现，并以 token 级区域注入证据而不输出像素坐标。专用 probe token 估计相似度分布，再对空间连通分量做重要性采样，生成多样且语义显著的证据轨迹。多个高分辨率和通用视觉推理基准实验显示，TiCO 持续优于先进坐标方法且推理效率显著更高；摘要未列具体数据集或分数。

Insight: 不必将视觉 grounding 强制压成点坐标，可在潜空间直接选择语义相关区域作为推理证据，缓解语义表达和空间分辨率错配。摘要未提供提升数值、计算开销或模型范围细节。

Abstract: Recent advances in multi-modal large language models have shifted toward the “thinking with images” paradigm to improve reasoning faithfulness, yet existing methods rely heavily on explicit coordinate- based grounding. We identify two fundamental limitations in this approach, i.e., a semantic gap where sparse regressions miss high-dimensional latent structures, and a resolution gap due to the granularity mismatch between patch-level tokens and pixel-level coordinates. To address these, we propose TiCO (Thinking with images without COordinates). It reformulates visual reasoning as a semantic evidence discovery process within the latent space, and evidence is injected as token-level regions without pixel coordinates. Specifically, TiCO utilizes a dedicated probe token to estimate similarity distributions and performs importance sampling over spatially connected components to generate diverse, semantically dominant evidence rollouts. Extensive experiments across multiple high-resolution and general visual reasoning benchmarks demonstrate that our TiCO consistently outperforms state-of-the-art coordinate-based methods with significantly improved inference efficiency.

[12] RoboFake: Are Deepfake Detectors Ready for Humanoid Face Robots? cs.CV | cs.CR | cs.ROPDF

Hao Chen, Junnan Xu

TL;DR: RoboFake 评估六种按 FaceForensics++ 固定阈值训练的 deepfake 检测器对实体/合成仿人机器人脸的反应；多数检测器将实体机器人误判为假，但反应不随合成而变化且易受轻微 JPEG 压缩影响。

Details

Motivation: 会口型同步、能做细致表情的仿人脸机器人若使用真实人物面孔并播放生成语音，会在摄像机前形成身份伪造，但画面记录的是未被编辑的真实物体。研究检验现有深伪检测器是否能正确识别这类场景。

Result: 研究在实体口型同步机器人视频、同一机器人的合成视频和渲染机器人脸上测试六种用 FaceForensics++ 训练的公开检测器，阈值固定于 FaceForensics++，并与经过同样处理的真实人脸对照。在误报率设为 5% 时，三种检测器几乎将实体机器人所有帧标为假，另两种低于 7%。匹配人脸控制和遮罩实验表明响应来自脸部而非处理或背景；但响应不追踪合成过程：只保留脸部时，六种检测器中的三种给实体机器人的假分数高于其合成视频。检测结果也不稳健，对部分检测器轻度 JPEG 压缩即可使多数机器人帧低于阈值。

Insight: 检测器标记机器人主要因为其外观不像人类，而非识别出伪造过程；因此部署在真实机器人场景时可能产生系统性误报。结论依据固定于 FaceForensics++ 的阈值和所测机器人/检测器，不代表所有 deepfake 检测系统。

Abstract: Humanoid face robots can now move their lips in sync with speech and show detailed facial expressions. Give such a robot the face of a real person and drive it with generated speech, and you get a face forgery that happens in front of the camera: the camera records a real object, and no pixel is edited. We call this a RoboFake. In this paper we ask what current deepfake detectors do when they see a robot face. We test six public detectors trained on FaceForensics++ on video of a physical lip-syncing face robot, on a synthesized video of the same robot, and on rendered robot faces. All thresholds are fixed on FaceForensics++, and every result is compared with real human faces that went through the same processing as the robot videos. At a threshold that flags 5% of real faces, three detectors flag almost every physical robot frame, while two flag fewer than 7%. A matched human control and masking show that the response comes from the face, not from the processing or the background. However, it does not track synthesis: with only the face visible, three of the six detectors give the physical robot a higher fake score than its synthesized video. The response is also fragile, and light JPEG compression is enough to push most robot frames below the threshold for some detectors. Our results suggest that current detectors flag today’s robots because their faces do not look human, not because they recognize a forgery. We release our evaluation protocol and code, and state three testable hypotheses about realistic, interactive RoboFakes.

[13] GeoFix: Geometry Fixing for 4D Consistent Generation cs.CV | cs.AI | cs.LGPDF

Yeobin Hong, Hyojun Go, Dohun Lee, Jeongjae Lee, Sunghwan Hong

TL;DR: GeoFix 用静态场景区域的跨视角深度一致性构造几何奖励，并校正视场重叠造成的奖励偏差；实验中 WorldScore 3D consistency 提高 39%，VBench Quality Score 提高 4%。

Details

Motivation: 视频扩散模型是 4D 生成的重要先验，但生成视频常缺乏几何一致性。强化学习几何奖励可能损害纹理和细节，需再叠加外观奖励并仔细调权；另外，更大相机视场重叠本身就会提高真实视频深度一致性，可能被误当作几何改进。

Result: GeoFix 是用于训练阶段和推理时奖励对齐的统一框架，只需一个几何奖励。奖励测量静态场景区域跨视角深度一致性，目标是结构错误，不直接约束外观，也不惩罚有效物体运动。作者进一步指出，即使在真实视频和 4D 数据集中，深度一致性也会随相机视场重叠增加，因此用重叠相关趋势校准奖励。实验报告 WorldScore 3D consistency 提升 39%，VBench Quality Score 提升 4%。

Insight: 几何奖励应对相机视场重叠这一混杂因素校准，否则更小的相机位移可能被错误奖励为更好几何。报告显示几何一致性和质量分数均升，但摘要未提供基线数值、多个模型结果或运动指标。

Abstract: Video diffusion models increasingly serve as priors for 4D generation, but the generated videos often lack geometric consistency. Recent approaches address this through post-training by designing geometric rewards, yet this can degrade texture and fine detail, requiring auxiliary appearance rewards and careful reward balancing. To address this challenge, we propose GeoFix, a unified framework that improves geometric consistency and visual quality with a single geometric reward, applicable to both training and inference-time reward alignment. Our reward measures cross-view depth consistency in static scene regions, targeting structural inconsistencies without directly constraining appearance or penalizing valid object motion. Crucially, depth agreement increases with camera field-of-view overlap even in real videos and 4D datasets, allowing a smaller camera movement to be mistaken for better geometry. We account for this effect by calibrating the reward against the overlap-dependent trend. GeoFix achieves a 39% improvement in WorldScore 3D consistency and a 4% increase in VBench Quality Score.

[14] Real-Time Traffic Density Estimation: A Hybrid Computer Vision and Predictive Machine Learning Pipeline cs.CVPDF

Prince Kwabena Appiah Boadu

TL;DR: 该交通密度管线先检测跟踪车辆并由道路掩码提取逐帧特征，再用 LSTM/GRU/TCN 预测未来五步；GRU 的汇总 R² 为 0.858，但评估预测的是检测器导出的 telemetry，非绝对交通密度精度。

Details

Motivation: 交通摄像头流量监测既需视觉车辆感知，也需短期趋势预测；要把目标检测、空间道路过滤和时间序列预测串接评估。

Result: 两阶段系统先检测跟踪车辆，再以预计算道路掩码筛除与道路重叠不足的框，提取逐帧交通特征；LSTM、GRU、TCN 使用十帧输入预测未来五步。十种、六类检测器在 5,716 个标注帧上测试（单 NVIDIA GPU）：YOLOv5l 的 mAP@50:95 最高为 0.524、56 FPS；YOLOv8n 为 0.487、258 FPS。FPS 只测检测器，不包含跟踪或空间筛选。时间评估使用 29,243 个留出窗口，GRU/LSTM/TCN 汇总 R² 分别为 0.858/0.856/0.840；这些是对检测器导出 telemetry 的预测，且主要受方向变化影响，不代表绝对交通密度预测准确度。评估流没有道路掩码，因此未验证完整空间过滤；第三路摄像头的 live run 用于观察完整管线行为。

Insight: 检测速度、目标检测精度和交通 telemetry 预测是不同层面的指标，不能把检测 FPS 当作完整管线速度，也不能将高 R² 解读为绝对密度准确。缺少评估流道路掩码限制了空间筛选组件验证。

Abstract: Traffic density estimation from camera feeds requires connecting visual perception with short-horizon forecasting. We present a two-stage pipeline that combines vehicle detection and tracking with a spatial logic layer, followed by a temporal estimator. The spatial layer filters bounding boxes by their overlap with a precomputed road mask and extracts per-frame traffic features. LSTM, GRU, and Temporal Convolutional Network (TCN) estimators use these features to predict five future steps from a ten-frame input window. We benchmark ten detector variants from six families on a dataset of 5,716 annotated frames using a single NVIDIA GPU. YOLOv5l achieves the highest mAP@50:95 of 0.524 at 56 frames per second, while YOLOv8n reaches 258 frames per second at 0.487 mAP@50:95; these rates measure detector inference without tracking or spatial filtering. Temporal evaluation on 29,243 held-out windows yields aggregate R-squared values of 0.858 for GRU, 0.856 for LSTM, and 0.840 for TCN. These metrics measure forecasts of detector-derived telemetry and are dominated by direction variance, rather than establishing absolute traffic-density accuracy. No road mask was available for the evaluation stream, limiting validation of the spatial filtering component. A live run on a third camera stream examines forecasting behavior within the complete pipeline. The results characterize detector speed-accuracy trade-offs and the limitations of short-window traffic telemetry forecasting.

[15] RealityBench: Simulated Environments to Evaluate a Vision Language Model’s Ability to Document Human Physical Work cs.CV | cs.AI | cs.CLPDF

Andrew Perrault, Anmol Takiar, Dominik Wienecki, Arnab Nandi

TL;DR: RealityBench 用 Blender 构建 11 种非结构化体力工作场景，评估 VLM 是否能读懂工作状态并判断画面是否可作凭证；26 个模型擅长读取画中文字，但测量值、小部件和空间关系仍弱。

Details

Motivation: 智能眼镜/相机下的实体工作辅助需要理解工人正在做什么，并协助记录工作。但真实屋顶太阳能安装、车辆维修等现场数据常属专有且变化大，难以控制变量构建基准。

Result: RealityBench 在 Blender 中模拟 11 种工作环境，系统变化画面是否可接受及记录内容。评估两项能力：判断视角是否适合作为完工凭证，以及根据视图理解任务相关状态。26 个 VLM 实验显示，若答案直接写在画面中，顶级模型大体能完成文档记录；但读取测量值、识别小部件及判断空间关系（如卷尺是否倾斜）较弱。画面可接受性难以精确衡量，因为人工评分者彼此有分歧。用与多数人工评分者一致率约等于单个人工评分者的评分器比较，最佳模型无法与一名非作者人工评分者区分，成本也在人类单次通过评估范围内。在三个可追踪错误的环境中，错误拒绝多数源于误读场景而非没理解要求。对三项有真实图像的问题，渲染图上模型排名可迁移到真实图像。

Insight: 基准针对高变现场工作而非结构化厨房/仓库，并揭示不同能力瓶颈：文字读取相对强，精细测量和空间关系薄弱；可接受性判断还受到人工标准分歧限制。真实图像迁移证据仅覆盖三项问题。

Abstract: Augmenting worker capabilities in the physical world through guidance has been a decades-long goal for AI agents. One key supporting capability is to understand what the worker is doing, seeing through the human’s camera or smart glasses, and assist them in documentation. Despite advances in vision language models, their capabilities in these settings remain uncertain and are challenging to benchmark because most physical work data is proprietary, and involves too much variability to study in a principled manner. Unlike structured settings such as kitchens and warehouses that have predictable lighting, geometry, and objects, field-facing tasks such as rooftop solar installation and vehicle repair are unconstrained and high variance. Here, we model real-world physical work in Blender across eleven environments, incorporating variants in both image acceptability and documentation. The acceptability task asks whether a particular view would be suitable proof-of-work and documentation asks whether an agent can understand the task-relevant state of the environment from that view. Across 26 vision-language models, we find that top models largely solve documentation when the answer is text in frame, but are weak at reading measurements and at judging small parts and spatial relations, such as whether a tape measure leans. Acceptability is harder to pin down because human raters disagree with each other. Measured against a scorer that agrees with the raters’ majority about as often as a single rater does, the best models cannot be distinguished from a single non-author rater, at a cost within the range of a human pass. In the three environments where we can trace acceptability errors, most wrong rejections come from misreading the scene rather than the requirement. On the three documentation questions for which we have real images, model rankings on rendered frames transfer to real images.

[16] Learning Transferable World-Action Models from Task-Paired Human Videos cs.CV | cs.ROPDF

Jeongeun Park, Taekyung Kim, Dongyoon Han, Sangdoo Yun

TL;DR: WATCH 用跨视频人类演示预测另一个演示的动作和未来帧，学习可迁移世界动作表征；机器人后训练后，Cosmos Policy/Cosmos3 的 MimicDroid 成功率分别由 40.9%/59.8% 升至 68.3%/77.7%。

Details

Motivation: 标准视频预训练只预测同一视频下一帧，可能依赖当前与未来共享的外观，而未学到能跨场景和具身形态迁移的交互规律。机器人数据稀缺，跨视频人类演示可提供可扩展监督。

Result: WATCH 在不同环境中收集任务配对的人类演示：给定一段演示的视频和动作，模型从另一段演示当前观察联合预测其动作和未来帧，使学习重点转向跨演示共享的任务交互。训练仅需可规模化的人—人配对，不需要匹配的人—机器人演示。机器人后训练前，分析显示模型在未见机器人上的动作和视频预测更好、跨形态运动对齐更强，表示更接近机器人配对数据训练出的表示。机器人后训练后，MimicDroid 的 in-context 成功率在 Cosmos Policy 下由 40.9% 升至 68.3%，Cosmos3 下由 59.8% 升至 77.7%，未见关卡增益最大；无上下文机器人后训练下增益也保留。方法超过同视频预训练和联合训练基线。

Insight: 让一个演示预测另一段演示的动作与未来，迫使模型学习跨场景共享的交互结构，而非单视频外观连续性。成功率数字来自 MimicDroid 的两种机器人策略配置，摘要未列具体机器人和任务数量。

Abstract: World-action models inherit broad visual and physical priors from video pre-training. However, standard within-video next-frame prediction allows them to rely on appearance shared between the context and future, rather than learning interactions that transfer across scenes and embodiments. We introduce WATCH, a cross-video pre-training that learns from task-paired human demonstrations collected across diverse environments. Conditioned on the video and actions of one demonstration, the model jointly predicts another demonstration’s actions and future frames from its current observation. Predicting one demonstration from another encourages the model to learn task-relevant interactions that transfer across scenes and demonstrators. Our pre-training shows that these capabilities can be learned from readily scalable human-human pairs alone, without requiring matched human-robot demonstrations. Analyses before robot post-training show improved motion alignment across embodiments and better action and video prediction on unseen robots, with representations closer to those learned from robot pairs. After robot post-training, WATCH improves MimicDroid in-context success from 40.9% to 68.3% with Cosmos Policy and from 59.8% to 77.7% with Cosmos3, with the largest gains on the unseen levels. Gains persist under context-free post-training across robot embodiments. WATCH outperforms within-video pre-training and co-training baselines, supporting cross-video prediction as a scalable approach to learning transferable world-action representations.

[17] SceneForesee: Learning Scene Evolution for Spatiotemporal Reasoning cs.CV | cs.AIPDF

Chaoyue Li, Shoudong Han, Haoyang Wu, Boxue Yang, Sijia Chen

TL;DR: SceneForesee 将场景演化作为时空推理目标，构建 144.57K 观察—未来配对及 1.05M 实体对应；EvoSee 在 Combined Point-F1 和 Event Global F1 上较 Qwen3.5-9B 分别高 10.45、17.24 分。

Details

Motivation: 多模态模型需在场景发展中跟踪同一实体的运动与交互，并把观察到的变化联系到后续状态。仅问答或静态定位难以测量实体连续性和事件演化理解。

Result: Scene Evolution Understanding（SEU）任务要求模型基于观察视频估计实体历史，并预测一种未来场景延续，包括未来轨迹、事件、来自已观察运动的视觉证据和场景描述。SceneForesee 基准包含 144.57K 时间配对、1.05M 实体对应及 272.49K 配有观察证据的事件。作者提出 EvoSee，通过共享候选状态和联合分配学习未来几何与事件语义；为各实体从观察中选一种延续，并用选中变化更新实体、保留关系和场景表示。SceneForesee 上相对 Qwen3.5-9B，Combined Point-F1 提高 10.45 分，Event Global F1 提高 17.24 分；Dyn-Bench 4D 推理上，问答准确率和 grounding J&F 分别较对应基础系统提升 4.8 和 3.58 分。

Insight: 预测的实体变化同时作为预测目标和后续推理的动态表示，使跟踪、未来几何和事件解释共同建模。基准结果使用不同指标/基线，需分别比较；摘要中的提升均按分数点表述。

Abstract: Large multimodal models (LMMs) need to understand how entities move and interact as a scene develops. This requires following the same entities over time and relating their observed motion to subsequent changes. We formulate Scene Evolution Understanding (SEU), a task for predicting and interpreting this change. Given observed video, a model estimates entity histories and predicts one scene continuation through future trajectories and Events, accompanied by Visual Evidence from observed motion and a scene description. We construct SceneForesee, a benchmark for evaluating entity continuity and scene evolution reasoning through paired observed and future scenes with shared entity identities. It contains 144.57K temporal pairs, 1.05M entity correspondences, and 272.49K Events paired with observed evidence. We propose a spatiotemporal reasoning framework, EvoSee, that learns future geometry and Event semantics through shared candidate states and joint assignment. It selects one continuation per entity from observation and uses the selected changes for entity refinement, retention, and scene representation. On SceneForesee, EvoSee outperforms Qwen3.5-9B by 10.45 points in Combined Point-F1 and 17.24 points in Event Global F1. On the 4D reasoning benchmark Dyn-Bench, it improves question answering accuracy and grounding J&F by 4.8 and 3.58 points over their respective base systems. These results support scene evolution as both a prediction target and a dynamic representation that brings predicted entity changes into spatiotemporal reasoning.The project is available at https://github.com/mikubaka88/SceneForesee-Spatiotemporal-Reasoning.

[18] VisualCoding: Learning Executable Skills for Reliable Visual Reasoning cs.CVPDF

Liu Jiaqi, Jin Ruinan, Li Xiaoxiao

TL;DR: VisualCoding 从有标签样例中迭代获取可执行视觉程序，并对适用输入直接返回程序答案；四模型、四基准、69 类任务中，相较最强基线平均准确率提高 4.45 或 14.78 个百分点。

Details

Motivation: 可复用技能能改善视觉推理智能体，但应用技能时仍需模型自行实现或调用程序并生成答案，执行步骤中的错误/幻觉可能让已学经验无法转化为可靠单次表现。

Result: VisualCoding 结合执行引导的可复用视觉程序获取与适用时直接采用程序答案。Execution-Guided Program Acquisition 将视觉原语及代码组合为任务族 solver，再在有标签训练样例上反复测试与修订。Scope-Aware Direct Execution 则检查输入是否在程序适用范围内，符合时运行保留 solver 并直接返回有效答案。四种模型、四项基准、69 个任务族实验中，Gemini-3.1-Pro 相较最强评测基线平均准确率高 4.45 个百分点；将冻结的 Pro 获取程序复用于 Flash-Lite 时高 14.78 个百分点。消融支持执行引导获取有效，并显示直接执行优于对同一程序做文本化或可调用复用。

Insight: 可靠性增益既来自程序在训练样例上执行反馈迭代，也来自推理时对适用范围内答案直接执行，而非让模型再次用自然语言复述。Flash-Lite 的程序来自冻结 Pro 获取流程，属于跨模型复用；摘要未给各任务族分数。

Abstract: Learning reusable skills can improve visual reasoning agents, yet applying a skill still requires the agent to implement or invoke a procedure and produce the final answer. Errors and hallucinations at these steps can prevent acquired experience from translating into reliable single-run performance. We propose VisualCoding, which combines execution-guided acquisition of reusable visual programs with direct adoption of their answers on applicable inputs. Execution-Guided Program Acquisition composes visual primitives and additional code into family-level solvers, then iteratively tests and revises them on labeled training instances. Scope-Aware Direct Execution runs a retained solver and returns its valid in-scope answer. We evaluate VisualCoding on four models across four benchmarks spanning 69 task families. VisualCoding exceeds the strongest evaluated baseline in mean accuracy by 4.45 percentage points with Gemini-3.1-Pro and by 14.78 points when reusing the frozen Pro-acquired artifacts with Flash-Lite. Ablations support execution-guided acquisition and show that direct execution outperforms textual or callable reuse of the same programs. Code is available at https://github.com/jiaq-liu/VisualCoding.

[19] FreeCHROMA: Generative Chroma Keying from Unaligned Background References cs.CVPDF

Jiayi Wu, Tianfu Wang, Rui Zhao, Florian Schiffers, Zhimin Chen

TL;DR: FreeCHROMA 用参数高效适配的视频扩散模型，从 1–3 张未对齐背景参考图生成合成视频背景并保留前景身份、运动与细节；相较基线优势在相机运动下尤为明显。

Details

Motivation: 传统影视抠像需要协调拍摄背景板并单独做光照协调。近期生成式方法常要求背景预先对齐、只做背景条件前景重光照，或仅用编辑后的首帧描述环境，限制场景构图与相机运动并增加准备成本。

Result: FreeCHROMA 参数高效适配预训练视频扩散模型，用一到三张未对齐背景参考图进行生成式视频合成。方法将前景融入参考环境，同时保留前景身份、动作和细节，并生成时空连贯背景；前景—背景协调在同一次生成中完成，无需额外重光照阶段。可选文本指令用于消歧复杂空间关系和相机运动。训练数据由现成真实单目视频构造伪配对样本：先分离前景/背景，再用基础模型增强，原视频作为训练目标，并通过自动质量校验规模化构造。多个参考忠实度、前景保留、场景融合、协调与整体视频质量指标上，FreeCHROMA 持续优于近期基线，尤其相机运动时收益明显；摘要另有机器人探索实验但未提供细节。

Insight: 从未对齐参考图到整段合成背景，并在同一生成过程中完成调和，放宽了传统合成对参考背景准备和静态相机的限制。结果覆盖多个质量维度，但摘要未给具体分数或机器人实验规模。

Abstract: In film production, chroma keying composites an approved foreground performance into a new environment, but conventional workflows require coordinated background plates and dedicated harmonization. Recent generative methods often require pre-aligned backgrounds, focus on background-conditioned foreground relighting, or represent the environment only through an edited first frame. These restrictions limit background selection, scene composition, and camera motion while adding preparation or editing effort. We present FreeCHROMA, which adapts a pretrained video diffusion model in a parameter-efficient manner for generative video compositing from one to three unaligned background references. The model integrates the foreground into the referenced environment, preserving foreground identity, motion, and fine visual details while generating a spatiotemporally coherent background. Foreground-background harmonization occurs within the same generation process, without a separate relighting stage. Optional text instructions help disambiguate complex spatial relationships and camera behavior. To train the model, we construct pseudo-paired data from readily available real monocular videos through foreground-background separation and augmentation with existing foundation models. The original videos serve as training targets, while automated processing and quality verification enable scalable data construction. Across reference fidelity, foreground preservation, scene integration, harmonization, and overall video quality, FreeCHROMA consistently outperforms recent compositing and conditional video-generation baselines, with particularly clear gains under camera motion. FreeCHROMA relaxes the constraints of existing tools to enable end-to-end generative chroma keying from unaligned inputs, with exploratory experiments demonstrating its potential in robotics.

[20] TartanMatch: Towards Universal Dense Correspondence Across Modalities cs.CV | cs.LG | cs.ROPDF

Hyeokjoon Kwon, Jiting Cai, Ruogu Li, Kritan Bhandari, Geethika Hemkumar

TL;DR: TartanMatch 用单一共享网络覆盖 RGB、事件、热成像、深度和投影 LiDAR 的 25 种有向模态配对；零样本真实数据上，跨模态与同模态 EPE 较最强基线分别降低 61.9% 和 49.6%。

Details

Motivation: 机器人多传感器平台需在复杂环境中做稠密对应，但现有方法通常只支持 RGB-RGB 或少数模态组合。为每个模态对训练独立模型会随传感器种类增长而低效，也无法利用不同模态对之间的正迁移。

Result: TartanMatch 将 RGB 预训练对应模型适配到 RGB、事件、热成像、深度和投影 LiDAR 的全部 25 种有序模态对，通过轻量模态适配和真实/合成混合数据在一个共享网络内完成。零样本真实世界基准上，Endpoint Error 相较最强基线在跨模态匹配降低 61.9%，同模态匹配降低 49.6%。联合训练所有模态对的平均匹配准确率高于逐对训练，尤其改善困难跨模态配对。

Insight: 共享模型不但降低多传感器部署所需模型数量，也利用模态对之间的数据共同性提升匹配精度。EPE 降幅与平均准确率是不同指标；摘要未列各模态组合的绝对结果和测试集细节。

Abstract: Robotic platforms combine multi-modal sensors to operate across diverse conditions, but most dense correspondence methods are limited to RGB-RGB or a small set of modality pairs. Training a model for each modality pair scales poorly with the number of sensors and forgoes potential positive transfer across modality pairs. We introduce TartanMatch, a unified model for dense correspondence across all 25 ordered pairs among RGB, event, thermal, depth, and projected LiDAR. TartanMatch adapts an RGB-pretrained correspondence model to all modality pairs within one shared network using lightweight modality adaptation and a mixture of real and synthetic data. On zero-shot real-world benchmarks, TartanMatch reduces Endpoint Error (EPE) over the strongest baselines by 61.9% for cross-modal matching and 49.6% for same-modal matching. We further show that joint training across all modality pairs improves average matching accuracy over pair-specific training, especially on challenging cross-modal pairs. These results demonstrate that a single model achieves broad multi-modal coverage and state-of-the-art matching accuracy.

[21] Adversarial Gradient Forcing for Robust Few-Step Autoregressive Video Generation cs.CVPDF

Chenxi Xia, Yixuan Xu, Shan Liu, Jiapeng Wang, Maohua Li

TL;DR: AGF 针对自回归视频生成中的局部历史鲁棒性缺口，用未来块分布匹配梯度构造对抗历史扰动并联合训练；在 60 秒生成中 VBench Total 最高、4 分钟生成中视觉质量最佳，训练仅需 5 秒 rollout。

Details

Motivation: Self Forcing 等方法通过模型自生成轨迹减轻训练/推理历史分布不匹配，但长程生成仍会累积误差。作者认为模型对自身生成历史中的小错误过于敏感，偏差会被后续块放大。

Result: Adversarial Gradient Forcing（AGF）使用未来块预测的分布匹配梯度，在自生成历史上选择对抗扰动，再让生成器同时在原始和扰动历史上训练，以提高后续生成鲁棒性。实验报告 AGF 在 60 秒视频生成上取得最高 VBench Total，在 4 分钟生成上取得最佳视觉质量；尽管生成时程很长，训练只需 5 秒 rollout。摘要没有列出各指标具体分数或测试模型。

Insight: 该方法直接训练模型对历史误差扰动保持稳健，而不仅是让训练轨迹匹配推理历史分布。5 秒指训练 rollout 时长，不是最终视频时长；长时程质量结论为不同基准/指标上的最佳结果。

Abstract: Autoregressive video generation models generate videos sequentially, with each new block conditioned on previously generated blocks. By training on the model’s own rollouts, Self Forcing and its follow-up methods substantially mitigate exposure bias caused by the mismatch between training- and inference-time history distributions. However, these models still suffer from error accumulation during long-horizon generation. We identify a key underlying limitation: the model remains highly sensitive to errors in its generated history, allowing small deviations to be amplified through subsequent generations. We refer to this underexplored problem as the local history robustness gap. To address this gap, we introduce Adversarial Gradient Forcing (AGF), a framework designed to improve the robustness of few-step autoregressive video models to perturbations and accumulated errors in their own generated histories. AGF uses distribution matching gradients from future block predictions to select adversarial perturbations of self-generated histories. It then jointly trains the generator on the original and perturbed histories to improve the robustness of future generation. Extensive experiments show that AGF achieves the highest VBench Total score on 60-second video generation and the best visual quality on four-minute video generation. Notably, AGF achieves such long-horizon robustness while requiring only 5-second rollouts for training.

[22] FrameFold: Response-Guided KV Cache Consolidation for Autoregressive Long Video Generation cs.CVPDF

Muyu Liu, Mingming Liu, Hong Gu, Yuyao Zhang, Qin Lu

TL;DR: FrameFold 在固定 KV 预算下比较合并前后同一历史空间查询的注意力响应，以此构造代表并选合并对；在最长 8 分钟自回归视频生成中比基线质量退化更少。

Details

Motivation: 自回归视频生成若保留所有历史 KV 状态，随着片段增长会增加内存和每步注意力开销；固定缓存预算虽可限成本，但需保留对未来一致性和画质真正重要的历史内容。

Result: FrameFold 是免训练 KV 缓存合并方法。作者发现空间查询依赖不同历史内容，当前注意力覆盖率高并不必然意味着长视频质量更好。FrameFold 比较同一批近期历史查询在合并前后检索到的内容，利用响应差异指导代表构造与合并选择。它合并相邻历史条目中时间冗余的 KV 状态为共享代表，同时保留完整空间网格；通过闭式解拟合 value 合并系数以匹配条件注意力响应，再选择剩余响应失真最低的合并。在匹配 KV 缓存预算下，Self Forcing/Causal Forcing 的 30/60 秒、LongLive 的 120 秒实验显示一致性和视觉质量改善；连续 8 分钟 Self Forcing rollout 中，质量退化也少于基线。

Insight: 缓存合并目标应是保留未来查询所需的条件注意力响应，而不是仅按当前显著性挑帧；完整空间网格避免破坏空间分辨率。各时长结果来自不同生成方法，摘要未给具体图像质量数值或缓存大小。

Abstract: In autoregressive video generation, retaining all historical key-value (KV) states increases memory usage and per-step attention costs as videos grow longer. A fixed cache budget bounds these costs but raises the challenge of preserving historical content that supports consistent, high-quality generation. Existing methods control cache size through history selection or consolidation, yet deciding what to retain and how to limit the loss of useful content remains a shared challenge. Our analysis shows that spatial queries depend on different historical content and that higher current attention coverage does not necessarily yield better long-video quality. We introduce FrameFold, a training-free KV cache consolidation method that compares the content retrieved by the same recent historical queries before and after merging to guide both representative construction and merge selection. Exploiting temporal redundancy between adjacent historical entries, FrameFold merges their KV states into shared representatives while retaining complete spatial grids. It fits value-merging coefficients in closed form by matching conditional attention responses, then selects the merge with the lowest remaining response distortion. Under matched KV cache budgets, experiments on Self Forcing and Causal Forcing at 30 and 60 seconds and on LongLive at 120 seconds demonstrate improved consistency and visual quality. Continuous eight-minute Self Forcing rollouts further show less quality degradation than the baselines.

[23] Attribute-Centric Vision-Language Alignment for Fine-Grained Gait Understanding cs.CVPDF

Saihui Hou, Chenye Wang, Wenpeng Lang, Qiannan Guo, Zhengxiang Lan

TL;DR: RA-GAR 发布含 1,001 名受试者、242,729 段步行序列和 15 项步态属性的大规模数据集；Gait-VL 将属性特定多帧视觉证据与文本类别原型对齐，并改善身份识别迁移。

Details

Motivation: 步行序列包含可解释的身体结构、姿态和运动属性，属性识别可支持细粒度步态理解。但现有数据集细粒度标注有限，且分类器通常仅用学习到的输出权重表示候选类别，而未直接对齐视觉证据和文字类别概念。

Result: RA-GAR 是大规模多模态步态属性基准，包含 1,001 名受试者、242,729 段步行序列和 15 项静态/运动属性标注。Gait-VL 用属性中心的视觉—语言对齐识别步态：属性专属可学习 query 聚合空间位置和多帧 RGB 特征，形成每属性视觉嵌入；语言侧将大模型扩展的多样描述与相关身体区域上下文结合为文本类别原型。作者还将属性预训练视觉编码器替换身份识别流程中的通用上游视觉模型；该预训练不使用身份分类或度量学习目标。在 RA-GAR 上属性识别达到 SOTA，并在多个流程和基准上持续改善下游身份识别。

Insight: 显式属性语义可作为步态表示的监督，并迁移至身份识别，即使预训练没有身份标签。摘要未提供具体属性识别分数、身份识别增益或基准名称，结论为跨多个设置的一致改善。

Abstract: Walking sequences reveal human-interpretable attributes related to body structure, posture, and motion. Gait attribute recognition provides a concrete route to fine-grained gait understanding by predicting these attributes. However, existing benchmarks provide limited fine-grained annotations, while classifier-based gait attribute recognition methods typically represent candidate classes through learned output weights rather than directly matching visual evidence with textual class representations. To address the annotation gap, we present RA-GAR, a large-scale, richly annotated multimodal benchmark for gait attribute recognition, containing 1,001 subjects, 242,729 walking sequences, and annotations for 15 static and motion gait attributes. We further propose Gait-VL, an attribute-centric vision-language alignment framework that recognizes gait attributes by matching attribute-specific multi-frame visual evidence with textual class prototypes. On the visual side, attribute-specific learnable queries aggregate RGB features across spatial locations and frames into per-attribute visual embeddings. On the language side, diverse descriptions expanded by a large language model are combined with context about relevant body regions to construct textual class prototypes. Beyond attribute recognition, we evaluate the transferability of the attribute-pretrained visual encoder by using it to replace generic upstream large vision models in identityoriented gait recognition pipelines; attribute pre-training uses no identity classification or metric-learning objectives. Experiments demonstrate state-of-the-art attribute recognition performance on RA-GAR and consistent improvements in downstream identity recognition across multiple pipelines and benchmarks, supporting the effectiveness and transferability of attribute-centric gait representation learning.

[24] Bidirectional Adaptive Transformers for Multimodal Anomaly Detection cs.CVPDF

Yuxin Jiang, Yunkang Cao, Weiming Shen

TL;DR: BAT 用双向互预测对齐 RGB 纹理与点云几何，并以轻量适配器微调预训练特征；MVTec 3D-AD 和 Eyecandies 实验报告异常检测 SOTA。

Details

Motivation: 多模态异常检测需有效结合 RGB 图像和 3D 点云。依赖冻结预训练编码器的静态融合虽能对齐空间特征，却缺乏潜在语义对应机制，容易产生跨模态干扰并降低异常区分能力。

Result: Bidirectional Adaptive Transformers（BAT）通过双向特征对齐模块，以互预测机制连接 2D 纹理模式与 3D 几何结构：增强点云描述的纹理细节，同时扩展 RGB 特征的空间上下文。为提高对齐表征的信息密度，框架加入轻量图像和点云适配器，对预训练模型进行参数高效专项微调，以减少冻结编码器的信息损失。在 MVTec 3D-AD 和 Eyecandies 上的实验显示，BAT 达到所比较方法中的 SOTA，并显著优于现有多模态异常检测方法；摘要未给出具体分数。

Insight: 双向互预测让两种模态彼此补充，而非静态拼接；轻量适配则允许任务特定特征对齐。性能结论限于两个基准，摘要未提供异常定位/检测分项指标或资源开销。

Abstract: The effective integration of features derived from RGB images and 3D point clouds presents a substantial challenge in the field of multimodal anomaly detection (MAD). Existing MAD methods typically rely on static fusion paradigms built upon frozen pre-trained encoders. While these frameworks facilitate spatial alignment, they lack the mechanism to establish latent semantic correspondence, thereby inducing cross-modal interference that degrades anomaly discriminability. To mitigate these limitations, this study proposes a Bidirectional Adaptive Transformers (BAT) framework, which constructs comprehensive and semantically aligned multimodal representations by leveraging complementary attributes across modalities. The BAT framework introduces a Bidirectional Feature Alignment (BFA) module, which employs a mutual predictive mechanism to bridge the gap between 2D textural patterns and 3D geometric structures. This bidirectional interaction explicitly enhances the textural granularity of point cloud descriptors while augmenting the spatial contextual awareness of RGB features. To ensure that the latent representations possess sufficient informational density for alignment, BAT incorporates a parameter-efficient adaptation architecture. By utilizing lightweight image and point cloud adapters, the framework facilitates specialized fine-tuning of pre-trained models, mitigating the information loss inherent in frozen encoders. Extensive experiments conducted on MVTec 3D-AD and Eyecandies confirm that BAT achieves state-of-the-art anomaly detection performance, significantly outperforming existing MAD methods.

[25] EviScene: Beyond “Is It AI?” for Agentic Image Factuality Assessment cs.CVPDF

ZiAn Wang, Ruiqi Liu, Zhengbo Zhang, Yi Han, Jialiang Shen

TL;DR: EviScene 将图像事实性判断定义为有证据支持/反驳的结论，而非仅判断是否 AI 生成；Qwen3.5-9B 在 1,527 张测试图上达到 79.05% 平衡准确率、52.26% 严格证据充分率。

Details

Motivation: 判断图像是否由 AI 生成，不能说明图中情境是否有事实证据支持。若没有给定文本主张，系统还需在调查过程中逐步细化核查问题，并将外部证据与图像呈现的具体情境对齐。

Result: 作者提出图像事实性评估任务：输入图像和通用核查指令，系统输出“支持/反驳”判断及基于证据的解释。EviLens 是专家审阅基准，含 10,174 张来自受控构造和网络收集的支持/反驳图像，覆盖 12 类事实错误，同时评估结论准确度和证据充分性。EviScene 是多模态 ReAct 智能体，交替进行视觉检查、外部检索和证据分析，并根据新发现细化问题、回看相关图像细节。训练先以审阅过的教师轨迹监督微调，再对策略自身调查得到的验证决策做特权 on-policy 自蒸馏。在 1,527 张测试图上，Qwen3.5-9B 的平衡准确率为 79.05%、严格证据充分率 SESR 为 52.26%，较同主干基础智能体分别提高 6.71 和 9.37 个百分点；其 SESR 为被评估智能体中最高。

Insight: 事实性评估要求不仅猜对，还需证明证据足以支持结论；迭代提问和回看图像连接外部来源与具体场景。严格证据充分率显著低于平衡准确率，说明证据质量仍是瓶颈；结果限于该测试集和智能体比较。

Abstract: Determining whether an image is AI-generated does not establish whether the situation it depicts is factually supported. We formulate Image Factuality Assessment: given an image and a generic verification instruction, a system returns a supported or refuted verdict with an evidence-grounded explanation. Without a supplied textual claim, the system must progressively refine factual questions during investigation and align external evidence with the specific situation depicted. We introduce EviLens, an expert-reviewed benchmark of 10,174 supported and refuted images from controlled construction and web collection, covering twelve factual error types and evaluating both verdict accuracy and evidence sufficiency. We develop EviScene, a multimodal ReAct agent that interleaves visual inspection, external retrieval, and evidence analysis, using new findings to refine its questions and revisit relevant image details. We train a compact policy with supervised fine-tuning on reviewed teacher trajectories, followed by privileged on-policy self-distillation applied to verified decisions from the policy’s own investigations. On the 1,527-image test set, the trained Qwen3.5-9B agent achieves 79.05% balanced accuracy and a Strict Evidence Sufficiency Rate (SESR) of 52.26%, improving over the same-backbone base agent by 6.71 and 9.37 percentage points, respectively. Its SESR is the highest among the evaluated agents.Code, models, and data will be released at https://github.com/wangzian828/EVISCENE and https://huggingface.co/ZiAnwang/EviScene.

[26] EarthAlign: Learning Language-Aligned Representations of Earth System Data cs.CV | cs.AI | cs.LGPDF

Fadillah Adamsyah Maani, Numan Saeed, Jiaqi Ma, Dat Tien Nguyen, Youssef Wehbe

TL;DR: EarthAlign 汇集 2000–2024 年、多地理区域的事件与常态天气样本，并以 63 项 ERA5 变量配四种 grounded caption；MeteoLIP 在跨模态检索、天气分类和物理量估计上超过相应 CLIP/SigLIP 基线。

Details

Motivation: 理解地球系统条件对气候与天气敏感行业决策很重要，但现有多模态天气系统多面向专用预测/生成/问答；缺少多样且与气象观测明确对应的语言监督，使通用气象—语言对齐研究不足。

Result: EarthAlign 覆盖 2000–2024 年和广泛地理区域，整合十类专家整理的自然事件/灾害来源及由气候态派生的常态天气样本。每条样本包含时间、地理位置和四种 grounded 描述风格，并对齐 63 个 ERA5 气象变量。作者提出 MeteoLIP 双编码器学习气象—语言表示，其中 MetTok 使用位置和时间条件 query 将变量特定 patch 嵌入聚合为紧凑气象 token。评测包括跨模态检索、天气状况分类和物理变量估计；实验中 MeteoLIP 在这些任务上超过对应 CLIP/SigLIP 基线。分析还强调气象输入、时空条件和多风格监督的重要性。

Insight: 显式地理/时间元数据和多种 grounded caption 为气象观测与语言描述建立对应，MetTok 则把大量变量压缩为紧凑表征。摘要没有给数据量或具体下游指标增益，因此只能报告相对基线的方向性结果。

Abstract: Interpreting Earth system conditions is central to understanding environmental risks and informing decisions across weather- and climate-sensitive sectors. Existing multimodal weather systems primarily target task-specific prediction, generation, or question answering, while general-purpose alignment between Earth system data and language remains underexplored due to the limited availability of diverse language supervision grounded in meteorological conditions. To fill this gap, we introduce EarthAlign, a large-scale dataset with broad geographic coverage spanning 2000-2024 that combines 10 expert-curated natural event and disaster data sources with climatology-derived normal-weather samples. Each sample includes temporal and geospatial information and four grounded caption styles aligned with 63-variable ERA5 meteorological data. Building on EarthAlign, we introduce MeteoLIP, a dual-encoder framework for learning aligned meteorology-language representations. MeteoLIP incorporates MetTok, which uses a geolocation- and time-conditioned query to aggregate variable-specific patch embeddings into a compact sequence of meteorology tokens. To comprehensively assess Earth system multimodal representations, we introduce three downstream tasks: cross-modal retrieval, weather-condition classification, and physical variable estimation. Our experiments show that MeteoLIP improves upon the corresponding CLIP and SigLIP baselines across these tasks. Further analyses highlight the importance of meteorological input, temporal and geospatial conditioning, and multi-style supervision for meteorology-language alignment. Thus, EarthAlign and MeteoLIP provide a foundation for learning and evaluating language-aligned representations of Earth system data. Our code, dataset, and weights are publicly available at https://github.com/mbzuai-oryx/EarthAlign.

[27] TAG: Calibrating SAM3 for Universal Image Segmentation without Dense Supervision cs.CVPDF

Hao Zhu, Heyu Sun, Shuo Jin, Jiayu Xiao, Feng Dai

TL;DR: TAG 认为 SAM3 下游分割瓶颈主要是概念是否存在的识别校准，而非已识别对象的掩码质量；仅用图像级类别标签和 3.1M 参数适配器即可提升多场景表现。

Details

Motivation: SAM3 将可提示分割从空间提示扩展到概念提示，有望实现通用图像分割，但在下游任务中表现不总可靠。作者发现模型通常能准确描绘已识别对象，却难稳健判断图像中到底有哪些概念。

Result: Text Anchor re-Grounding（TAG）面向目标域、无需密集标注来校正 SAM3 概念识别。Anchor Residual Pathway 使用图像条件低秩适配器，将文本锚点重新 grounding 到目标分布，并调节后续文本条件交互。训练只依赖图像级类别标签，联合使用二元交叉熵与 Soft Boundary Ranking，以校准类别存在置信度并区分真实概念和易混负例。Confidence-Weighted Distillation 进一步将 MLLM 丰富语义上下文与 SAM3 细粒度空间先验结合，在没有人工密集标注时完成校准。作者报告只需 3.1M 可训练参数，方法在多种场景提升 SAM3 能力，并在若干指标上超过领先掩码监督方法；摘要未给具体指标值。

Insight: 若掩码已能精确描绘对象，错误更可能来自“目标概念是否存在”的判断。TAG 因此以图像级标签改善文本概念锚定，并用置信加权蒸馏结合语义和空间先验；超越掩码监督方法仅报告于若干指标，非所有设置。

Abstract: The recent Segment Anything Model 3 (SAM3) extends powerful promptable segmentation from spatial cues to concepts, raising the prospect of a single foundation model for universal image segmentation. However, this promise often fails to translate into reliable performance on downstream segmentation tasks. In this work, we empirically identify concept recognition, rather than mask quality, as the primary bottleneck: SAM3 accurately delineates the objects it recognizes yet struggles to robustly determine which concepts are present. Driven by this, we present \textit{\underline{T}ext \underline{A}nchor re-\underline{G}rounding (\TAG{})}, a simple yet effective framework that rectifies the concept recognition of SAM3 on the target domain without dense annotations. Specifically, \TAG{} introduces an \textit{Anchor Residual Pathway} comprising image-conditioned, low-rank adapters to re-ground text anchors in the target distribution and modulate subsequent text-conditioned interactions. Guided solely by image-level category labels, \TAG{} jointly optimizes binary cross-entropy with \textit{Soft Boundary Ranking} objectives to calibrate presence confidence and distinguish present concepts from confusable negatives. Building upon this, we devise \textit{Confidence-Weighted Distillation} to couple rich semantic context of a multimodal large language model with fine-grained spatial priors of SAM3, thereby achieving effective calibration in the absence of manual annotations. Extensive experiments demonstrate that \TAG{} unlocks SAM3’s segmentation potential with only 3.1M trainable parameters, achieving impressive performance across diverse scenarios, and notably, even surpassing leading mask-supervised methods on several metrics. \href{https://github.com/MiSsU-HH/TAG_SAM3}{Our code will be publicly available at GitHub \faGithub}.

[28] Prospective Forcing: Learning to Look Ahead for Autoregressive Long Video Diffusion cs.CV | cs.AIPDF

Yifan Shen, Yewei Xia, Gongxu Luo, Peiyuan Zhu, Boyang Sun

TL;DR: Prospective Forcing 用多个未来时间偏移预测头拓展自回归视频训练视野，并在推理时经候选验证跳过完整滚动窗口执行；摘要提供了机制但未给实验结果数值。

Details

Motivation: 自回归视频生成按片段滚动，但通常只训练预测紧邻的下一个片段，预测视野短会使长时 rollout 难以维持质量和时间一致性。

Result: Prospective Forcing 在滚动窗口生成器上加入轻量、偏移特定的多头预测。各头共享相同条件信息，分别预测不同未来时间偏移的片段，以较长视野监督促使共享历史表示保留长程预测所需信息。推理时可使用被接受的未来片段预测，跳过完整窗口主干执行；头部通过可学习嵌入生成下一个片段的多个候选，学习式验证器选择一个，或全部拒绝并回退到标准自回归生成。摘要没有给出视频质量、加速幅度或模型/基准实验结果。

Insight: 训练阶段增加远期片段监督，推理阶段再以候选和验证器安全跳过部分窗口计算。缺少摘要实验结果时，只能描述框架设计，不能确认其实际质量和效率收益。

Abstract: Autoregressive video generators synthesize long videos through chunk-wise rollouts, yet they are typically trained to predict only the immediate future. This limited prediction horizon makes it challenging to maintain video quality over extended rollouts. To address this problem, we introduce Prospective Forcing, a multi-head prediction framework that injects long-range predictive signals into autoregressive video representations. Built on a rolling-window generator, Prospective Forcing attaches lightweight, offset-specific heads that share the same conditioning information and independently predict chunks at different future temporal offsets. By extending supervision beyond the next chunk, these heads encourage the shared history representation to retain information useful for longer-horizon prediction, supporting temporally consistent generation. At inference time, we accelerate sampling by using accepted future-chunk predictions to skip full-window backbone executions. We equip the head with learnable embeddings to generate multiple candidates for the next chunk, and a learned verifier either selects a candidate or rejects all and falls back to standard autoregressive generation.

[29] TGuide: Faster Autoregressive Video Diffusion via Historical Trajectory Guidance cs.CVPDF

Zhuo Chen, Bin Zou, Fanyue Wei, Lixin Duan, Angela Yao

TL;DR: TGuide 用当前片段前一去噪步特征与前一视频片段同去噪步特征共同预测速度，在 Self Forcing 等模型上最高加速 1.89×；Self Forcing 可从 81 帧训练泛化到 333 帧生成。

Details

Motivation: 自回归视频扩散仅剩四个去噪步时仍有加速空间，但相邻去噪步间特征变化大，使直接缓存或轻量预测器难以复用。另一方面，视频不同片段在匹配去噪步的特征仍高度相关。

Result: TGuide 使用轻量 Anchor-Transport-Correct（ATC）预测器从历史采样轨迹加速生成。当前片段前一去噪步特征保留空间结构，前一片段相同去噪步特征提供目标时间阶段信息；通过空间检索与自适应修正融合两类参照，提高大去噪步间隔下的速度预测。为减轻跨去噪步和跨视频片段误差累积，预测器在自身自回归 rollout 上训练，适配由先前预测塑造的状态；可选置信度调度器调整质量—效率折衷。在 Self Forcing、Causal Forcing 和 HY-WorldPlay 上，质量保持同时相对原四步生成器最高加速 1.89×。Self Forcing 从 81 帧训练视频泛化到 333 帧生成，约为训练长度四倍，且 VBench 分数超过原模型。

Insight: 跨片段同一步信息补充跨去噪步的局部缓存，两种时空参照共同提升速度预测；自生成轨迹训练让预测器适配自身误差分布。1.89×为最高加速，长视频泛化结果限定于 Self Forcing。

Abstract: Can autoregressive video diffusion be accelerated further when only four denoising steps remain? Existing feature caching approaches struggle in this regime, as large feature changes between denoising timesteps undermine both direct reuse and lightweight predictors. We observe that features across video chunks remain strongly correlated at matching denoising timesteps, offering a complementary source of guidance. Building on this insight, we introduce TGuide, a framework that accelerates autoregressive video diffusion through a lightweight Anchor-Transport-Correct (ATC) predictor guided by historical sampling trajectories. Our ATC combines two complementary references: the current chunk’s preceding-step features preserve its spatial structure, while the previous chunk’s same-timestep features provide information at the target denoising stage. By integrating these references through spatial retrieval and adaptive correction, ATC improves velocity prediction accuracy despite the large gaps between denoising steps. To mitigate error accumulation across both denoising steps and video chunks, we train the predictor on its own autoregressive rollouts, allowing it to adapt to states shaped by its own earlier predictions. An optional confidence-aware scheduler further adjusts the quality-efficiency trade-off. Experiments on Self Forcing, Causal Forcing, and HY-WorldPlay show that TGuide preserves generation quality while achieving denoising speedups of up to $1.89\times$ over the original 4-step generators. On Self Forcing, TGuide generalizes from 81-frame training videos to 333-frame generation, approximately four times the training length, while exceeding the original model’s VBench score.

[30] TETRIS-4DGS: Efficient Post-Training Compression of 4D Gaussian Splatting cs.CVPDF

Chengzhen Wu, Kelei Liu, Sicheng Li, Lu Yu, Yiyi Liao

TL;DR: TETRIS-4D 用重要性自适应量化和运动/重要性驱动打包，将 4D Gaussian Splatting 后训练压缩为通用 codec 兼容流程；数秒内可实现 40× 以上压缩并保持较高质量。

Details

Motivation: 4D Gaussian Splatting 能实时高质量渲染动态场景，但存储开销很大。已有压缩方法多按场景分别优化，不易泛化，也不适用于要求低延迟的压缩应用。

Result: TETRIS-4D 利用现有图像与视频 codec，提出通用的 4DGS 后训练压缩框架。对 Gaussian 按重要性级别做自适应量化，以优化码率分配且兼容现有编码器；再利用 4DGS 统计特性，以运动和重要性感知打包重新组织属性，映射为适合 codec 的二维图。作者报告该方法可在数秒内实现超过 40× 压缩，同时保持较高质量，摘要未给具体场景、文件大小或渲染指标。

Insight: 压缩不需每个场景单独迭代优化，而通过通用属性量化和 codec 友好布局实现快速后处理。40×以上为压缩倍率，质量描述为“较高”，没有具体率失真曲线或场景依赖数据。

Abstract: 4D Gaussian Splatting (4DGS) has emerged as a powerful dynamic scene representation, supporting real-time high quality rendering but suffers from prohibitive storage overhead. Existing methods typically attempt to solve this by treating compression as a per-scene optimization problem, limiting generalizability and applicability where low-latency compression is required. In this work, we leverage existing image and video codecs and propose a generalized post-training compression framework for 4DGS to address these issues. Specifically, we apply importance-based adaptive quantization to Gaussians according to their importance levels, effectively refining bitrate allocation while remaining fully compatible with existing codecs. To further improve compression efficiency, we utilize the statistical properties of 4DGS and propose motion- and importance-aware packing to reorganize attributes into codec-friendly 2D maps. Our approach can achieve over 40x compression rate in seconds while maintaining high quality, facilitating the practical deployment of 4DGS.

[31] SPARE: Shared Perception and Adaptive Retrieval of Evidence for Streaming Video Understanding cs.CVPDF

Senbin Xu, Zijun Zhang, Xinhua Wu, Li He, Jubo Feng

TL;DR: SPARE 先判断当前片段证据是否足够，再按需检索历史视频并训练模型在证据仍不足时弃答；在 OVO-Bench 任务组和 StreamingBench 上分别达 77.78% 与 73.31%。

Details

Motivation: 流式视频问答所需证据可能来自当前片段或较早事件。历史上下文有助回忆，却可能干扰基于近期帧的判断；是否引入历史以及选择哪些片段都重要，相关性本身也不保证证据充分。

Result: SPARE 结合自适应历史检索与视觉编码器共享。轻量门控判断近期观察是否足以回答；需要更多证据时，多模态嵌入模型检索相关历史片段补充上下文。LoRA 指令微调让理解模型联合利用近期观察和检索历史，并在仍无充分证据时弃答。为减少冗余视觉参数，理解与检索分支共享视觉编码器、保留独立语言主干；四个线性适配器通过视觉 token 和最终检索嵌入两级蒸馏对齐共享特征。仅用视觉输入，SPARE 在 OVO-Bench 所评 Real-time/Backward 任务组得 77.78%，在 StreamingBench 得 73.31%；相较 2fps 评测的 Qwen3-VL-8B-Instruct，分别提高 20.68 和 14.15 个百分点。

Insight: 按需历史检索和显式证据充分性门控，可避免无差别引入长历史；共享视觉主干则降低检索/理解重复编码。OVO 分数只覆盖文中所评任务组，基线对比固定为 Qwen3-VL-8B-Instruct、2fps。

Abstract: Streaming video understanding requires answering questions from the observed video, with evidence drawn from recent observations or earlier events. Historical context can improve recall but may interfere with judgments based on recent frames. Its usefulness depends on both when it is introduced and which clips are selected, while relevance alone does not guarantee sufficient evidence for an answer. We propose SPARE, a framework that combines adaptive historical retrieval with visual encoder sharing. A lightweight gate predicts whether recent observations provide sufficient evidence to answer the question. When additional evidence is predicted to be necessary, a multimodal embedding model retrieves relevant historical clips to supplement the current context. LoRA instruction tuning trains the understanding model to use recent observations and retrieved history jointly and abstain when the evidence remains insufficient. To reduce redundant visual parameters, the understanding and retrieval branches share a visual encoder while retaining separate language backbones. Four linear adapters align the shared features with the retrieval model through distillation at the visual token and final retrieval embedding levels. With visual-only inputs, SPARE achieves 77.78% on the evaluated Real-time and Backward task groups of OVO-Bench and 73.31% on StreamingBench. Compared with Qwen3-VL-8B-Instruct evaluated at 2 fps, SPARE improves accuracy by 20.68 percentage points on OVO-Bench and 14.15 percentage points on StreamingBench.

[32] Competition Domains Matter: Selective Head Globalization for Training-Free Sparse Attention in Video Diffusion Transformers cs.CVPDF

Tiantian Liu, Meng Wan, Zhiqiang Liang, Jue Wang, Ningming Nie

TL;DR: EG-Sparse 根据注意力熵为低熵头选用跨头全局竞争池、为高熵头保留逐行局部选择；四种配置下视频质量更高且注意力密度与 GPU 去噪延迟更低。

Details

Motivation: 免训练稀疏注意力可加速视频扩散，但现有方法通常让所有注意力头使用固定 query 行内竞争域，主要研究如何在该域中估计 token 重要性。不同注意力头的异质性可能意味着竞争范围也应自适应。

Result: EG-Sparse 是免训练框架：低熵头路由到分组跨头候选池，高熵头继续逐行局部选择，仅使用稀疏管线已计算出的分数。控制诊断固定排序分数、只改变竞争规则：跨头分组使低熵头重建误差降低 9.8%，而强迫高熵头进入全头全局池使误差增加 7.9%。在四种配置下，EG-Sparse 相较强免训练稀疏注意力基线提升视频质量，同时降低注意力密度和 GPU 去噪延迟。机制消融显示，选择性适配优于统一、随机或固定竞争域分配。

Insight: 稀疏选择不仅要估计 token 排名，也要为不同熵/头异质性选择合适的候选竞争域；统一全局或局部规则可能分别伤害某些注意力头。摘要未列四配置的绝对延迟和质量分数。

Abstract: Training-free sparse attention is a practical route to efficient video diffusion transformer inference, yet existing methods typically fix the competition domain at each query row for every attention head. Prior work primarily addresses how to estimate token-level importance within this fixed domain; whether the domain itself should adapt to head heterogeneity remains under-specified. We introduce EG-Sparse, a training-free framework that routes low-entropy heads to a grouped cross-head pool while retaining per-row local selection for high-entropy heads, using only scores already materialized by the sparse pipeline. A controlled diagnostic fixes the ranking score and varies only the competition rule: grouped cross-head selection reduces low-entropy-head reconstruction error by 9.8%, whereas forcing high-entropy heads into an all-head global pool increases it by 7.9%. Across four configurations, EG-Sparse achieves higher video quality while reducing attention density and GPU denoising latency compared with strong training-free sparse-attention baselines. Mechanism ablations further show that selective domain adaptation outperforms uniform, random, and fixed domain-assignment policies. These results support competition-domain adaptation as a distinct design axis within centroid-guided training-free sparse attention.

[33] CAST-ASD: Causal Attention over Speakers and Time for Online Active Speaker Detection cs.CV | cs.SDPDF

Pratham Gupta, Haotian Qi, Tushar Sandhan, Gabriel Skantze

TL;DR: CAST-ASD 在零前视因果推理中结合当前帧跨人脸注意力与各人历史时间注意力，既共享说话者信息又保持低延迟；MSDWild 上相较单脸因果基线，竞争说话者误报减少 32%。

Details

Motivation: 联合分析场景中所有人脸可提升主动说话人检测，但通常需要整段视频，无法实时输出。低延迟因果方法通常逐脸独立处理，缺少跨说话者信息共享。

Result: CAST-ASD 将场景级当前帧跨脸注意力与每张人脸自身历史的因果时间注意力结合，在零 lookahead 条件下共享说话人信息；参数量不随脸数增长。在多数负脸帧同时出现竞争说话者的 MSDWild 上，相较其扩展的逐脸因果检测器，竞争说话者误报降低 32%，且场景人数越多收益越大。模型在两个验证划分上超过所有因果基线；与三块固定 roster 卷积相比，跨说话者参数量仅为其十分之一。

Insight: 当前帧内的跨脸信息可在不等待未来帧的情况下消除竞争说话者造成的误报，同时模型规模不依赖场景人数。32%为相对指定因果基线的误报减少，数据来自 MSDWild。

Abstract: We present CAST-ASD (Causal Attention over Speakers and Time), a causal joint model for active speaker detection (ASD). Recent ASD models improve performance by jointly processing all faces in a scene, but operate offline, with the full clip available before any frame is labelled. Low-latency causal detectors instead process each face independently. CAST-ASD combines joint face reasoning with fully causal inference. It applies scene-level attention across faces within the current frame and causal temporal attention over each face’s history, enabling information sharing across speakers with zero lookahead. Its parameter count does not grow with the number of faces. On MSDWild, where most negative face-frames contain a visible competing speaker, CAST-ASD reduces competing-speaker false alarms by 32% relative to the causal per-face detector it extends, with larger gains as scene occupancy increases. It outperforms all causal baselines on both validation splits while using only one-tenth the inter-speaker parameters of a three-block fixed-roster convolution. Code and models are available at https://github.com/prathamg007/CAST-ASD.

[34] HumanPhys: A Human-Centric Physical-World Perception Benchmark for Real-World Visual Reasoning cs.CVPDF

Yiming Zhang, Enzhi Zhou, Junan Lin, Dulan Zhou, Weiqiang Wang

TL;DR: HumanPhys 用传感器标定、源数据元信息或渲染器状态直接生成标签，覆盖六类以人为中心的物理感知量；多种开源/闭源 MLLM 在六个维度上均面临挑战。

Details

Motivation: 评估多模态模型理解人物所处物理世界的能力，需要超越一般对象识别，直接测量深度、运动、视线、眨眼、受力变形、照明和相机位姿等量。

Result: HumanPhys 是以人为中心的真实物理世界感知基准，并由同一流程生成训练语料。问题询问深度、人与人距离、注视角、位移、速度、朝向—步态轨迹、眨眼次数和接触变形强度；渲染探针还测量从身体/面部线索推断照明，以及画面有/无人时的相机自运动。标签直接由传感器标定、源元数据或渲染器状态程序化生成，而非标注者手写或在基准图像上运行估计器。基准系统测空间状态、运动学、眨眼事件、变形、照明和相机位姿六个维度。多种 MLLM 实验显示这些能力在当前模型中仍困难，且六类能力仅部分可分离。

Insight: 直接元数据/渲染标签减少人工判断和基准图像估计器引入的标签噪声，适合评估具体物理量。但摘要未列题量、模型分数或维度间相关系数，且六种能力在模型中并非完全独立。

Abstract: We introduce HumanPhys, a human-centric benchmark for real-world physical perception, and a training corpus emitted by the same pipeline. Each question targets a physically grounded quantity related to people, including depth, inter-person distance, gaze angle, displacement, velocity, heading-step trajectory, blink count, and contact-deformation strength. The rendered probes additionally test illumination inference from body and face cues and camera ego-motion, both with and without a person in the frame. All labels are derived programmatically from sensor calibration, source metadata, or renderer state, rather than written by annotators or produced by estimators on the benchmark frames. HumanPhys systematically probes six dimensions of human-centric physical perception: spatial state, kinematics, blink events, deformation, illumination, and camera pose. To the best of our knowledge, no existing MLLM benchmark-and-training corpus jointly supervises this set of human-centered physical quantities under direct metadata- or renderer-derived labels. The six dimensions appear only partially dissociable across the tested MLLMs. We benchmark a range of open- and closed-source MLLMs on HumanPhys and find that all six capability dimensions remain challenging for current models. We hope HumanPhys will support future research on multimodal systems that perceive and reason about people in the physical world.

[35] Demiurge: Forging Navigable 4D Worlds via Autoregressive Video Rerendering cs.CVPDF

Hanxin Zhu, Jianqi Ma, Cong Wang, Zhihao Hu, Junyao Gao

TL;DR: Demiurge 用动态 4D 点云记忆和自回归视频重渲染保持不同相机轨迹一致，支持 180° 无视角重叠的大视差运动，并能将结果优化为低伪影 3DGS。

Details

Motivation: 视频扩散可从单段输入视频合成新视角片段，用于世界建模、具身智能和机器人仿真；但逐条独立处理相机轨迹，缺乏跨轨迹一致性，难以构建可导航的 4D 世界。

Result: Demiurge 是带动态 4D 点云记忆缓存的自回归视频重渲染管线，包含四项组件：以 4D 记忆为条件生成逼真新视角视频的外观扩散模型；与缓存几何对齐已知区域深度并补全缺失区域的记忆一致视频深度补全；提供高保真视觉引导和遮挡线索的记忆感知双流变形；以及改善感知质量的几何感知视频精化。实验报告在密集多视角轨迹上构建连贯时空网格；在相机剧烈运动下进行大视差重渲染，包括输入视图无重叠的 180° 旋转；输出几何一致，可直接优化为高质量 3D Gaussian Splatting 表示且伪影较少。作者称整体达到 SOTA，摘要未给量化指标。

Insight: 跨轨迹缓存几何和双流视觉引导共同约束新视角生成，使世界表示可跨相机路径复用，而不仅是逐轨迹生成好看的视频。摘要主要给能力案例和定性 SOTA 结论，未提供基准分数或计算成本。

Abstract: Recent advances in video diffusion enable novel-view video synthesis from a single input video, supporting world modeling, embodied intelligence, and robotic simulation. However, existing methods process camera trajectories independently, lacking the cross-trajectory consistency needed for coherent, navigable 4D worlds. We introduce Demiurge, an autoregressive video rerendering pipeline built on a dynamic 4D point-cloud memory cache. It promotes cross-trajectory consistency through four components: (1) an appearance video diffusion model that synthesizes photorealistic novel-view videos conditioned on the 4D memory; (2) a memory-consistent video depth completion model that aligns known-region depths with cached geometry and completes missing regions, mitigating scale ambiguity and discontinuities during memory updates; (3) memory-aware dual-stream warping that provides high-fidelity visual guidance and occlusion cues for novel-view synthesis; and (4) geometry-aware video refinement that enhances perceptual quality. Extensive experiments demonstrate state-of-the-art performance and three key capabilities: coherent spatiotemporal lattice construction over dense multi-view trajectories; large-parallax video rerendering under challenging camera motions, including 180-degree rotations with no input-view overlap; and geometry-faithful outputs that can be directly optimized into high-quality 3D Gaussian Splatting representations with minimal artifacts. Project page: https://hanxinzhu-lab.github.io/demiurge_project_page/.

[36] PURE: Refining Feed-forward 3D Reconstruction with Unlabeled Post-training cs.CVPDF

Yaolong Zhu, Peng Li, Jiahao Lu, Chengfeng Zhao, Mengfei Li

TL;DR: PURE 用无标签后训练将 2D 结构先验蒸馏到前馈 4D 重建架构，通过相对深度、仿射不变几何和特征对应位姿优化，改善动态野外视频的深度与相机估计。

Details

Motivation: 前馈 3D 重建模型依靠大量训练和先验取得进展，但在非刚性运动、遮挡和照明变化普遍存在的动态野外视频上仍脆弱。依赖昂贵深度及相机姿态标注限制了真实动态数据的规模和泛化。

Result: PURE 是无需真值监督的无标签后训练精化框架，将先进 2D 先验中的稳健结构知识蒸馏进 4D 前馈架构。方法结合相对深度约束、仿射不变几何蒸馏，以及基于特征对应的位姿精化，以稳定困难动态条件下的几何预测。在合成和真实动态基准上，对多个先进架构均报告一致改善，并缓解域偏移、释放基础模型在野外动态场景中的能力。摘要未给具体架构名称、数据集分数或提升幅度。

Insight: 无标签精化利用相对结构而非精确深度/位姿真值，降低动态真实数据监督成本。摘要只报告多个架构和基准上的一致提升，尚无数值来判断不同先验组件的独立贡献。

Abstract: Feed-forward 3D reconstruction models have achieved remarkable progress through large-scale training and strong learned priors. However, they remain fragile when applied to dynamic in-the-wild videos, where non-rigid motion, occlusions, and illumination changes are prevalent. A key limitation lies in their reliance on expensive depth and camera pose annotations, which hinders scaling to diverse real-world dynamic data and restricts generalization. To address this issue, we propose \textbf{PURE}, an unlabeled post-training refinement framework that improves depth and camera estimation without requiring ground-truth supervision. Instead of relying on annotations, PURE elegantly distills robust structural knowledge from state-of-the-art 2D priors into 4D feed-forward architectures. PURE leverages relative depth constraints, affine-invariant geometry distillation, and pose refinement based on feature correspondences to stabilize geometric predictions under challenging dynamic conditions. Extensive experiments on both synthetic and real-world dynamic benchmarks demonstrate that PURE delivers consistent improvements across multiple state-of-the-art architectures. Furthermore, our data-efficient paradigm effectively mitigates domain shifts, successfully unlocking the architectural limits of these base models in challenging, in-the-wild dynamic scenes.

[37] LightVAE: Towards Compact and Efficient Video Autoencoders cs.CVPDF

Shiqiao Gu, Yang Yong, Fengjun Zhong, Ruihao Gong, Chengtao Lv

TL;DR: LightVAE 将因果 Conv3D 重参数化为稠密空间响应加低秩时间修正，并结合宽度保持剪枝和教师特征对齐；Wan VAE 在 H100/RTX 5090 解码吞吐最高为基线的 2.24×/1.11×。

Details

Motivation: 潜扩散视频生成质量很高，但视频 VAE 编解码计算开销仍大。现有轻量化要么同时限制空间与时间计算，要么完全去掉时间建模，未利用预训练 Conv3D 中二者对压缩敏感度不同这一点。

Result: LightVAE 解析重参数化可适用的因果 Conv3D 层，将其精确改写为聚合空间响应与差分驱动时间修正。Temporal Low-Rank（TLR）保持聚合空间核稠密，仅对时间修正做逐层低秩近似；代数等价的折叠实现把算术减少转化为实际加速。另通过保持宽度的块剪枝减小深度，并以冻结教师后缀的探针引导特征对齐恢复压缩质量；方法也扩展到视频编码器。三种视频 VAE 主干、两种 GPU 平台上均提升解码速度。Wan VAE 的吞吐相较先进高效解码器，在 NVIDIA H100/RTX 5090 分别最高为 2.24×/1.11×，DAVIS PSNR 更高、LPIPS 更低。MiniMax-H3 上吞吐最高为原解码器 2.97×，整体 VBench-2.0 分数保留 99.51%。

Insight: 空间核保留、时间修正低秩压缩针对不同冗余结构，折叠实现确保理论算术节省变为实际加速。速度比较存在硬件差异，且 Wan 对比高效基线、MiniMax 对比原解码器，不能混用基线。

Abstract: Latent diffusion models enable high-quality video generation, but video VAE encoding and decoding remain computationally expensive. Existing methods use stage-specific lightweight operators, yet either constrain spatial and temporal computation jointly or remove temporal computation altogether, without distinguishing their compression sensitivities in pretrained Conv3D. To address these limitations, we present LightVAE, a compression framework that exploits redundancy in temporal modeling and network depth. LightVAE exactly reparameterizes eligible causal Conv3D layers into an aggregate spatial response and a difference-driven temporal correction. Based on their distinct compression sensitivities, temporal low-rank compression (TLR) keeps the aggregate spatial kernel dense and applies layer-specific low-rank approximation only to the temporal correction. An algebraically equivalent folded implementation converts the arithmetic savings into practical acceleration. LightVAE further uses width-preserving block pruning to reduce depth and probe-guided feature alignment, guided by feedback from a frozen teacher suffix, to recover quality after compression. The framework also extends to video encoders. Across three video VAE backbones and two GPU platforms, LightVAE consistently accelerates decoding. For the Wan VAE family, it achieves up to $2.24\times$ and $1.11\times$ the decoding throughput of state-of-the-art efficient VAE decoders on NVIDIA H100 and RTX~5090, respectively, with higher DAVIS PSNR and lower LPIPS. On MiniMax-H3, it achieves up to $2.97\times$ the original decoding throughput while retaining 99.51% of the original model’s overall VBench-2.0 score. Our code is available at https://github.com/ModelTC/LightVAE.

[38] Align-VAE: Aligning Attention and Geometry in VecSet VAEs for 3D Reconstruction cs.CVPDF

Huaizhi Qu, Mingxin Yang, Xintong Han, Fangneng Zhan, Chunchao Guo

TL;DR: Align-VAE 识别 VecSet VAE 几何对齐的四类问题，并以共享解码 token、距离偏置注意力、法线对齐和屏蔽 Poisson 正则修复；ABO 上 Chamfer distance 降低 26.6%。

Details

Motivation: VecSet VAE 用少量潜 token 表示 3D 形状，每个 token 源自采样表面位置，但这些潜变量是否真正与几何对应并不明确。

Result: 分析发现四类问题：类似 attention sink 的 token 吸引来自全形状的 query；解码注意响应在空间上周期重复；编码器和解码器 query 特征的法线信息不匹配；SDF 与表面法线内在关系缺少监督。Align-VAE 对应作出四项修改：共享可学习解码 token 作为全局计算空间；距离偏置交叉注意力引导检索邻近预测锚点；编码器 query 不含法线、但 key/value 保留法线，使两阶段访问法线信息对齐；受屏蔽 Poisson 启发的正则监督表面值和法线方向导数。潜变量维度不增加。在 ABO 数据集上，相较标准 VecSet VAE，Chamfer distance 降低 26.6%。

Insight: 问题不在潜变量数量，而在 query/latent 检索和场函数预测是否按几何关系对齐；该方法用注意力偏置和 SDF—法线关系监督提升几何一致性。报告的定量结果为 ABO 上 CD 降幅。

Abstract: Set-based variational autoencoders (VecSet VAEs) represent 3D shapes as compact sets of latent tokens. Although each token originates from a query at a sampled surface location, whether these latents truly align with geometry remains unclear. We investigate this alignment and identify four issues: (i) attention-sink-like tokens that attract queries from across the shape; (ii) spatially periodic repetitions in decoder attention responses; (iii) mismatched normal information in encoder and decoder query features; and (iv) missing supervision of the intrinsic relation between signed distance fields and surface normals. To address these issues, we introduce Align-VAE with four corresponding modifications: (i) a shared bank of learnable decoder tokens provides a workspace for global computation; (ii) distance-biased cross-attention guides retrieval toward nearby predicted spatial anchors; (iii) normal-free encoder queries align the two stages’ access to normal information while retaining normals in the encoder keys and values; and (iv) a screened Poisson-inspired regularizer supervises surface values and normal-direction derivatives. These changes align latent retrieval and field prediction with geometry without enlarging the per-shape code. On the ABO dataset, Align-VAE reduces Chamfer distance by 26.6% compared to the standard VecSet VAE.

[39] YOUDrive: Driving Style as a Composable Direction for End-to-End Autonomous Driving cs.CV | cs.AI | cs.ROPDF

Guantian Zheng, Lidong Yu, Tianyu Gao, Jiaxing Chen, Jiashu Li

TL;DR: YOUDrive 将教师驾驶风格转成可缩放、可组合的 task vector，在不重训主干时控制自驾策略风格；NAVSIM navtest 上六种 persona 的 assertiveness 随单一系数单调变化，PDMS 始终高于 0.90，最高 0.954。

Details

Motivation: 端到端驾驶策略可达到较好任务性能，却难在推理时直接控制加速、制动和跟车距离等驾驶风格。缺乏显式偏好条件的策略没有可组合的风格调节接口。

Result: YOUDrive 是视觉—语言—动作规划器，将不同教师的行为蒸馏成 task vector，即作用于规划器表示的可学习参数更新方向。推理时可缩放和组合这些方向，无需重训主干。条件化后，flow-matching 解码器建模多模态轨迹分布并生成运动学可行的备选方案。作者提出 Style Alignment Score（SAS），在安全和任务约束后衡量相对人类参考的运动学行为。NAVSIM navtest 上，单一系数可在六个教师 persona 间单调控制 assertiveness；所有所测单独及混合配置 PDMS 均高于 0.90，最高 0.954。控制场景上下文后，分析发现策略本身仍有系统性运动学差异，而评估的离散风格标签额外解释力有限。

Insight: 风格 task vector 提供连续可组合的推理时控制，不必在“安全/任务达标”和风格之间重训策略；SAS 通过安全/任务门控评估行为。结论限于 NAVSIM navtest 和六个教师 persona，离散标签的解释力并不强。

Abstract: End-to-end driving policies can achieve strong task performance while offering limited control over their driving style. Driving style describes how a vehicle accelerates, brakes, and maintains distance. Policies trained without explicit preference conditioning lack a direct mechanism for adjusting these behaviors at inference. We propose YOUDrive, a vision–language–action planner for continuous and composable driving-style control. Each teacher’s behavior is distilled into a task vector, a learned parameter update that induces a style direction in the planner’s representation. These directions can be scaled and combined at inference without retraining the backbone. Conditioned on these directions, a flow-matching decoder models a multimodal trajectory distribution and generates kinematically feasible motion alternatives. We introduce the Style Alignment Score (SAS) to describe human-referenced kinematic behavior after safety and task gating. On NAVSIM navtest, a single coefficient monotonically controls assertiveness across six teacher-derived personas, while PDMS remains above 0.90 across the tested individual and blended configurations, reaching 0.954. Further analyses reveal systematic policy-specific kinematics after controlling for scene context, whereas the evaluated discrete style labels provide limited additional explanatory power. Code, checkpoints, and the metric will be released.

[40] [Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning cs.CVPDF

Anton Nuzhdin, Andrei Gruzitski, Balázs Egyed, Aleksandr Raudvee

TL;DR: 该复现研究重做 CLIP 对抗微调提升可解释性的关键主张，并增加显著图引导的 Energy Pointing Game 正则；显式定位监督改善定位指标，但零样本准确率和对抗鲁棒性存在随权重变化的代价。

Details

Motivation: 先前 ICLR 2025 工作提出无监督对抗微调与范数正则可提升 CLIP 图像编码器视觉可解释性。作者尝试复现显著图质量、概念对齐、OOD 迁移及零样本准确率权衡等主张，并进一步测试直接空间监督。

Result: 研究复现无监督 AFT 方法及其范数正则，并提出 saliency-guided regularization 扩展：加入 Energy Pointing Game 损失，直接监督 Simple Gradient 显著图与目标物体的空间对齐。实验扫描不同显著图损失权重，显示显式定位监督可显著改善定位指标；在最低权重 w_s=0.2 时，零样本准确率和对抗鲁棒性下降较小，最强 w_s=0.6 时下降为中等。摘要未具体说明原论文哪些主张成功复现或各指标数值。

Insight: 直接对齐显著图可提升定位，但可解释性改善伴随零样本准确率和对抗鲁棒性折衷，且强弱依赖损失权重。摘要没有提供原方法复现完整度及数据集结果，不能将扩展结论等同于原论文主张全部复现。

Abstract: This paper presents a reproducibility study of ``Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning’’, published at ICLR 2025, which proposes an unsupervised adversarial fine-tuning (AFT) method with norm regularisation to enhance the visual interpretability of CLIP’s image encoder. We attempt to reproduce the key claims regarding improved saliency map quality, increased concept alignment, transferability to out-of-distribution datasets, and the trade-off with zero-shot accuracy. Beyond reproduction, we propose a saliency-guided regularisation extension that introduces an Energy Pointing Game loss, directly supervising the spatial alignment of Simple Gradient saliency maps with target objects. We evaluate our extension across a range of saliency-loss weights and show that explicit saliency supervision \rev{substantially} improves localisation metrics at a modest cost: the reduction in zero-shot accuracy and adversarial robustness is small at the lowest saliency-loss weight ($w_s=0.2$) and moderate at the strongest setting ($w_s=0.6$). We restructure the codebase for configurable, systematic experimentation.

[41] Getting Space Right: Agentic Rewards for Image Generation cs.CV | cs.AIPDF

Xiaofeng Tan, Jiaxin Lin, Haobo Yuan, Tiancheng Shen, Zihan Su

TL;DR: SpatialCritic 用带空间工具的智能体奖励模型验证文生图位置关系，配合能力图课程式 RL 组织训练；SpaRW-Eval 验证准确率 95.1%，8B 主干优于 320B GLM-5.3-Flash。

Details

Motivation: 空间文生图 RL 的提升常限于域内，转到域外会退化，暴露奖励模型不可靠和 rollout 缺少由简单到复杂进阶两项问题。

Result: 作者提出课程引导 RL 框架。SpatialCritic 是带空间感知 harness 的智能体奖励模型，结合可复用验证指南与专门视觉工具；针对性反事实扰动构造空间偏好数据，并进行智能体—harness 对齐，以提升工具选择和证据整合。课程学习把基础能力及其组合依赖组织为能力图，支持掌握度引导 rollout、定期复习和针对性补救。在 SpaRW-Eval 上，SpatialCritic 验证准确率为 95.1%，用 8B 主干超过参数量 40 倍的 320B GLM-5.3-Flash。其奖励用于 SD3.5-Medium 后，平均基线分由 78.4 升至 83.9；用于 BAGEL 后由 83.2 升至 86.5。

Insight: 空间奖励的可靠性和学习难度进阶同时影响 RL 泛化；SpatialCritic 以显式工具检查关系，能力图课程逐步组合技能。95.1%是奖励验证准确率，生成提升对应两个指定模型和基线分数。

Abstract: Reinforcement learning (RL) is promising for spatial text-to-image generation, but existing work largely focuses on policy optimization, assuming accurate rewards and effective online rollouts. We observe in-domain gains with out-of-domain degradation, exposing two limitations: unreliable spatial rewards in open-world settings and rollouts lacking basic-to-complex progression. We propose a curriculum-guided RL framework. First, we introduce SpatialCritic, an agentic reward model equipped with a spatial perception harness that combines reusable verification guidance with specialized vision tools. We then construct spatial preference datasets through targeted counterfactual perturbations and perform agent-harness alignment to improve tool selection and evidence integration efficiently. Second, curriculum learning organizes shared foundational capabilities and their compositional dependencies in a capability graph for mastery-guided rollouts, periodic review, and targeted remediation. On SpaRW-Eval, SpatialCritic achieves 95.1% verification accuracy, outperforming 320B GLM-5.3-Flash with a 40x smaller 8B backbone and raising average baseline scores from 78.4 to 83.9 on SD3.5-Medium and 83.2 to 86.5 on BAGEL. Project page: https://xiaofeng-tan.github.io/projects/spatialcritic/

[42] CAST: Real-Time Motion Capture for Any Skeleton Topology cs.CVPDF

Pu Pang, Zhongfei Qing, Ling-Hao Chen, Songnan Li, Hua Huang

TL;DR: CAST 面向目标骨架拓扑与视频主体不同的单目实时动捕，将学习的局部关节旋转与全局骨方向几何校正结合；因果 CAST-B 在单张 A100 上超过 51 FPS。

Details

Motivation: 从单目视频恢复动作时，目标骨架拓扑可能与视频人物不同；关节旋转误差会沿运动学层级耦合传播，使跨拓扑转移尤其脆弱。

Result: CAST 是拓扑感知动捕框架，以预测的全局骨方向对学习到的局部旋转进行几何校正。可观测性分析刻画哪些旋转分量能由出骨方向约束；基于此，GRACE 按父到子的顺序执行可靠性门控对齐，仅校正有足够信息的合成姿态，对不可约束分量则保留学习式旋转预测。运动预测器与校正模块使用由动作数据导出的监督联合训练，图像编码器冻结。TopoCap 评估显示，目标拓扑不同（包括未见结构）时动作恢复均有改善；整理的 Mixamo 基准和用户偏好研究进一步评估跨骨架迁移。因果版 CAST-B 在单张 NVIDIA A100 上超过 51 FPS，计时包括图像编码和动作预测。

Insight: 先依据骨方向可观测性判断哪些旋转可校正，再按层级修复，可减少误差沿骨架传播，同时保留不可观测分量的预测。摘要未给各拓扑精度数值，51 FPS 指 CAST-B 与所述计时范围。

Abstract: This paper studies the problem of recovering motion from monocular video, even when the target skeleton topology differs from the subject observed in the video. Although previous methods have made efforts to resolve this problem, they remain vulnerable to joint-rotation errors that are coupled across the skeleton and can propagate along its kinematic hierarchy. To mitigate these errors, we present CAST, a topology-aware motion capture framework that combines learned local rotations with geometric correction guided by predicted global bone directions. Our observability analysis characterizes which rotational components can be constrained by outgoing bone directions. Guided by this analysis, GRACE uses reliability-gated alignment to correct composed orientations in parent-to-child order, leaving unconstrained components to the learned rotation predictor. The motion predictor and correction module are trained jointly with motion-derived supervision, using a frozen image encoder. Experiments under the TopoCap evaluation protocol show that the proposed framework improves motion recovery across different target topologies, including unseen structures. Reference-based cross-skeleton evaluation on a curated Mixamo benchmark and a user preference study further assess transfer to different target structures. A causal CAST-B variant achieves over 51 FPS on a single NVIDIA A100 GPU, with timing covering image encoding and motion prediction. Code repository: https://github.com/fankewen/CAST.

[43] MORE: Learning Motion-Oriented Registers for Driving World Models and Planning cs.CVPDF

Yingdong Gu, Haoyu Wang, Shaocheng Yan, Zhenjun Zhao, Hao Han

TL;DR: MORE 在视频预训练阶段学习运动导向 registers，把紧凑状态接口同时用于驾驶世界模型预测和轨迹生成/评分；NAVSIM-v2 上 EPDMS 超过 SOTA 1.6 分，推理延迟为 DriveLaW 的十分之一。

Details

Motivation: 端到端驾驶世界模型借未来预测获得丰富自监督，但常继承稠密视觉编码器特征作为状态接口，运动条件建模与规划的适配被推迟到下游训练；高分辨率下稠密未来预测计算成本也高。

Result: MORE 是基于 V-JEPA 的运动导向 register 学习框架，将时间对齐的自车运动注入一小组 registers，并将其作为掩码 patch 预测的唯一上下文，使稠密 patch 监督通过紧凑状态瓶颈传递；分块 register 掩码目标进一步塑造 register 空间预测结构。下游训练时，同一 register 空间定义世界模型预测的未来状态，并作为轨迹生成和评分共享的场景记忆。NAVSIM-v2 上，MORE 的 EPDMS 超过先进方法 1.6 分，推理效率更高，延迟仅为 DriveLaW 的十分之一。

Insight: 运动状态接口在视频预训练期直接学习，避免下游世界模型再从稠密视觉特征中抽取紧凑运动表征。性能差距为 EPDMS 分数点，速度比较相对 DriveLaW；摘要未给绝对延迟或额外基准表现。

Abstract: By predicting future states, latent world models provide rich self-supervision for end-to-end driving, but often inherit dense features from pretrained visual encoders as their world-state interface. This creates a mismatch: visual pretraining produces information-rich, spatially dense features, but their adaptation for motion-conditioned world modeling and planning is deferred to downstream training. The dense grid also makes future prediction more computationally expensive as visual resolution increases. We therefore argue that video pretraining should directly learn state representations suited to downstream world modeling and planning, rather than relying on downstream training to adapt inherited visual features. We propose MORE, a V-JEPA framework that learns Motion-Oriented REgisters during video pretraining as a shared state interface for driving world models and planning. MORE injects temporally aligned ego motion into a small set of registers and uses them as the sole context for masked patch prediction, channeling dense patch-level supervision through a compact state bottleneck. A complementary blockwise register-masking objective further shapes predictive structure in the register space. During downstream training, the same register space defines the future states predicted by the world model and provides shared scene memory for trajectory generation and scoring. On NAVSIM-v2, MORE outperforms the state of the art by 1.6 in EPDMS, while achieving superior efficiency with only 1/10 the inference latency of DriveLaW. These results support learning the state interface for world modeling and planning directly during video pretraining.

[44] Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations cs.CVPDF

Jie Gui, Hang Tu, Yunfeng Dou, Wen Sha, Xiuquan Du

TL;DR: RADAR 用 BiomedCLIP 语义锚定、MedSAM2 掩码解码和 LoRA 适配进行文本引导医学分割，加入跨层残差聚合、稠密对齐表示和复合多模态提示；仅需 0.74M 可训练参数。

Details

Motivation: 自然语言指定分割目标可提供灵活语义条件，但将高层文本概念稳定落到像素级提示仍困难，尤其在低对比病灶、边界模糊和相似解剖背景中。

Result: RADAR 参数高效地组合 BiomedCLIP 医学语义锚定能力与 MedSAM2 提示式掩码解码。Guided Residual Depth Aggregation 在语义引导下自适应汇聚跨层特征，缓解文本注入过晚及深层语义稀释；Structured Dense Representation Alignment 使用冻结 BiomedCLIP 构建裁剪级目标、边界和上下文锚点，提升稠密特征局部区分度及自动提示可靠性；Multi-modal Prompt Generation 把跨模态语义响应转成组合文本、点和掩码提示供 MedSAM2 解码。在三个公开医学分割数据集上，RADAR 仅用 0.74M 可训练参数达到有竞争力结果；摘要未给各数据集分数。

Insight: 将语义定位和掩码生成明确拆分，并用边界/上下文锚点增强局部表示，适合以少量参数迁移医学分割。结论是三数据集竞争力而非明确 SOTA，摘要没有报告计算/标注成本对比。

Abstract: Specifying segmentation targets with natural language provides flexible semantic conditioning for medical image segmentation. However, existing methods still struggle to reliably ground high-level textual semantics into accurate pixel-level spatial prompts, especially under low-contrast lesions, ambiguous boundaries, and visually similar anatomical contexts. In this paper, we propose RADAR (Residual Aggregation and Dense Alignment Representations), a parameter-efficient framework for text-guided medical image segmentation. RADAR explicitly decouples and synergizes the medical semantic anchoring capability of BiomedCLIP with the prompt-driven mask decoding ability of MedSAM2, while enabling low-cost task adaptation via LoRA. Specifically, Guided Residual Depth Aggregation adaptively aggregates cross-layer features under semantic guidance to alleviate delayed textual injection and deep semantic dilution. Structured Dense Representation Alignment leverages a frozen BiomedCLIP to construct crop-level target, boundary, and context anchors, improving the local discriminability of dense features and the reliability of automatic prompts. Multi-modal Prompt Generation further converts cross-modal semantic responses into composite text, point, and mask prompts for MedSAM2 decoding. Experiments on three public medical image segmentation datasets show that RADAR achieves competitive performance with only 0.74M trainable parameters, demonstrating its effectiveness in parameter-efficient adaptation, semantic grounding, and dense representation enhancement. The source code is available at https://github.com/JGuiidev/RADAR.

[45] CatPortrait: Taming Video Generative Models for Portrait Animation via In-context Conditioning cs.CVPDF

Ziqi Ni, Zhou Fang, Shijian Wang, Rui Li, Yi Zhou

TL;DR: CatPortrait 将结构化 3D 面部指导空间拼接到预训练图像到视频模型输入中，无需改主干或增设条件模块即可统一支持视频驱动与音频驱动肖像动画。

Details

Motivation: 视频生成模型已有高质量、时间连贯的视频生成能力，但将其适配为可控肖像动画仍复杂；既有方法常改架构或增加条件模块，降低对预训练模型的兼容性。

Result: CatPortrait 用 3D 面部引导作为 in-context 条件，在输入空间进行空间拼接，保留原视频生成主干及其预训练先验。统一框架支持视频驱动和音频驱动两类肖像动画。大量实验显示，方法在两种设置下达到有竞争力或更优结果；摘要未给具体数据集、质量指标或分数。

Insight: 把 3D 脸部结构化指导作为输入上下文，避免修改视频生成骨干，降低适配复杂度并兼容预训练模型。摘要只报告两种任务下的综合比较结论，没有量化结果。

Abstract: Recent video generative models have shown strong capability in synthesizing high-quality and temporally coherent videos. Yet, adapting them for controllable portrait animation remains nontrivial, as existing methods typically rely on architectural changes or extra conditioning modules, increasing complexity and reducing compatibility with pretrained models. We present CatPortrait, a simple and effective framework that enables controllable portrait animation via 3D face guidance as in-context conditioning. CatPortrait preserves the original video generation backbone and injects structured 3D facial guidance into the input space through spatial concatenation, allowing pretrained image-to-video models to animate portraits without sacrificing their pretrained priors. CatPortrait supports both video-driven and audio-driven portrait animation within a unified framework. Extensive experiments demonstrate that CatPortrait achieves competitive or superior results across both settings.

[46] Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning cs.CV | cs.LGPDF

Thanh Nguyen, Hee Suk Yoon, Tung Minh Luu, Eunseop Yoon, Chang D. Yoo

TL;DR: 该 RL 框架用 Object Consistency Score 衡量物体描述的正确性与覆盖率，并加 KL 引导和轻量 PPO 微调冻结骨干；InstructBLIP 上对象幻觉在 COCO/Visual Genome 分别最多降低 41%/29%。

Details

Motivation: LVLM 图像描述可能提及图中不存在的对象。已有数据增强监督或人类反馈 RL 方法有潜力，但依赖昂贵的人力，难以规模化；仅用 NLP 指标做 RL 又容易奖励破解、反馈稀疏和训练成本高。

Result: 框架以 NLP 指标驱动图像描述强化学习，并提出 Object Consistency Score（OCS）同时评估对象描述正确性和覆盖率；KL 正则提供更稠密 token 级引导，轻量 PPO 变体仅微调适配器、冻结 LVLM 主干，以提高效率和稳定性。基于 InstructBLIP 的广泛实验显示，COCO 上对象幻觉最高降低 41%，Visual Genome 上最高降低 29%。加入 BERTScore、METEOR 等互补评价指标还能改善描述质量。

Insight: 结构化对象一致性奖励试图弥补单纯语言质量指标不能识别虚构实体的问题，同时降低人类标注依赖。41%和29%是数据集上的最高幻觉降幅；摘要未给基线水平、检测指标或是否伴随其他描述质量折衷。

Abstract: Large Vision-Language Models (LVLMs) have excelled in joint visual and language understanding, particularly in image captioning. However, LVLMs remain prone to object hallucination, describing non-existent entities in the given image. Previous remedies, including dataset-enhanced supervised training and reinforcement learning from human feedback, show promise but rely heavily on costly human involvement, limiting scalability. To address this, we propose a reinforcement learning framework to mitigate object hallucination in LVLM-based image captioning, driven entirely by natural language processing (NLP) metrics. While optimizing reinforcement learning solely on NLP metrics offers scalability, it also introduces pitfalls: reward hacking, sparse feedback, and costly training. Our framework sidesteps these by introducing an Object Consistency Score (OCS) capturing both correctness and coverage of objects, incorporating KL-regularization for denser token-level guidance, and designing a lightweight PPO variant that fine-tunes adapters atop a frozen LVLM backbone, achieving efficiency and stability. Extensive experiments on the InstructBLIP baseline demonstrate the effectiveness and generalizability of our framework, achieving a substantial hallucination reduction of up to 41% on the COCO dataset and 29% on the Visual Genome. Furthermore, the framework can enhance caption quality by incorporating complementary evaluation metrics such as BERTScore and METEOR. These results reveal a broader insight in multimodal learning: LVLMs can effectively self-correct hallucinations using metric feedback, motivating deeper exploration into scalable reinforcement frameworks that significantly reduce human effort.

[47] Retraining-Free Depth Pruning of Vision Transformers: A Cross Domain Study of Layer Redundancy cs.CVPDF

Alessandro Viespoli, Loris Nanni

TL;DR: 该研究基于层输入输出相似度分析 ViT 深度冗余，发现 DINOv2 可免重训删除 20% block；ImageNet-1K 吞吐提升 1.24×、每图能耗下降 20%，相对精度损失低于 1%。

Details

Motivation: Transformer 部署受到计算、显存和存储成本限制。相比 LLM，视觉 Transformer 层冗余尚缺少系统分析及跨架构的免重训剪枝评估。

Result: 作者用层输入与输出的余弦相似度衡量变换对表示的影响，扩展相似度冗余分析到多种视觉模型。DINOv2、SwinV2 等深层 ViT 同样呈现显著冗余。不同剪枝策略体现推理速度与精度退化之间的折衷；在 DINOv2 上免重训去掉 20% block，ImageNet-1K 吞吐提升 1.24×、每图能耗减少 20%，相对精度损失低于 1%。作者还提出模型选择流程，自动为给定数据集确定合适剪枝配置；跨架构平均 FLOPs 减少 13%–27%，平均精度变化介于 -0.3 到 +1.0 分。

Insight: 中度深度剪枝在部分视觉主干上可直接利用层级冗余，无需重训；实际压缩配置仍应按目标数据集和精度/速度折衷选择。吞吐/能耗数字专指 DINOv2 的 20% block 剪枝和 ImageNet-1K。

Abstract: Transformer architectures have achieved strong results in natural language processing and computer vision, but their deployment is limited by high computational, memory and storage costs, motivating research into model compression techniques, such as pruning, while preserving performance. Prior work has shown substantial redundancy in large language models (LLMs), and redundancy in vision transformers has also been investigated, but it remains less systematically analyzed than in LLMs. In this work, we extend similarity-based redundancy analysis to vision models by measuring cosine similarity between layer inputs and outputs to identify transformations with minimal impact. We study redundancy patterns across several transformer architectures and find that deep vision transformers such as DINOv2 and SwinV2, like language models, exhibit significant redundancy. We then evaluate how this redundancy can be exploited to improve efficiency, analyzing the trade-off between inference speed and performance degradation under different pruning strategies. Our results show that moderate pruning can yield substantial acceleration with limited impact on accuracy, and that 20% of the DINOv2 blocks can be removed without any retraining, for a 1.24x throughput gain and a 20% lower energy per image at a relative accuracy loss below 1% on ImageNet-1K. We further propose a model selection pipeline that automatically identifies the best pruned configuration for a given dataset, reducing FLOPs by 13% to 27% on average across architectures with an average accuracy change between -0.3 and +1.0 points.

[48] CowCare: A Unified Visual Framework for Viewpoint-Robust Cattle Tracking and Spatiotemporal Heat-Stress Analysis cs.CVPDF

Xin Liu, Weiran Li, Minghao Zhu, Mina Han, Zhenbo Li

TL;DR: CowCare 将跨视角牛只跟踪与时空热应激识别整合；MVCow 上全局跟踪 HOTA/IDF1 为 83.48%/87.72%，HSCow 上热应激 Top-1 准确率为 87.72%。

Details

Motivation: 精准畜牧需要连续个体监测以支持奶牛福利和早期热应激预警。多摄像头遮挡、视角变化、相机振动及农业数据有限会造成身份断裂，并掩盖快速呼吸、侧腹起伏等细微热应激线索。

Result: CowCare 将视角鲁棒跟踪和时空热应激识别整合。CowTrack 的 VDA-Net 在严重遮挡下聚合局部纹理而非依赖静态全局表征，VDT 抑制短暂视角差异，并以因果在线自监督持续适应动态农场场景。后续 TSDF-Net 做细粒度视频理解，其中 TSAM 抑制高频抖动，冻结预训练扩散模型语义作为 DSP 先验以增强细微行为信号并缓解过拟合。两个真实农场基准上，CowTrack 在 MVCow 的全局多目标跟踪 HOTA 为 83.48%、IDF1 为 87.72%；TSDF-Net 在 HSCow 的热应激 Top-1 为 87.72%。

Insight: 跟踪的跨视角身份连续性为后续行为诊断提供稳定轨迹，时间稳定和语义先验用于保留细微热应激模式。摘要称其为端到端监测框架，但汇报指标分别来自两个基准及两个子任务。

Abstract: Reliable individual monitoring is essential for dairy cattle welfare and early heat stress detection in Precision Livestock Farming (PLF). However, translating raw multi-camera videos into automated diagnostics remains challenging due to disconnected tracking and recognition pipelines and complex physical environments. Dense mutual occlusions and cross-camera viewpoint variations fragment identities, while environmental noise, including camera vibrations, and limited agricultural data obscure subtle heat stress cues such as rapid breathing and flank undulations. To bridge low-level visual features and high-level behavioral diagnostics, we propose CowCare, a unified framework combining viewpoint-robust tracking with spatiotemporal heat stress recognition. Its tracking module, CowTrack, maintains cross-camera identity consistency through a View-Decoupled Aggregation Network (VDA-Net). Instead of relying on static global representations, VDA-Net aggregates localized textures under heavy occlusion and uses a Viewpoint Decoupling Transformer (VDT) to suppress transient viewpoint variations. CowTrack further incorporates a causal online self-supervision mechanism for continuous adaptation to dynamic farm scenes. The resulting tracklets are subsequently processed by TSDF-Net, a fine-grained video understanding network. Its Temporal Stabilization and Aggregation Module (TSAM) suppresses high-frequency jitter, while a Diffusion Semantic Prior (DSP) uses frozen pre-trained diffusion semantics to enhance subtle behavioral signals and mitigate overfitting. Extensive evaluations on two real-world farm benchmarks demonstrate the framework’s effectiveness: CowTrack achieves 83.48% HOTA and 87.72% IDF1 for global multi-object tracking on MVCow, and TSDF-Net achieves 87.72% Top-1 accuracy on HSCow. These results establish CowCare as a robust end-to-end framework for automated livestock health monitoring.

[49] What Really Works for Universal Multimodal Embeddings? A Controlled Study of Contrastive Training Strategies cs.CVPDF

Duo Li, Haojun Pan, Zuhao Yang, Xintong Wang, Chengfu Huo

TL;DR: 受控 MMEB 研究发现，多数复杂对比训练策略相对标准对齐基线收益有限甚至退化；结合 hard-negative、false-negative 缓解和 hard-batch 构造时，Qwen2-VL-2B 平均提升 6.0%。

Details

Motivation: 多模态嵌入对比学习已有多种负例选择、batch 构造和梯度调制策略，但在统一可复现条件下这些复杂方法是否稳定带来下游收益尚不清楚。

Result: 作者在 MMEB 统一框架下系统评估代表性对比策略。直接结果显示，多数策略相较标准对齐基线只有边际提升，部分还降低总体表现。为定位原因，作者提出 MLLM-as-a-Judge 诊断，将 top-1 检索错误分为标注正例、未标注的有效匹配（false negative）及困难/简单负例错误。分析发现：提高高相似负例梯度权重会损害泛化；正负相似度分布拉得更开不必然提升检索；单独缓解 false negative 或构造 hard batch 收益有限，但联合 hard-negative mining、false-negative 缓解与 hard-batch 构造效果最佳，Qwen2-VL-2B 在 MMEB 平均提升 6.0%。

Insight: 对比损失或相似度分布看似更理想，不代表检索能力更好；应联合诊断负例标注和批内难度。6.0%是指定模型及 MMEB 上组合策略的平均提升，其他单项策略大多收益有限。

Abstract: Contrastive learning has become the standard paradigm for multimodal representation learning, aligning semantically matched pairs while separating mismatched ones in a shared embedding space. Although various contrastive training strategies have been proposed across negative selection, batch construction, and gradient modulation, their actual effectiveness under controlled, reproducible conditions remains poorly understood. This motivates a fundamental question: Do sophisticated contrastive strategies consistently bring meaningful downstream gains in multimodal embedding learning? To address this, we systematically evaluate representative strategies under a unified benchmarking framework on MMEB. Surprisingly, direct evaluations reveal that most existing strategies offer marginal gains over a standard aligned baseline, with several even degrading overall performance. To uncover the underlying failure modes, we introduce an MLLM-as-a-Judge diagnostic framework that dissects top-1 retrieval behaviors into labeled positives, false negatives (unlabeled valid matches), and hard/easy negative errors. Our systematic analysis yields three key insights: (1) Gradient distortion: Upweighting gradients from high-similarity negatives consistently hurts downstream generalization; (2) Distribution illusion: Wider separation between positive and negative similarity distributions does not necessarily translate to better retrieval performance; and (3) Synergistic training: While false-negative mitigation or hard-batch construction yields limited gains in isolation, integrating hard-negative mining, false-negative mitigation, and hard-batch construction achieves optimal performance, delivering an average improvement of 6.0% on MMEB with Qwen2-VL-2B. Overall, our findings demystify popular contrastive strategies and provide actionable recipes for training universal multimodal embedding models.

[50] Use Of Convolutional Neural Networks In Detection Of Branch Dieback Disease In Pequi Trees cs.CV | cs.CE | cs.DBPDF

Ronilson Rodrigues Pinho, Elua Ramos Coutinho, Nilza de Lima Pereira Sales, Sildenir Aalves Ribeiro, Sergio Manuel Serra da Cruz

TL;DR: 研究比较 AlexNet、Inception、LeNet、VGG16、MobileNet 等 CNN 对巴西 pequi 树枯梢病图像进行健康/病株分类；摘要说明研究目标和方法，但没有提供哪种架构最佳或定量结果。

Details

Motivation: 枯梢病会严重影响 pequi 树发育和果实产量，造成经济和生态损失。研究希望评估计算机视觉是否可基于叶、枝和花序图像识别病征。

Result: 作者比较 AlexNet、Inception、LeNet、VGG16 和 MobileNet 等 CNN 架构，利用 Python 视觉与机器学习库分析叶片、枝条和花序图像中的纹理与颜色变化，区分健康植株和感染枯梢病的植株。摘要表示将识别最有效架构，但没有给出数据规模、测试划分、准确率或最终胜出模型，因而不能补述比较结果。

Insight: 该条目摘要描述了要评估和比较的模型与病害图像类型，但没有提供实验结果；解析应保留缺失，不把研究目标误写成已验证性能。

Abstract: This study aims to evaluate and compare convolutional neural network (CNN) architectures, specifically AlexNet, Inception, LeNet, VGG16, and MobileNet, for diagnosing branch dieback (seca-dos-ponteiros) in pequi trees (Caryocar brasiliense). This disease represents a severe threat to tree development and fruit yield, causing substantial economic and ecological losses. The methodology implements computer vision techniques using Python libraries, including Scikit-learn, NumPy, TensorFlow, and Matplotlib, to process and evaluate visual data. The CNN models analyze images of leaves, branches, and flower spikes to detect pathological patterns, such as alterations in texture and coloration. Ultimately, this comparative analysis identifies the most effective architecture for accurately distinguishing healthy plants from diseased ones.

[51] BrainAid: Efficient Brain Tumor Detection Using Lightweight CNNs and Explainable AI for Resource-Constrained Environments cs.CVPDF

Sai Sriram Gonthina, Ross Gore, Bikash Chandra Singh, Pronaya Bhattacharya, Sachin Shetty

TL;DR: BrainAid 将 CNN/ViT 用于 MRI 四分类脑肿瘤识别，并部署到 Raspberry Pi；公开数据集上报告 precision、F1 和 accuracy 均为 98%。

Details

Motivation: 人工 MRI 评估可能因专业人员标准不同而不一致且耗时，深度模型又常有较高计算要求。作者希望探索可在资源有限环境部署的快速脑肿瘤图像分类模型。

Result: BrainAid 使用 CNN 和 Vision Transformer 对 MRI 做胶质瘤、脑膜瘤、垂体瘤和无肿瘤四类分类，并与 VGG16、Inception-V3、ResNet-50 等标准模型比较，另集成 XAI 提高解释性。作者报告公开数据集 precision、F1 和 accuracy 均为 98%，并将系统部署到 Raspberry Pi，展示边缘设备运行可行性。摘要未给数据集名称、患者级切分方式、外部验证或解释性评估指标。

Insight: 报告结果表明轻量边缘部署有可行性，但公开数据集上的 98% 不能直接代表临床人群表现；缺少患者级切分和外部验证信息，摘要也未证明其可独立诊断。

Abstract: A brain tumor is an unusual growth of cells in the brain that can be cancerous or not, affecting normal brain function and causing serious risk to health. Existing solutions for brain tumor classification, such as expert manual evaluations of magnetic resonance imaging (MRI) scans, are generally inconsistent as different professionals have different criteria and can be resource-intensive, leading to delayed diagnoses. Additionally, while deep learning models have demonstrated potential, they typically require significant computational resources, motivating the need for more efficient and accessible diagnostic tools that can be deployed in real-time healthcare settings. To address these limitations, this study presents BrainAid, which makes efficient use of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for effective classification of four-class classification of glioma, meningioma, pituitary tumor, and no-tumor MRI scans, outperforming all the standard models such as VGG16, Inception-V3, ResNet-50 and so, while integrating Explainable AI (XAI) to improve interpretability. The BrainAid is trained to classify tumors with a high precision of 98%, achieving an F1 score of 98% and an accuracy of 98% using publicly available dataset. The model is deployed on a Raspberry Pi, demonstrating efficiency in low-resource environments. This system shows the feasibility of real-time, edge-based tumor classification, providing quick diagnostic support for medical professionals and enhancing patient care in resource-limited healthcare facilities and underserved areas.

[52] TIDE: Trajectory-Informed Skill Distillation for Efficient Text-to-Video Prompt Optimization cs.CVPDF

Chengxing Zhou, Yuliang Huang, Jingwei Song

TL;DR: TIDE 将昂贵的文生视频提示搜索蒸馏为可复用、可解释的规则式提示改写技能；只用 100 个提示、14 小时演化后，效果可比搜索方法，并能迁移到未见视频生成模型。

Details

Motivation: 文生视频模型对提示细节高度敏感，常见迭代优化需在推理时反复生成视频并调用视觉语言模型评估，计算成本高。

Result: TIDE 将提示改写表示为按来源条件化的规则模块，每个模块针对不同语义维度。技能冻结后，按需启用规则，一次 LLM 调用同时生成正向和负向提示，随后只进行一次视频生成，无需测试时视觉搜索或微调生成主干。技能仅用 100 个提示、14 小时演化，提示优化效果可与计算密集的搜索方法相当；迁移到训练未见的视频生成模型时，TIDE 还超过专为这些模型设计或训练的既有方法。

Insight: 把优化搜索过程转成可解释、可迁移的规则技能，可将重复搜索成本前置到有限演化阶段。迁移结论来自未见生成模型，摘要未给质量指标或生成主干清单。

Abstract: Text-to-video (T2V) models are highly sensitive to prompt specificity, while iterative prompt optimization typically improves generation quality only through repeated video generation and vision-language model evaluation at inference time. We propose TIDE, a framework that distills this costly search process into an interpretable and reusable prompt-rewriting Skill. TIDE formulates prompt rewriting as a set of source-conditioned rule modules, each targeting a distinct semantic dimension. Once distilled, the frozen Skill activates the relevant rules and generates both positive and negative prompts in a single LLM call, followed by only one T2V generation, without test-time visual search or backbone fine-tuning. Evolved using only 100 prompts over 14 hours, TIDE achieves prompt-optimization performance comparable to computationally expensive search-based methods. Moreover, when transferred to previously unseen video-generation models, TIDE outperforms existing approaches specifically designed or trained for those models. Code is available at https://github.com/xiningin/TIDE_T2V.

[53] ExeWorldBench: Benchmarking Code-Based Executable 3D World Modeling cs.CVPDF

Xinge Peng, Zhen Fang, Ruibing Wang, Yongxuan Lv, Haodong Li

TL;DR: ExeWorldBench 评测代码智能体能否从视觉观察理解机制并构建可执行 3D 世界，覆盖单机制复现、组合系统和目标驱动合成；前沿模型实验显示此类可执行世界建模仍有明显局限。

Details

Motivation: 代码智能体正从复现世界外观转向构建能体现物理交互和因果规律的可执行世界。现有基准往往只评估物理推理、视觉保真或行为正确性，未充分评估理解与代码实现的结合。

Result: ExeWorldBench 被作者称为首个全面评估代码式 3D 世界构造中 world understanding 与 executable realization 的基准。它包含三类递进任务：从视觉观察复现单个机制的 Mechanism Executability；包含多阶段空间、机械和因果依赖的 Compositional Executability；以及从高层目标和物理需求构建可执行世界的 Intent-Driven Executable World Synthesis。对前沿模型的实验揭示其可执行世界建模存在明显限制，为发展物理 grounding 更强、因果连贯的世界构造系统提供分析依据；摘要未列模型、任务分数或失败类别。

Insight: 基准评估的不只是生成画面或回答物理问题，而是能否把理解变成实际运行、由物理交互产生行为的代码系统。摘要只报告总体能力限制，没有量化排名。

Abstract: Coding agents are increasingly evolving beyond reproducing what a world looks like toward constructing executable worlds that capture how a world works. This shift calls for executable world modeling, which integrates two core capabilities: world understanding, reasoning about the underlying structural, mechanical, and causal principles of a world, and executable realization, translating that understanding into a code-based system whose behavior emerges from physical interactions. However, existing benchmarks provide only fragmented evaluations centered on physical reasoning, visual fidelity, or behavioral correctness, leaving the integration of these capabilities insufficiently assessed. To bridge this gap, we introduce ExeWorldBench, to our knowledge the first benchmark to comprehensively evaluate world understanding and executable realization through code-based 3D world construction. ExeWorldBench comprises three progressively challenging tasks: Mechanism Executability, which evaluates the realization of individual mechanisms from visual observations; Compositional Executability, which extends to multi-stage systems with spatial, mechanical, and causal dependencies; and Intent-Driven Executable World Synthesis, which requires constructing executable worlds from high-level goals and physical requirements. Extensive experiments on frontier models reveal substantial limitations in executable world modeling, offering insights into the development of more physically grounded and causally coherent world construction systems.

[54] GeoBridge: Evolving Transferable Decoder Topologies for Geometric Vision cs.CV | cs.AIPDF

Rui Wang, Haoyu Zhang, Ran Cheng

TL;DR: GeoBridge 用可搜索的阶段内残差/串行/并行解码器拓扑和无训练梯度敏感性代理进行多目标演化；在 NYU Depth V2/KITTI 的深度任务优于四种适配基线，同一拓扑在 RealEstate10K 达 27.11 dB PSNR。

Details

Motivation: 深度估计与 Gaussian 重建都需多尺度特征处理，但预测接口和训练目标不同。是否能一次优化解码器内部计算拓扑，并在不同几何视觉任务中复用，仍是结构设计问题；只搜索原语配置会把变换组织方式固定。

Result: GeoBridge 将阶段内拓扑设为可搜索对象，在固定多尺度骨架接口内组合残差、串行和并行原语。函数树编码支持阶段感知交叉和层级变异；梯度敏感性对齐代理无需训练候选即可评估，再与计算量、参数量做多目标演化。选定拓扑在目标任务中实例化并针对专用预测接口训练，单次源搜索约需 12 GPU 小时。GeoBridgeNet-A3 在统一编码器及训练协议下，于 NYU Depth V2、KITTI 的深度准确率超过四种适配解码器基线；同一拓扑在 RealEstate10K Gaussian 新视角合成达到 27.11 dB PSNR，其精化表示还支持 LERF 开放词汇分割。

Insight: 解码器拓扑可作为跨几何任务复用的设计对象，而不只是固定结构中的超参数；源搜索使用梯度代理以避免逐候选训练。比较控制了编码器/训练协议，跨任务结果涵盖深度、Gaussian 渲染和开放词汇分割。

Abstract: Depth estimation and Gaussian reconstruction share a need for multi-scale feature processing, yet use different prediction interfaces and training objectives. Whether a decoder’s internal computation can be optimized once and reused across these systems remains a structural design question. Search over primitive configurations alone leaves the organization of transformations fixed. To address this gap, we propose GeoBridge, an evolutionary framework for transferable decoder topologies. First, GeoBridge makes stage-internal topology searchable through residual, sequential, and parallel composition of configurable primitives. Shape-preserving composition retains the stage interfaces of a fixed multi-scale scaffold. Second, a functional-tree encoding supports stage-aware crossover and hierarchical mutation, while a gradient sensitivity-alignment proxy evaluates candidates without training. Multi-objective evolution balances this proxy with computation and parameter cost. Third, selected topologies are instantiated and trained in target systems with task-specific prediction interfaces. A single source search takes approximately 12 GPU hours. GeoBridgeNet-A3 achieves the best depth accuracy among four adapted decoder baselines under a common encoder and training protocol on NYU Depth V2 and KITTI. The same topology reaches 27.11 dB PSNR in Gaussian novel-view synthesis on RealEstate10K, and its refined representations support open-vocabulary segmentation on LERF. These results support decoder topology as a reusable design object across the evaluated geometric vision systems. Code and pretrained weights are available at https://github.com/EMI-Group/GeoBridge-NAS.

[55] TempoKV: Predictive KV Cache Quantization for Autoregressive Video Generation cs.CVPDF

Jiahui Hu, Weile Mo, Peiliang Cai, Jiacheng Liu, Jinkui Ren

TL;DR: TempoKV 利用视频 KV 的时间可预测性差异，分别以历史预测 key、用当前 key 和短历史预测 value，只低精度保存残差；Self/Causal Forcing/LongCat 上最高实现 6.71× KV 压缩。

Details

Motivation: 自回归视频扩散支持长视频流式生成，但 KV 缓存随时间快速增长。现有语言模型 KV 量化方法不易迁移到视频，因为激活统计不同且存在强时空冗余。

Result: 作者观察到视频 KV 随时间总体可预测且 key/value 依赖不同：key 高度可由时间历史预测，value 主要依赖当前 key，并需要短段前序 value。训练免费框架分别按这些依赖预测 key 和 value，只将预测残差低精度存储。两个轻量预测器通过少量 BF16 样本离线线性回归校准一次，之后在不同提示和生成任务间固定。Self-Forcing、Causal Forcing 和 LongCat-Video-13B 实验中，在相当或更高压缩率下，生成质量持续优于既有 KV 量化方法；最高压缩倍率为 6.71×，且总体质量—效率折衷最佳。

Insight: 视频 key 和 value 的时间依赖不对称，分别建模比直接移植通用 LLM 缓存量化更合适。6.71×是最高压缩率，质量结论针对三种生成框架，摘要未给具体指标或显存/速度数字。

Abstract: Autoregressive (AR) video diffusion models enable streaming generation of long videos but suffer from significant memory overhead due to the rapidly growing KV cache. Existing KV cache quantization methods are primarily designed for LLMs and transfer poorly to video models, where activations exhibit different statistics and strong spatiotemporal redundancy. We observe that the video KV cache is largely predictable over time, with a clear asymmetry. Keys are highly predictable from their temporal history, while values depend primarily on the current key and need a short history of preceding values. Based on these observations, we propose \textbf{\method}, a training-free predictive KV cache quantization framework for AR video generation. \method predicts keys and values according to their distinct dependencies and stores only the residuals at low precision. Both predictors are lightweight, calibrated once offline by linear regression on a small set of BF16 samples, and remain fixed across prompts and generation runs. Experiments on Self-Forcing, Causal Forcing, and LongCat-Video-13B show that \method consistently achieves higher generation quality than existing KV cache quantization methods at comparable or higher compression ratios, reaching up to $6.71\times$ KV cache compression with the best overall quality–efficiency tradeoff. Our code will be released on GitHub.

[56] BaT-SAM: Introducing Inductive Bias into SAM via a Mixture of Convolutional Experts for Brain Tumor Segmentation cs.CVPDF

Sumiaya Jannat, Md Mahedi Hasan, Md Shamimul Islam, Md Nur Hossain

TL;DR: BaT-SAM 为冻结 SAM ViT 加入稀疏门控多尺度卷积专家，并联合训练自动 box 回归器，无需人工提示；BRISC2025 加权 mIoU 0.832，Figshare/LGG Dice 为 0.702/0.784。

Details

Motivation: MRI 脑肿瘤分割对诊断和治疗规划重要，但人工勾画耗时；任务专用方法泛化有限，SAM 类方案推理时通常依赖用户提示。

Result: BaT-SAM 用 Conv-LoRA 将局部卷积归纳偏置注入冻结 SAM ViT 编码器，稀疏门控混合多尺度卷积专家处理图像特征，补足普通 ViT 平移等变性不足；可训练参数少于 5M，不到完整模型 1%。另加入轻量 self-prompting box 回归器，从图像嵌入直接预测肿瘤位置，并与适配器通过 scheduled sampling 联合训练，取消用户框提示。在 BRISC2025 上加权 mean IoU 为 0.832，超过论文所比先进基准；跨数据集 Figshare 和 LGG MRI Segmentation 的 Dice 分别为 0.702 和 0.784。

Insight: 卷积专家提供局部归纳偏置，自动框回归则去除交互提示依赖；少于 5M 参数实现参数高效适配。分数来自三个不同数据集/指标，不能直接横向比较；摘要未报告患者级外部临床验证。

Abstract: Accurate brain tumor segmentation from magnetic resonance imaging (MRI) is essential for diagnosis and treatment planning; however, manual delineation is labor-intensive and time-consuming. Existing approaches are either task-specific with limited generalization or rely on user-provided prompts during inference. We present BaT-SAM, a parameter-efficient adaptation framework that enables fully automatic brain tumor segmentation with the Segment Anything Model (SAM). BaT-SAM injects local inductive bias into SAM’s frozen Vision Transformer (ViT) encoder via Conv-LoRA, where image features are processed by a sparsely gated mixture of multi-scale convolutional experts. This design compensates for the limited translation-equivariant inductive bias of vanilla ViTs while fine-tuning fewer than 5M parameters (less than 1% of the full model). To remove the need for user prompts, we further introduce a lightweight self-prompting box regressor that is jointly optimized with the adapter using scheduled sampling and predicts tumor locations directly from image embeddings. On the BRISC2025 benchmark, BaT-SAM achieves a weighted mean IoU of 0.832 while outperforming state-of-the-art benchmarks. It further demonstrates strong cross-dataset generalization, achieving Dice scores of 0.702 on Figshare and 0.784 on the LGG MRI Segmentation dataset.

[57] Across Viewpoints and Time: Trajectory-Generalizable Physical Attacks on World Models and Video Understanding Models cs.CV | cs.AIPDF

Song Xia, Meiwen Ding, Chenqi Kong, Wenhan Yang, Xudong Jiang

TL;DR: TARA 面向跨相机轨迹的实体可打印对抗纹理，用轨迹感知梯度对齐和任务精化降低实例间优化冲突；作者在 EPIC-KITCHENS-100、Something-Something V2 及实拍重摄数据上报告物理条件下可持续干扰。

Details

Motivation: 世界模型和视频理解模型用于动态视觉场景，但对实体对抗威胁的鲁棒性研究不足。单一物理纹理需在物体—相机视角沿不可控真实轨迹变化时仍能误导模型，异质视频语义和轨迹会使不同实例的优化梯度难以对齐。

Result: TARA（Trajectory-aware adversarial Alignment and Refinement Attack）利用轨迹感知优化刻画打印纹理在真实场景下随几何变化的效果，并通过轨迹对比对齐让不同视频/轨迹实例的攻击表示趋向共享对抗区域，再进行任务级精化以形成预测错误。为可复现评估，作者在 EPIC-KITCHENS-100 和 Something-Something V2 上构建几何 grounding 的追踪管线，生成物理合理、时间一致的攻击区域轨迹；同时采集真实打印—重摄视频验证物理部署效果。摘要报告单一可打印纹理能在物理条件下持续干扰世界模型和视频理解模型，但未提供攻击成功率或扰动大小。

Insight: 攻击重点是使同一可打印纹理跨不同对象—相机轨迹保持效用，梯度对齐与任务精化分别处理实例优化冲突和目标预测错误。摘要未量化攻击强度或防御结果，结论限于所列视频数据和实拍测试。

Abstract: World models and video understanding models are increasingly deployed in dynamic visual environments, yet their robustness to physical adversarial threats remains largely underexplored. We study whether a single physical adversarial texture can mislead the victim model across different real-world scenarios where observing viewpoints vary along uncontrolled object–camera trajectories. We identify that heterogeneous video semantics and trajectories induce poorly aligned gradients for optimizing the adversarial texture. To address that, we propose the Trajectory-aware adversarial Alignment and Refinement Attack (TARA), which incorporates trajectory-aware optimization to capture the evolving geometric variations of a printed texture under real-world scenarios. To improve cross-instance optimization compatibility, TARA incorporates trajectory-aware contrastive alignment to encourage attacked representations from diverse video–trajectory instances to approach a shared adversarial region and then performs trajectory-aware task refinement to translate the aligned representation into task-level prediction errors. For reproducible evaluation, we develop a geometry-grounded tracking pipeline that provides physically plausible adversarial regions with temporally consistent trajectories on EPIC-KITCHENS-100 and Something-Something V2. We further collect real-world print-and-recapture videos to validate attack effectiveness under physical deployment conditions. Together, the annotated benchmarks and real-world captured data form a reproducible testbed for future evaluation. Notably, our results reveal that a single printable adversarial texture can persistently disrupt world models and video understanding models under physical conditions, highlighting the fundamental security challenge.

[58] Video-Based Sexual Harassment Detection Using Fine-Tuned Video Transformers cs.CVPDF

Jubayer- Al- Mahmud, Md Shamimul Islam, Md Nur Hossain

TL;DR: 研究将 SHV 扩展至 600 个短视频并比较 VideoMAE、ViViT、TimeSformer 二分类；TimeSformer 测试准确率最高，为 92.27%。

Details

Motivation: 性骚扰视频检测可能帮助及时发现和干预，但衣着、光照、遮挡、低分辨率及相机视角变化都会造成困难，且大规模标注视频稀缺。

Result: 作者将 Sexual Harassment Video（SHV）数据集扩展至 600 个来自 YouTube 的短视频，按统一标注指南分为骚扰与非骚扰两类。使用在 Kinetics-400 预训练的 VideoMAE、ViViT、TimeSformer 三种视频 Transformer 微调做二分类。测试集准确率分别为 TimeSformer 92.27%、VideoMAE 91.71%、ViViT 91.16%。分析指出数据量、视频质量、帧抽样策略和模型架构都会显著影响结果。

Insight: 在小规模扩展数据集上三种模型表现接近，TimeSformer略高；摘要未提供类别平衡、跨来源验证或误报/漏报细节，不能据准确率直接推断真实监控部署可靠性。

Abstract: Sexual harassment represents a pervasive societal challenge, with a substantial proportion of incidents remaining unreported due to insufficient evidence and delayed intervention mechanisms. Smart surveillance systems capable of detecting such incidents in real-time could assist law enforcement and serve as a deterrent. However, video-based harassment detection poses significant challenges, including variations in clothing, illumina- tion, partial occlusions, low resolution, and diverse camera view- points. Despite its social importance, this problem has received limited research attention, primarily due to the scarcity of large- scale annotated video datasets. To address this gap, we extend the Sexual Harassment Video (SHV) dataset to 600 short video clips sourced from YouTube and annotated into harassment and non-harassment categories using consistent labeling guidelines. We fine-tune three state-of-the-art video transformer models, VideoMAE, ViViT, and TimeSformer, pretrained on Kinetics- 400, for binary harassment classification. Experimental results show that all three models achieve competitive classification performance, with TimeSformer attaining the highest test accu- racy of 92.27%, outperforming VideoMAE 91.71% and ViViT 91.16% on the extended SHV test set. Our analysis reveals that performance is notably influenced by dataset size, video quality, frame sampling strategy, and model architecture. This work presents a comprehensive evaluation of modern video transformers for this sensitive real-world task.

[59] VideoEvoHarness: Self Evolving Agent Harness from Unlabeled Videos for Video Understanding cs.CVPDF

Weixin Xu, Yifan Liu, Yi Wang, Mingfei Rong, Zhenyu Yang

TL;DR: VideoEvoHarness 从无标签视频联合进化 Questioner 和 Solver harness，固定模型参数，通过验证集比较候选并保留改进；在与进化视频不重叠的基准上准确率逐轮提升且可跨骨干迁移。

Details

Motivation: 视频理解智能体依赖 harness 获取证据、调用工具和推理，但改进 harness 通常需要人工。现有自进化方法常依赖带问答标注的视频或直接更新模型参数，未充分探索从无标签视频演化 harness。

Result: VideoEvoHarness 在冻结模型参数下联合进化提问者和求解器。Questioner 从视频记忆记录中生成问题和参考答案，问题质量评估与 Solver 执行轨迹反馈用于修订其生成策略；Solver 则依据执行反馈进化技能、策略、超参数和工具。可执行策略用代码定义工作流，减少反复解释自然语言指令。每轮在相同验证问题上比较候选 harness，只保留经过验证的改进。使用与进化视频不重叠的视频基准及专门问题集实验显示，整体准确率高于初始 harness，且收益可迁移到不同主干；专门问题集的准确率随进化轮数逐步提高。

Insight: 冻结模型、只演化提问与工具流程，说明无标签视频也可提供自监督式的 harness 改进信号。摘要未给具体基准名称、分数或进化成本，报告的是相对初始系统的整体改善和跨主干迁移。

Abstract: Video understanding agents rely on a harness to acquire evidence, use tools, and reason, but improving it requires substantial effort. Recent self evolution methods seek to optimize harnesses through execution feedback to reduce human intervention. However, existing methods often rely on annotated video questions and answers or focus on updating model parameters, leaving harness evolution from unlabeled videos underexplored. VideoEvoHarness jointly evolves a Questioner and a Solver harness from unlabeled videos with frozen model parameters. The Questioner constructs questions and computes reference answers directly from video memory records. Question quality assessments and Solver execution trajectories guide question generation policy revisions. The Solver supports evolving skills, strategies, hyperparameters, and tools using execution feedback. Executable strategies define workflows in code, reducing repeated interpretation of natural language instructions. Each round, candidate harnesses are compared on identical validation questions, with validated improvements retained for later use. Experiments on benchmarks and a dedicated benchmark question set, using videos disjoint from evolution videos, show higher overall accuracy than initially and transfer gains across backbones. Overall accuracy on this set also improves progressively across rounds. This supports our foundational framework for harness evolution from unlabeled videos, opening a promising direction for self evolving video agents. Code is available at https://github.com/Wei-Xin-Xu/VideoEvoHarness.

[60] OmniRSI: Taming Latent Diffusion for Unified Multimodal Remote Sensing Image Generation cs.CVPDF

Hebaixu Wang, Jing Zhang, Haonan Guo, Di Wang, Jiayi Ma

TL;DR: OmniRSI 用遥感专用 VAE、适配 SD3.5 的 ControlNet 和免训练循环潜对齐，统一支持多模态遥感重建/生成/翻译；作者构建 1,150 万张图的 RSI-MM 数据集。

Details

Motivation: 遥感数据常来自异构传感器且观测不完整，多模态生成可补充互补信息；现有方法多对自然图像生成模型逐传感器微调，缺少共享表示和稳健条件机制，难以统一生成。

Result: OmniRSI 提出统一多模态遥感图像生成模型。遥感专用 VAE 学习多模态先验并形成局部平滑潜空间，以支持不同观测数据的统一表征和稳健重建；随后完整适配 Stable Diffusion 3.5 并加入 ControlNet，利用文本和视觉条件统一完成多模态生成与翻译。训练免费的循环潜空间对齐机制迭代用生成的多模态信息补全不完整视觉条件，逐步提高引导可靠性，同时保持观测测量。作者构建 RSI-MM 数据集，含 11.5M 多分辨率图像、文字注释及部分配对数据。实验报告 OmniRSI 在多模态重建、生成和翻译方面优于比较方法，但摘要未列具体指标。

Insight: 方法把遥感专属潜空间、文本/图像条件控制及迭代条件补全结合起来，目标是跨传感器统一建模并保留已有观测。数据部分配对，摘要未给各任务数据划分或定量增益。

Abstract: Multimodal remote sensing image generation emerges as a promising paradigm for recovering complete and complementary observations from heterogeneous and partially observed remote sensing data. However, existing methods predominantly depend on sensor-specific fine-tuning of natural image generation models, limiting unified multimodal generation capacity due to the absence of shared representations and robust conditioning mechanisms. To address these challenges, we propose OmniRSI, a unified multimodal remote sensing image generation model. Specifically, we develop a remote sensing-specific variational autoencoder that captures multimodal priors and promotes a locally smooth latent space for unified representation and robust reconstruction across diverse observed data. Building upon that, we fully adapt Stable Diffusion 3.5 with a ControlNet branch to support unified multimodal generation and translation conditioned on textual and visual cues. Besides, we devise a training-free cyclic latent alignment mechanism that iteratively enriches incomplete visual conditions with generated multimodal information, progressively improving guidance reliability while preserving observed measurements. We also construct RSI-MM, a large-scale dataset comprising 11.5M multi-resolution images with text annotations and partially paired data for training and evaluation. Extensive experiments verify that OmniRSI achieves superior performance in multimodal reconstruction, generation and translation.

[61] Harnessing SPAD Noise: Test-Time Gated Attention for Indoor Depth Estimation cs.CV | cs.LG | cs.RO | eess.SPPDF

Yiquan Gao

TL;DR: HSN 将原生 SPAD 噪声视为环境先验，推理时通过门控注意力调整单目深度特征，不用稠密 3D 输入；真实室内基准上域内度量深度达到 SOTA，零样本跨环境更有潜力。

Details

Motivation: 单目深度估计成本较低，但尺度歧义、边缘扩散和未见环境泛化限制米制精度。已有 SPAD 融合方法多把噪声当作需要去除的伪影或忽略，未利用噪声中可能编码的确定环境信息。

Result: Harnessing SPAD Noise（HSN）结合一维 SPAD 直方图和 Gate Adapter，在推理时适配室内场景。三个优化阶段生成按通道门控的注意力向量，动态调节编码器特征图，从而应对尺度和边界歧义，无需其他测试时适配方法所需的高计算或稠密 3D 模态。真实室内基准实验显示，HSN 在域内米制深度精度达到 SOTA，在未见环境零样本泛化上表现出更大潜力。可解释性分析显示，SPAD 噪声越强，门控注意力越偏向较深语义层。摘要未给具体数值或数据集名称。

Insight: 作者将 SPAD 噪声重新解释为携带环境信息的先验，并在测试时用轻量门控利用它，而非先去噪。结果为域内 SOTA 和零样本潜力的摘要结论，没有精度/鲁棒性数值。

Abstract: Monocular depth estimation (MDE) has emerged as a cost-effective alternative to hardware-dependent 3D sensing modalities, yet predicting precise metric depth remains a persistent challenge due to intrinsic scale ambiguity, edge bleeding, and limited zero-shot generalization in unseen environments. Recent multi-modal approaches incorporate 1D/3D Single-Photon Avalanche Diode (SPAD) data to resolve these ambiguities, but they mainly treat SPAD noise as a harmful artifact required for denoising or leave the noise ignored. In this paper, we present an essentially different insight: native SPAD noise is an informative prior that implicitly encodes deterministic environmental characteristics. Driven by this paradigm shift, we introduce a novel, lightweight Test-Time Gated Attention framework by Harnessing SPAD Noise, therefore dubbed HSN. Tailored for indoor scenarios, HSN leverages the synergistic combination of 1D SPAD histograms and Gate Adapters (GAs). Through three complementary optimization phases, it generates channel-wise gated attention vectors to dynamically adjust encoder feature maps at inference time, effectively resolving ambiguous scale and boundary challenges without the heavy computational overhead or dense 3D modalities that other TTA methods suffer from. Extensive experiments on challenging real-world indoor benchmarks show that HSN achieves state-of-the-art metric depth accuracy in in-domain settings, while exhibiting even greater potential under zero-shot generalization to unfamiliar environments. Furthermore, interpretability analysis reveals that the proposed gated attention mechanism adaptively shifts attention toward deeper semantic layers under intensifying SPAD noise. Code will be made publicly available.

[62] Rethinking Supervision in Document Unwarping: A Self-consistent Flow-free Approach cs.CVPDF

Shaokai Liu, Hao Feng, Wengang Zhou

TL;DR: DRNet 以畸变图像自身几何一致性进行粗到细文档校正，并用未畸变文档监督而无需真实 warp 网格；在 DocUNet 上报告多项 SOTA。

Details

Motivation: 手机拍摄文档常有几何形变，影响数字存档和下游使用。现有方法通常需要网格形变真值，获取成本高且可能限制数据规模。

Result: DRNet 利用文档内在几何一致性指导矫正学习，并设计粗到细网络复用畸变图像表征，逐步改善校正结果。监督方式直接使用无畸变文档图像，不依赖已有方法中的真实形变网格。低层像素对齐和高层语义对齐共同监督畸变文档到平整图像的映射。在 DocUNet 挑战基准上，作者报告取得多项 SOTA，优于现有学习式方案；并通过一系列消融验证设计作用。摘要未给具体指标值。

Insight: 用校正目标图像监督映射而不要求逐像素形变场标注，可降低标注依赖；像素与语义对齐共同约束结果。摘要只说 DocUNet 多项 SOTA，未列具体指标和其他数据集泛化。

Abstract: In recent years, the proliferation of smartphones has led to an upsurge in the digitization of document files via these portable devices. However, images captured by smartphones often suffer from distortions, thereby negatively affecting digital preservation and downstream applications. To address this issue, we introduce DRNet, a novel deep network for document image rectification. Our approach is based on three key designs. Firstly, we exploit the intrinsic geometric consistency inherent in document images to guide the learning process of distortion rectification. Secondly, we design a coarse-to-fine rectification network to leverage the representations derived from the distorted document image, thereby enhancing the rectification result. Thirdly, we propose a unique perspective for supervising the learning of rectification networks, where undistorted document images are employed for supervision, which is free of warping mesh as ground truth in existing methods. Technically, both low-level pixel alignment and high-level semantic alignment jointly contribute to the learning of the mapping relationship between deformed document images and distortion-free ones. We evaluate our method on the challenging DocUNet Benchmark dataset, where it sets a series of state-of-the-art records, demonstrating the superiority of our approach compared to existing learning-based solutions. Additionally, we conduct a comprehensive series of ablation experiments to further validate the effectiveness and merits of our method.

[63] EmoTrace-Bench: Diagnosing Emotional Expression in Text-to-Video Generation through Controlled Cues and Intensity cs.CVPDF

Yeheng Wang, Liudan Shi, Yi Zheng, Jingke Li, Shiyuan Liu

TL;DR: EmoTrace-Bench 用 448 个受控提示、8 种情绪和两级强度评测 12 个文生视频模型；角色显式情绪反应的提示更可靠，增加情绪线索或电影化措辞并不稳定带来改善。

Details

Motivation: 文生视频模型可生成定向情绪表达，但整体视频—文本对齐指标难区分失败来自目标情绪、强度控制还是具体情感线索实现。

Result: EmoTrace-Bench 是细粒度诊断基准，包含 448 个控制提示，覆盖八种情绪、两个强度级别和不同情感线索条件；用 12 种 T2V 模型生成 5,376 段视频，评估目标情绪是否出现、感知强度以及十项视听情绪表达维度。自动评估器在部分数据上用人工标注校准和验证。结果显示模型和提示设计间存在显著差异；明确描述角色情绪反应的提示比只描述事件或场景更容易被遵循。增加情绪线索数量或使用更明确的电影化视听语言并不总能改善生成。

Insight: 整体相似度会掩盖目标情绪、强度及线索实现的局部差异；受控因素使错误更可追溯。自动评估器仅在数据子集与人工对齐，摘要未列不同模型的具体分数。

Abstract: Text-to-video (T2V) models are increasingly used to generate videos with targeted emotional expressions, yet existing evaluation methods mainly rely on holistic video-text alignment and provide limited insight into whether failures arise from target-emotion expression, intensity control, or specific affective cues. We introduce EmoTrace-Bench, a fine-grained diagnostic benchmark that controls emotion intensity and selected affective cues in prompts and evaluates generated videos at multiple levels. EmoTrace-Bench contains 448 controlled prompts covering eight emotions, two intensity levels, and diverse cue conditions. Using 12 T2V models, we generate 5,376 videos and evaluate target-emotion presence, perceived intensity, and ten audiovisual dimensions of emotional expression with an automated evaluator calibrated and validated against human annotations on a subset of the data. Results reveal substantial differences across models and prompt designs. Prompts that explicitly describe a character’s emotional reaction are followed more reliably than prompts that describe only events or scenes, whereas adding more affective cues or using more explicit cinematic audiovisual language does not consistently improve emotional generation. These results illustrate how aggregate alignment scores may overlook localized differences in target-emotion expression, intensity control, and cue realization. EmoTrace-Bench provides a traceable and interpretable framework for diagnosing emotional video generation without modifying existing generation pipelines, and we release the prompts and the automated evaluation rules.

Ziqi Wen, Shravan Murlidaran, Jonathan Skaza, William Wang, Miguel P. Eckstein

TL;DR: AUTO-SUM 用 MLLM 描述原图与逐对象移除图，再比较语义变化估计对象对场景理解的重要性；预测人工场景理解显著区域和自由观看注视的表现优于低层显著模型、接近 DeepGaze。

Details

Motivation: 传统低层图像显著性模型依赖颜色、边缘等视觉特征，难覆盖影响人类注意和眼动的高层认知因素。

Result: Automated Scene Understanding Map（AUTO-SUM）不使用人类数据拟合参数，而自动分割并逐一移除场景对象，让多模态大模型描述原图和修改图，再通过文本嵌入比较描述语义差异，估计该对象对场景理解的贡献。AUTO-SUM 能预测人类描述构造的场景理解地图（H-SUM）中最重要的对象；在自由观看及场景描述任务中，其对人类注视对象的预测优于 GBVS、Itti-Koch 等低层显著模型以及基于人类意义评分训练的 Meaning Map，并与用人类注视训练的 DeepGaze 相当。

Insight: 通过对象移除引发的场景语义变化估计高层显著性，无需拟合人类标注；它可补充而非取代低层显著模型。对 DeepGaze 的比较为相当水平，摘要未给具体准确率或场景规模。

Abstract: Image-computable low-level image saliency has shaped perceptual psychology and computer science, but is limited by its inability to capture human high-level cognitive factors that influence attention and eye movements. Here, we propose a new fully automated method to compute semantic saliency by quantifying each object’s contribution to scene comprehension without fitting any parameters to human data: the Automated Scene Understanding Map (AUTO-SUM). AUTO-SUM automatically segments and removes individual objects from a scene, uses multimodal large language models (MLLMs) to describe the original and modified scenes, and measures the semantic similarity between these descriptions using text embeddings. We show that AUTO-SUM can predict which object is most important in a human-measured scene understanding map (H-SUM) estimated using human scene descriptions. We also show that AUTO-SUM can predict the object most fixated by human observers during free viewing and scene description tasks better than a low-level saliency model (GBVS, Itti-Koch), a saliency model trained based on human meaning ratings (Meaning Map), and comparable to DeepGaze, which is trained on human fixation data. We contend that AUTO-SUM can be used as a semantic saliency model that complements lower-level saliency models. Code are available at https://github.com/starsky77/AUTO-SUM-Source-Code .

[65] Structured Stateful Relational Scripts with a High-Quality Dataset for Multi-Shot Audio-Visual Video Generation cs.CVPDF

Chen Zhao, Lehan Shen, Qilong Huang, Hongyu Li, Zhuoliang Kang

TL;DR: SSRS 将多镜头视频中的持久身份/声音与镜头级视听状态分开，并显式标注声源绑定及跨镜头事件；MSAV-Script-100K 提供 10 万规模结构化脚本数据，实验显示跨镜头一致性提升。

Details

Motivation: 连贯多镜头视听生成需要跨镜头保持身份、描绘状态变化并同步声音。密集字幕虽能详细描述每个镜头，却通常不说明哪些实体持续存在、其状态如何改变，以及哪些实体/动作产生声音。

Result: Structured Stateful Relational Scripts（SSRS）是一种文本监督表示，将持久实体身份和声音身份与每个镜头的视觉/音频状态分离，显式编码声源绑定及时间关系，并用全局音视频事件序列连接局部动作与声音，包括跨镜头延续的事件。作者发布 MSAV-Script-100K，包含整理后的多镜头视听视频及详细 SSRS 标注，支持学习身份持续、状态转换和事件连续性。实验报告该表示和数据可改善跨镜头一致性、对指定状态变化的遵循及音画对齐；摘要未提供指标数值。

Insight: 仅描述镜头内容不足以监督跨镜头实体和事件演化，SSRS 将身份、局部状态及全局时间关系显式拆开。结果为摘要级定性优势，没有给模型、数据划分或提升幅度。

Abstract: Coherent multi-shot audio-visual video generation requires maintaining consistent identities while depicting evolving states and coordinating sounds across shot boundaries. Although dense captions describe individual shots in detail, they rarely specify which entities persist, how their states change, or which entities and actions produce sounds. To address this limitation, we propose Structured Stateful Relational Scripts (SSRS), a textual supervision representation that separates persistent entity and voice identities from shot-specific visual and auditory states. SSRS explicitly encodes sound-source bindings and temporal relations, while a global audio-visual event sequence connects local actions and sounds, including events that continue across shot boundaries. We further introduce MSAV-Script-100K, a large-scale dataset of curated multi-shot audio-visual videos paired with detailed SSRS annotations, enabling models to learn identity persistence, state transitions, and event continuity from structured supervision. Experiments demonstrate improvements in cross-shot consistency, adherence to specified state changes, and audio-visual alignment, highlighting the effectiveness of our SSRS and MSAV-Script-100K for coherent multi-shot audio-visual generation. Our proposed dataset is available at \href{https://huggingface.co/datasets/shenlehan/MSAV-Script-100K}{MSAV-Script-100K}.

[66] A Plausible Future Is Not a Causal Future: Can Video Models Generate How the World Unfolds? cs.CVPDF

Jialiang Li, Zhengning Xu, Zongyang Ma, Shaobo Cui

TL;DR: CausalCascade 用 50 个 Blender 事件场景和 2,832 个受控反事实变体，测试九种 I2V 模型能否生成完整因果事件链；模型更能定位运动，而难以正确实现时空演化和物理干预。

Details

Motivation: 视频生成模型常被视为世界模型，但现有评估多看视觉质量、提示遵循或宽泛物理合理性，尚不清楚模型能否完整呈现起始事件如何经一系列因果交互到达终局。

Result: CausalCascade 是基于模拟器的因果级联基准，包含 50 个 Blender 事件丰富场景、结构化事件图、由图支持的诊断 VQA、三种未来事件指定粒度及 2,832 个受控反事实变体；评估九种图像到视频模型，覆盖运动、事件和反事实诊断。结果显示，生成运动的空间定位比正确实现时空演化更可靠；更丰富的未来事件描述可改善事件恢复，但物理实现仍有限。反事实评测发现模型对物理干预的选择性响应较弱，多数审查失败发生在基础层，如场景一致性或事件是否发生。

Insight: 因果级联生成是过程级世界建模测试，不只问最后画面是否合理，还检验未来事件、物理演化、干预响应和长期场景一致性。摘要未列模型级指标，报告的是跨九种模型的一般性失败模式。

Abstract: Large-scale video generation models are increasingly viewed as a promising path toward world models, yet their ability to capture causally coherent physical dynamics remains poorly understood. Existing evaluations largely emphasize visual quality, prompt adherence, or broad physical plausibility, leaving open whether a model can faithfully realize a complete causal cascade, in which an initiating event propagates through causally dependent interactions to a terminal outcome. We introduce CausalCascade, a simulator-grounded benchmark centered on 50 event-rich Blender scenes with structured event graphs, graph-grounded diagnostic VQA, three levels of future-event specification, and 2,832 controlled counterfactual variants, and use it to evaluate nine image-to-video models with complementary motion-, event-, and counterfactual-level diagnostics. Across the evaluated models, generated motion is more reliably localized spatially than realized with correct spatiotemporal evolution, while richer future-event specification improves event recovery but leaves physical realization limited. Counterfactual evaluation further reveals weak selective responsiveness to physical interventions, while most audited failures are foundational, arising at scene consistency or event occurrence. Together, these results position causal cascade generation as a process-level test of world modeling and highlight the need to jointly recover the intended future, faithfully realize its physical evolution, selectively respond to physical interventions, and maintain a coherent physical world over time. Code and dataset are available at https://github.com/SJTU-DeepDeltaLab/CausalCascade .

[67] Genesis: Visual Augmentation Primitives for Multimodal Chain-of-Thought Reasoning cs.CVPDF

Fan Yang, Lifeng Chen, Yipu Wang, Han Wang, Feilong Chen

TL;DR: VAP 用视觉生成模型合成多样图像与对应推理文本，并以多模态 CoT 规则 RL 评价中间步骤；作者报告 4B 模型超过 8B 基线，规模增至 32B 时持续改善。

Details

Motivation: 多模态大模型在 STEM 视觉推理基准上已有表现，但遇到复杂真实场景时能力下降。作者认为 CoT 多停留在语言空间，推理步骤少有视觉证据 grounding，训练数据缺少逐步视觉逻辑监督。

Result: Visual Augmentation Primitives（VAP）是训练数据扩充流程，用视觉生成模型对种子图像施加基本操作产生变体，并同步构造连贯多模态推理链，要求模型每一步依据视觉变化进行观察和逻辑推断。作者还提出多模态 CoT rubric RL，奖励不只看文字推理，也评估中间步骤的逻辑有效性、完整性和一致性。为评估超真实场景不同难度的视觉推理，构建 UNREAL 基准。实验报告基于该框架的 4B 模型超过 8B 基线，模型规模增加到 32B 时性能持续改善；摘要未提供分数或具体任务类别。

Insight: VAP 同时合成视觉样本和与之对齐的推理过程，RL 再校验中间步骤，使推理链逐步绑定图像证据。4B 超 8B 是摘要中的总体比较，具体 benchmark、模型配置和增益数值未列。

Abstract: Recent multimodal large language models achieve promising results on visual reasoning tasks, approaching human-level performance on existing STEM benchmarks. However, these capabilities degrade under complex real-world scenes. We identify that chain-of-thought reasoning remains largely confined to the linguistic space, rarely anchoring each step in visual evidence. The main cause is the absence of such data during training. In this work, we introduce a training data augmentation pipeline, \textbf{Visual Augmentation Primitives} (VAP), which leverages state-of-the-art vision generative models to synthesize diverse photorealistic training samples and corresponding reasoning text. VAP introduces some basic operators and applies them to a seed image and gets a variant. It also constructs coherent multimodal reasoning chains throughout the generation process. Such training data force models to perceive visual content and perform logical reasoning according to them at every step. We further introduce multimodal chain-of-thought rubric-based reinforcement learning to enhance this. It rewards not only text reasoning but also visual reasoning by scoring intermediate reasoning steps for logical validity, completeness, and consistency. Moreover, we construct the UNREAL benchmark to systematically evaluate visual reasoning on hyperrealistic scenes across controllable difficulty levels. Experiments show that a 4B model built on this framework surpasses 8B baselines, with performance improving consistently as the model scales to 32B, validating the scaling law for visual reasoning.

Hongbo Ding, Xiang Chen, Jiangxin Dong, Jinshan Pan

TL;DR: FireNav 用导航决策效用门控是否接纳重建特征，而不是一概最大化图像重建保真；在室内火灾 VLN 的多种设置中，作者报告成功率和路径效率均优于先进基线。

Details

Motivation: 室内火灾会严重退化视觉，影响指令 grounding、航点预测和长程探索。图像恢复看似可解决退化，但传统方法优化重建保真不一定改善导航，盲目恢复甚至会损坏任务关键线索。

Result: FireNav 提出面向决策的重建证据路由。FireVLN-Bench 提供成对室内火灾导航数据、保留任务语义的干净参考、安全监督，以及连续导航和含模拟动力学机器人执行的匹配设置。FireNav 保留原始视觉 token，将潜在重建特征作为辅助证据；轻量路由器依据视觉特征、任务上下文和先前执行状态，决定是否纳入以及对视觉 token 的贡献强度。Decision Utility Calibration 在同一状态下比较有/无候选证据时的动作预测损失，用其对航点、朝向和到达预测的贡献监督证据选择。实验显示 FireNav 在各评测设置的导航成功率与路径效率上持续优于先进基线，摘要未给出具体数值。

Insight: 恢复质量不等于导航效用；重建特征应按是否改善当前决策来选择性使用。摘要没有报告基准规模和定量收益，性能结论是跨评测设置的总体描述。

Abstract: In real-world indoor fire environments, visual degradation severely impairs instruction grounding, waypoint prediction, and long-horizon exploration in vision-language navigation (VLN). Although image restoration offers a natural means to mitigate such visual degradation, existing methods primarily optimize reconstruction fidelity, which does not necessarily translate into better navigation decisions and may even corrupt task-critical cues when applied indiscriminately. In this paper, we present FireNav, a decision-oriented reconstruction evidence routing framework that aligns visual restoration with navigation utility in VLN. To advance this task, we first contribute FireVLN-Bench, a paired indoor fire navigation benchmark with clean references preserving task semantics, safety supervision, and matched settings for continuous navigation and robot execution with simulated dynamics. Our FireNav retains native visual tokens and treats latent reconstruction features as auxiliary evidence, where a lightweight router leverages visual features, task context, and prior execution states to determine whether such evidence should be admitted and how strongly it should contribute to the visual tokens. To learn when reconstruction is beneficial, we further introduce decision utility calibration, which compares action prediction losses with and without candidate evidence under the same state, thereby supervising evidence selection according to its contribution to waypoint, heading, and arrival prediction. Experiments demonstrate that FireNav consistently outperforms state-of-the-art baselines in both navigation success and path efficiency across evaluation settings.

[69] Lesion-Aware Attention for Mammography Segmentation with Heterogeneous Supervision cs.CVPDF

Di Dai, Bo Liu, Liwei Wang

TL;DR: 该方法在乳腺 X 光片上联合训练掩码分割和图像级分类，使用类别原型与病灶注意力将弱标签传给局部区域；多个数据集上改善新类别开放词汇分割。

Details

Motivation: 乳腺病灶像素级标注成本高且稀少；公开分割集通常只覆盖少量病灶类型，而大规模乳腺影像数据往往只有图像级诊断标签，阻碍罕见/弱标注病灶的泛化分割。

Result: 作者研究混合监督下的开放词汇乳腺病灶分割：基础病灶类别有像素掩码，更广泛类别只有图像级标签。双分支框架联合学习：分割分支从掩码学习边界，分类分支提供额外图像级语义监督。Lesion-wise Attentive Classifier 为各类别维护特征原型并生成病灶注意力图，用于前景特征汇聚，减少背景组织影响并促使病灶表示更紧凑。多个乳腺摄影数据集实验中，方法在基础和新类别上持续优于开放词汇分割基线，并缩小与需要目标域密集掩码的全监督模型之间差距；定性结果显示小、低对比和细微病灶定位更完整。摘要未给具体指标。

Insight: 图像级标签由类别专属原型与注意力转成病灶区域证据，补充有限像素掩码。结论覆盖多个数据集和新类别，但摘要未说明疾病类别数、增益幅度或与全监督模型的具体差距。

Abstract: Accurate lesion segmentation in mammography is important for computer-aided breast cancer diagnosis, but dense pixel-level annotations remain costly and limited. Existing public segmentation datasets usually cover only a small set of lesion types with manually delineated masks, whereas larger mammography datasets often provide only image-level diagnostic labels. This annotation mismatch makes it difficult to train segmentation models that can generalize to rare or weakly annotated lesion categories. To address this problem, we study open-vocabulary mammographic lesion segmentation under a mixed-supervision setting, where pixel-level masks are available for base lesion categories and image-level labels are available for a broader set of categories. We propose a two-branch framework that jointly learns from segmentation data and classification data. The segmentation branch learns lesion boundaries from annotated masks, while the classification branch introduces additional image-level semantic supervision. To better transfer image-level supervision to lesion-relevant regions, we further introduce a Lesion-wise Attentive Classifier that maintains category-specific feature prototypes and generates lesion-aware attention maps for foreground feature pooling. This design reduces the influence of background tissue and encourages more compact lesion representations. Experiments on multiple mammography datasets show that the proposed method consistently improves open-vocabulary segmentation baselines on both base and novel lesion categories, while reducing the gap to fully supervised models that require dense target-domain masks. Qualitative results further indicate that our method produces more complete lesion localization for small, low-contrast, and visually subtle mammographic findings.

[70] SurgCast: Action-Conditioned Future Skeletons for Controllable Surgical World Models cs.CVPDF

Wanhao Liu, Rulin Zhou, Liangjing Shao, Zhaocheng Lin, Dongyue Li

TL;DR: SurgCast 先按当前结构、机器人状态和未来动作预测器械骨架，再用其控制手术视频生成；SutureBot/SRTH-Porcine-Cholei 上七项质量/一致性指标均优于直接动作基线，教师 FVD 分别降低 37.6%/67.3%。

Details

Motivation: 手术世界模型需将未来机器人动作转成视觉逼真、结构一致的器械运动，单有动作条件可能不足以约束器械的铰接结构和运动语义。

Result: SurgCast 根据当前器械结构、机器人状态和未来动作预测未来器械骨架，再用预测骨架控制视频生成。双路注入把骨架几何与器械专属运动描述符结合，通过几何残差和条件调制接入视频模型；两条路径补充原生动作条件，并保留在四步因果学生模型中。在 SutureBot 和 SRTH-Porcine-Cholei 上，教师和学生相较各自直接动作条件基线的七项质量及器械一致性指标均改善；教师 FVD 分别降低 37.6% 和 67.3%。消融显示未来动作能改善骨架预测，在预测骨架条件下，两路注入组合在六项指标上最佳。

Insight: 显式预测器械结构作为中间控制接口，补充动作指令与视频外观间的结构约束。FVD 改善对应教师模型两数据集结果，七项/六项指标统计范围不同；摘要未列绝对分数。

Abstract: Surgical world models must turn prospective robot actions into visually faithful, structurally consistent instrument motion. We introduce SurgCast, which predicts future instrument skeletons from current structure, robot state, and future actions, then uses the forecast to control video generation. Dual-path injection combines skeleton geometry with derived, instrument-specific motion descriptors through geometric residuals and conditioning modulation. Both paths complement native action conditioning and are retained in a four-step causal student. On SutureBot and SRTH-Porcine-Cholei, the teacher and student improve all seven reported quality and instrument-consistency metrics over their respective direct-action baselines; teacher FVD decreases by 37.6% and 67.3%. Ablations show that future actions improve skeleton prediction and that combining both injection paths performs best across all six metrics under predicted skeletons. These results support using predicted articulated structure and its motion semantics as a control interface for surgical video prediction. Project videos and examples are available at https://surgcast.netlify.app/.

[71] Sparse All-Layer Connector for Domain Generalised Semantic Segmentation cs.CVPDF

Ruoyu Guo, Xin Kun Lin, Maurice Pagnucco, Yang Song

TL;DR: SALC 让分割解码器各层可访问另一基础模型的所有前置层，再稀疏选择有用跨层信息，并用 candidate dropout 提升选择稳健性；四种模型组合、三种 DGSS 设置上优于层对齐融合。

Details

Motivation: 域泛化语义分割常组合视觉语言与视觉基础模型，但现有方法通常只融合深度对应层，隐含不同目标预训练出的模型在同层拥有互补特征；作者认为表示层级和域敏感性可能不同。

Result: Sparse All-Layer Connector（SALC）允许每一层访问另一模型前面各层的信息，以放宽固定深度对齐。由于可访问层随深度增加、密集聚合可能损害泛化，SALC 会自适应选择稀疏信息子集。Candidate dropout 正则进一步强化稀疏性并鼓励探索不同候选层，以提高选择稳健性。四种基础模型组合、三种 DGSS 评估设置中，SALC 泛化表现强于传统深度对齐融合；摘要没有给各场景具体指标。

Insight: 跨模型互补信息不一定出现在相同深度，稀疏 all-layer 检索将层级对应从固定结构变为可学习选择。结论基于所测四种组合和三种评估设置，未提供域外分数或计算开销。

Abstract: Domain generalised semantic segmentation (DGSS) has recently benefited from combining multiple foundation models, such as vision-language and vision foundation models, to leverage their diverse representations for generalisation. Existing multi-model approaches share a common design choice: cross-model fusion is restricted to depth-aligned layers, implicitly assuming that mutually beneficial information resides at the same depth across models. However, foundation models pretrained under different objectives may develop distinct representation hierarchies, making the optimality of depth-aligned fusion questionable. Moreover, we observe that the foundation models exhibit different depth-wise domain sensitivity. Motivated by these observations, we propose a Sparse All-Layer Connector (SALC) that enables each layer to access information from preceding layers of another model. As the number of accessible layers grows with depth, dense aggregation may hurt generalisation. SALC therefore learns to adaptively select a sparse subset of informative layers. We further introduce a candidate dropout regularisation that strengthens sparsity and encourages SALC to explore diverse layers, leading to more robust selection. Across four foundation model combinations and three DGSS evaluation settings, the proposed design demonstrates stronger generalisation than conventional depth-aligned fusion. The code is available at https://github.com/RuoyuGuo/SALC.

[72] A Persian Vision-Language-Action Dataset and Benchmark for Ambiguity- and Infeasibility-Aware Instruction-to-Plan Grounding cs.CV | cs.CLPDF

Abdollah Pourasghari, Saeedeh Sadat Sadidpour, Mohammad Roustaei

TL;DR: 该工作发布首个波斯语视觉—语言—动作规划数据集，含 258 个实例/230 张真实家居办公图片，并显式标注歧义与不可执行指令；检索增强是强且显著因素，其他新增组件在当前规模下未显著。

Details

Motivation: 将自然语言指令和场景图像转成可执行计划，是具身人机交互的重要能力；现有基准多为英语、模拟或视频环境，且常假定指令清楚可执行。

Result: 数据集包含 258 个样例、230 张真实家居/办公室照片，标注参考计划、刻意歧义/不可行标签、消歧问答对和自动生成 CoT 理由。研究消融检索增强、CoT 提示和智能体消歧步骤，并用 LLM 裁判协议和新指标 PGOS-FA 评估；该指标在既有 PG2S 上加入步骤顺序敏感性、过度生成惩罚和适配波斯语的嵌入。36 组配对检验显示，检索增强带来大且稳健显著的效果；细粒度新增组件方向上通常有利，但在当前样本规模下统计上无法与噪声区分。79 个真实输出的盲法人工相关性评估显示，PGOS-FA 至少和 LLM 裁判一样接近人工判断。跨主干泛化分析认为开放权重模型与强闭源规划器的大部分差距来自开放模型生成能力；最佳配置也显著超过独立复现的外部基线，尽管使用较弱主干。

Insight: 小规模真实场景数据显式纳入含糊/不可执行指令，并把步骤次序与过度生成纳入评估。统计证据最强的是检索增强；其他组件趋势尚不显著，需扩充样本验证。

Abstract: Converting natural-language instructions into executable action plans conditioned on a scene image is a core capability for embodied human-robot interaction, but existing benchmarks target English, simulated or video-grounded environments, and largely assume instructions are clear and executable. We introduce the first Persian vision-language-action dataset for this task: 258 instances over 230 real household/office photographs, annotated with reference plans, deliberate ambiguity and infeasibility labels, disambiguating question-answer pairs, and an auto-generated chain-of-thought rationale for training. Using this resource, we ablate retrieval-augmentation, chain-of-thought prompting, and an agentic disambiguation step, evaluated with an LLM-as-judge protocol and a new metric, PGOS-FA, extending prior work (PG2S) with step-order sensitivity, an over-generation penalty, and Persian-adapted embeddings. Paired statistical testing across 36 comparisons shows retrieval-augmentation is a large, robustly significant effect, while finer-grained additions are directionally consistent but not statistically distinguishable from noise at this scale, and the LLM-judge metric has near-zero power where PGOS-FA succeeds. A blind human-correlation study on 79 real outputs shows PGOS-FA tracks human judgment at least as closely as the LLM-judge. Testing generalization beyond the ablation’s single backbone, we find that most of the gap between an open-weight and a substantially stronger closed planner is attributable to open-weight model generation rather than to the closed model’s capability specifically, and that our best configuration significantly outperforms an independently reproduced external baseline despite a weaker backbone. The dataset, PGOS-FA implementation, and evaluation code will be made available upon reasonable request.

[73] ProtoPur: Prototype-Purity Pseudo-Labeling for Weakly Supervised Instance Learning in Pathology cs.CV | cs.LGPDF

Chen Shu, Chengyang Zhang, Boyu Fu, Ting Yin, Wenchuan Zhang

TL;DR: ProtoPur 用冻结病理基础模型特征构造原型区域，以类别均衡局部纯度生成弱监督伪标签；三个全切片数据集、两种编码器上均改善实例级性能并保持有竞争力的 bag 级分类。

Details

Motivation: 全切片图像可靠实例标注昂贵且有歧义，而弱监督 MIL 常只有 bag 标签。基于注意力、模型预测、负 bag 或空间细化得到的实例监督，未必能找到类别特异的局部证据。

Result: ProtoPur 在冻结基础模型特征构造的原型区域上估计类别均衡局部纯度，对与目标类别强关联的高纯度区域赋伪标签。其“纯度优先”设计将类别特异证据与负例/共享形态区分，覆盖率只用于保证获得足够伪标签。PAC 风格分析将实例分类器风险拆为一般一致收敛项，以及 ProtoPur 特有的伪标签噪声项（选中区域错配和覆盖残差）。三种公开全切片数据集、两种病理编码器实验中，ProtoPur 在多个 MIL 主干上持续提升实例级表现，同时保持有竞争力的 bag 级表现；两数据集伪标签质量对比也显示其选择更强。

Insight: 弱 bag 标签可能通过基础模型局部原型特征提取实例级监督，无需稠密实例标注。理论误差项显式区分选区错配与覆盖不足，实验覆盖三数据集和两编码器，但摘要没有量化提升。

Abstract: Weak bag-level labels are widely available in computational pathology, whereas reliable instance-level annotations for whole-slide images remain expensive and ambiguous. Weakly supervised multiple instance learning (MIL) achieves strong bag-level classification, but instance-level supervision from attention scores, model predictions, negative bags, or spatial refinement may fail to identify class-specific local evidence. We propose ProtoPur, a prototype-purity pseudo-labeling method for weakly supervised instance-level learning in pathology whole-slide images. ProtoPur estimates class-balanced local purity over prototype regions constructed from frozen foundation-model features, and assigns pseudo-labels to high-purity regions strongly associated with a target class. This purity-first design separates class-specific evidence from negative or shared morphology, with coverage used only to ensure sufficient pseudo-label assignment. We provide a PAC-style analysis showing that the risk of an instance classifier trained with ProtoPur pseudo-labels can be decomposed into a standard uniform-convergence term and a ProtoPur-specific pseudo-label noise term, comprising selected-region mismatch and a coverage residual. Experiments on three public whole-slide image datasets with two pathology foundation encoders show that ProtoPur consistently improves instance-level performance across multiple MIL backbones while maintaining competitive bag-level performance. Pseudo-label quality comparisons on two public datasets further demonstrate that prototype-purity selection provides stronger instance supervision. These results suggest that the foundation-model features contain exploitable instance-level supervision signals, and ProtoPur is a way to reveal and use these signals from weak bag labels without requiring dense instance annotations.

[74] Rethinking Human Intent-to-CAD: Parametric CAD Model Generation via Cooperative Multi-Task Alignment and Spatial-Aware Reinforcement Learning cs.CV | cs.AIPDF

Qingwang Zhang, Jiahao Li, Xiangdong Zhou

TL;DR: HiCAD 发布草图、文本和参数化 CAD 代码对齐的数据集，并用多任务对齐连接异构意图、空间感知 RL 约束几何/拓扑；多个任务上显著超过基线。

Details

Motivation: 概念设计阶段的人类设计意图往往以不完整草图和文字表达，从这些异构输入生成可执行参数化 CAD 代码仍困难，特别是不假设已有目标 CAD 模型时。

Result: 作者构建 HiCAD，包含手绘草图、文本描述和参数化 CAD 程序三者对齐的数据，并提出直接从多层级人类意图映射到可执行代码的统一方法。两阶段框架首先用 Cooperative Multi-Task Alignment 缩小不同输入模态的表示差距，再用 Spatial-Aware Reinforcement Learning 强化几何与拓扑一致性。实验显示其在多个任务上显著超过既有基线，可从多样意图生成高保真参数化 CAD 模型；摘要未给数据集规模、具体分数或各阶段增益。

Insight: 任务目标是可执行的参数化模型，而非单纯生成几何外观；对齐阶段处理异构输入，空间感知 RL 处理结构约束。当前摘要只提供总体结论和项目数据集名称。

Abstract: Parametric Computer-Aided-Design (CAD) modeling from human intent remains challenging, particularly during the conceptual design stage, where design goals are expressed through incomplete and unstructured modalities (e.g., hand-drawn sketches and textual descriptions). In this work, we rethink the human intent-to-CAD pipeline and propose a unified method that directly maps multi-level human intents to executable codes, without assuming the prior existence of target CAD models. To support our study, we construct HiCAD, the first large-scale dataset aligning hand-drawn sketches, textual descriptions, and parametric CAD codes. Based on this, we introduce HiCAD, a two-stage framework comprising Cooperative Multi-Task Alignment to bridge the representational gap between heterogeneous inputs, and Spatial-Aware Reinforcement Learning to enforce geometric and topological consistency. Extensive experiments demonstrate that our method significantly outperforms existing baselines across multiple tasks, validating its effectiveness and robustness in transforming heterogeneous human intents into high-fidelity parametric CAD models. Our project page: https://zqwlearning.github.io/HiCAD.

[75] PhySimCode: A Benchmark and Evaluation Method for Physics Video to Code Generation cs.CVPDF

Souraja Kundu, Aditya Gupta, Joel Julin, Yizhou Zhao, Liuyue Xie

TL;DR: PhySimCode 要求模型仅从视频推断物理规律、参数并生成能复现画面的 Python 仿真；10 种 MLLM 能陈述物理定律却难从视频归纳，最佳模型仅 7.7% 参数在真值 ±20% 内。

Details

Motivation: 视频描述运动或判断其是否合理，不等于模型学会控制运动的物理规律、估计定量参数或生成可运行仿真代码。已有视频物理基准较少测试视觉动力学到可执行代码的完整转化。

Result: PhySimCode 基准要求输入原始视频，模型推断相关物理、恢复参数并生成可运行 Python 代码重建视频。数据集含 160.6K 程序生成样本，覆盖 162 种物理现象，包括解析 SciPy 仿真和 PyBullet 渲染的刚体、铰接、流体及接触丰富场景；每例包含仿真代码、物理参数、视频及结构化 CoT（观察、定律、ODE、数值方法、验证检查）。十种开闭源 MLLM 按定律正确性、仿真等价、代码可执行和参数恢复评估。模型较能陈述物理定律（4.27–4.93/5），但从视频推断定律仅 1.51–2.47/5；最佳系统 Claude Sonnet 4.6 只有 7.7% 真值参数落在 ±20% 内。

Insight: 该基准把视觉理解、物理推断、参数估计、代码执行和结果复现连成端到端测试，揭示“会说定律”与“能从视频识别定律并仿真”间巨大差距。评估集含 2,430 视频，性能数字对应摘要所述评分协议。

Abstract: Multimodal large language models (MLLMs) are good at video understanding and code generation, but it remains unclear whether they can recover the physical laws underlying videos, a capability essential for converting videos into executable simulations and for providing structured physical supervision in robotics. Captioning motion video does not imply that a model has inferred its governing dynamics, estimated quantitative parameters, or can regenerate the observed scene from code. Existing video-physics benchmarks evaluate plausibility, property estimation, or future-state prediction, but offer limited insight into whether models can translate visual dynamics into runnable simulation code. We introduce PhySimCode, a benchmark for physics video-to-code generation: given only a raw video, a model must infer the relevant physics, recover parameter values, and generate runnable Python that reproduces the input video. The dataset contains 160.6K procedurally generated samples spanning 162 physics phenomena across analytic SciPy simulations and PyBullet-rendered rigid-body, articulated, fluid, and contact-rich scenes; each sample bundles the simulation code, physical parameters, rendered video, and a structured chain-of-thought covering observations, laws, ODEs, numerical methods, and validation checks. We evaluate ten state-of-the-art closed- and open-source MLLMs along physical-law correctness, simulation equivalence, code executability, and parameter recovery. Models reliably state physics laws (4.27-4.93/5) yet score only 1.51-2.47/5 on inferring them from video, and the best system (Claude Sonnet 4.6) recovers just 7.7% of ground-truth parameters within $\pm 20%$, exposing a substantial video-to-simulation gap. The 2,430-video evaluation set is available at https://huggingface.co/datasets/SourajaKundu123/PhySimCode and the evaluation code at https://github.com/SourajaKundu/PhySimCode

[76] Visually Legible but Adversarial: Watermark-Constrained Attacks on End-to-End VLMs for Document Understanding cs.CVPDF

Qingzhe Han, Fan Dang, Shengqi Chen, Xinyi Sun, Haoyu Xiong

TL;DR: AR-AWM 用模型感知的空间分布水印文本像素扰动攻击端到端文档 VLM；Qwen2.5-VL-7B 的 DocVQA EM 从 0.899 降至 0.150，但该结果为白盒设置且对模型迁移性弱。

Details

Motivation: 可见水印能标示文档版权、机密或状态，也可能成为对抗信号。研究评估常规水印形状区域在白盒攻击下是否可扰乱端到端文档理解模型。

Result: Auto-Regional Adversarial Watermarking（AR-AWM）是针对端到端文档 VLM 的模型感知白盒攻击，只优化小面积、空间分散的水印文字支持区域像素；攻击无目标且与具体问题无关，生成时不需问题、答案、解析提示、语言模型梯度或语言模型输出，每页可复用于多个下游查询。Qwen2.5-VL-7B 上，覆盖页面 7.98% 像素的水印支持区域使 DocVQA exact match 从 0.899 降至 0.150、ANLS 从 0.939 降至 0.212。相同区域不优化或使用匹配随机扰动时 ANLS 分别为 0.932、0.933，说明下降主要来自对抗优化；同布局优化矩形 ANLS 为 0.276，文字形状几何带来额外影响。OmniDocBench 报告的文本、公式、阅读顺序、表格指标均下降。InternVL2.5-8B 的单独白盒优化使 ANLS 从 0.922 降至 0.046，但跨模型迁移较弱，显示强编码器特异性。

Insight: 实验说明视觉上可读的常规水印区域在特定白盒模型下可成为攻击面；不优化或随机扰动的对照排除了普通遮挡为主因。结果受白盒访问和编码器特异性限制，摘要未给防御评估。

Abstract: Visible watermarks convey document ownership, confidentiality, and status, but can also carry adversarial signals. We introduce Auto-Regional Adversarial Watermarking (AR-AWM), a model-aware white-box attack on end-to-end visual language models (VLMs) for document understanding. AR-AWM constructs a small, spatially distributed watermark-text support and optimizes only its pixels to disrupt visual representations. The attack is untargeted and question-agnostic: generation uses no questions, answers, parsing prompts, language-model gradients, or language-model outputs. Each attacked page is reused across downstream queries. On Qwen2.5-VL-7B, a watermark support covering 7.98% of page pixels under the reporting threshold reduces DocVQA exact match from 0.899 to 0.150 and average normalized Levenshtein similarity (ANLS) from 0.939 to 0.212. Selected qualitative cases retain visually legible source fields. The same support without optimization and with matched random perturbations retains ANLS of 0.932 and 0.933, respectively, identifying adversarial optimization as the primary cause of degradation. Layout-matched optimized rectangles yield ANLS of 0.276, showing an incremental benefit from text-shaped geometry. AR-AWM also degrades all reported text, formula, reading-order, and table metrics on OmniDocBench. Separate white-box optimization on InternVL2.5-8B reduces ANLS from 0.922 to 0.046, while weak cross-model transfer reveals strong encoder specificity. These results identify conventional watermark-shaped supports as an adversarial surface for document VLMs in the evaluated white-box setting.luated white-box setting.

[77] ScenePlacer: Reformulating Scene-Level Object Pose Estimation as Visibility Field Learning cs.CVPDF

Ken Deng, Jiarui Liu, Guibiao Liao, Kunming Luo, Mochu Xiang

TL;DR: ScenePlacer 将生成物体的场景摆放拆为物体空间可见性场预测和刚体位姿对齐，训练仅需物体网格与渲染视图；3D-FUTURE 估计深度下 Chamfer/F-score 为 0.0077/92.40。

Details

Motivation: 将新生成的 3D 物体放入部分观测场景时，既需保留生成模型的 canonical shape prior，又要估计场景级位置姿态。直接在场景坐标生成可能损害先验；从完整生成形状回归姿态对尺度歧义和部分观察敏感。

Result: ScenePlacer 将物体生成与场景位姿估计解耦，重构为条件可见性场学习：输入重建得到的局部点云和生成网格，预测物体坐标系中的可见性场，再将部分观测与网格可见区域对齐求刚体变换。骨干是预训练 3D shape VAE，保持其 mesh prior，并单独编码场景观察作为条件。可见性感知、半径约束监督及仅用物体级网格与渲染视图合成训练对，避免场景位姿标注。3D-FUTURE 上在真值深度和单目估计深度下均优于已有放置/位姿精化基线；生成物体在估计深度时 Chamfer distance 0.0077、F-score 92.40，真值几何时为 0.0023 和 98.27。

Insight: 在物体空间学习可见性场可把部分观察与 canonical mesh 对齐，同时避免改动生成流程或依赖场景级位姿标签。两组数值分别对应估计深度和真值几何，存在明显输入质量差异。

Abstract: We present ScenePlacer, a framework for scene-level pose estimation of objects generated by state-of-the-art native shape generation models, placing generated 3D objects into scenes given a partial geometric observation without modifying the underlying generative pipeline. Existing scene generation approaches either generate objects directly in scene coordinates, compromising their canonical-space priors, or regress poses from complete generated shapes in the shape generation process - the latter being a poorly conditioned problem that is sensitive to scale ambiguity and partial observations. ScenePlacer decouples object generation from pose estimation by reformulating alignment as conditional visibility-field learning. Given an observed partial point cloud from reconstruction and a generated mesh, it predicts a visibility field defined in object space, then estimates the rigid transformation by aligning the partial observation to visible parts on the mesh. We propose a pretrained 3D shape VAE as our backbone, preserving its mesh prior while conditioning it on a separately encoded observation. We further introduce visibility-aware, radius-constrained supervision and synthesize training pairs using only object-level meshes and rendered views, eliminating the need for scene-level pose annotations. On 3D-FUTURE, ScenePlacer consistently outperforms existing placement and pose-refinement baselines under both ground-truth and monocular depth estimation. With generated objects, it achieves a Chamfer distance of 0.0077 and an F-score of 92.40 under estimated depth and improves to 0.0023 and 98.27, respectively, under ground-truth geometry.

[78] RSAJEV: Training-Free Abstaining Prototype Memory for Eliminating Silent Hallucination in JEV-Based Remote Sensing Classification cs.CVPDF

Man Chen, Antai Chen, Shuai Xu, Chengjie Gu, Huaping Zhou

TL;DR: RSAJEV 发现基于类别字母绑定的遥感 MLLM 分类虽在词汇内 UCM 达 97.71% OA，跨数据集可从 7.35% 到 96.40% 且会对未见类高置信；免训练原型检索读出恢复开放集识别和弃答校准。

Details

Motivation: 遥感场景分类可将候选类名映射为提示中的字母选项并通过单次前向打分，但这种槽位绑定在训练词表外可能不稳定，且高置信度不一定表示类别已知。

Result: 作者发现字母绑定在词汇内表现好（UCM OA 97.71%），但四种跨数据集方向零适配准确率波动 7.35%–96.40%；开放集上未见类置信度反而高于已见类，AUROC=0.38，称为 silent hallucination。RSAJEV 保留槽位表示但替换读出为免训练、可弃答的原型记忆：同一提示下缓存每类 K 张参考图的槽表示，通过最近邻检索分类，并用收缩 Mahalanobis 几何校正细化。该读出在 UCM 跨数据集 gate 上每类 16 张参考图、零梯度更新达到 89.43%，AID 达 96.90%；开放集 ID-ness 分数在两数据集上 AUROC 为 0.91–0.99，四种设置均高于符号置信度（0.38–0.96）。通过离线重评分缓存矩阵支持零额外推理成本的类别增删。

Insight: 问题来自词汇绑定槽位而非单纯视觉识别不足；将读出替换为参考图原型匹配并显式估计 ID-ness，可恢复可校准弃答能力。准确率/开放集指标依赖参考图数量和数据方向，开放集结果与原始字母置信度不可混为一谈。

Abstract: Scene classification of remote sensing imagery with multimodal large language models (MLLMs) is formulated as a single-forward-pass decision problem: candidate class names are lettered options in the prompt, a registered answer slot is conditioned on the prompt, and classification is read out by scoring letter tokens with the language-modeling head. We show that although this letter binding is highly effective within its training vocabulary (97.71% OA on UCM), it behaves erratically out of vocabulary: zero-adaptation accuracy across four cross-dataset directions fluctuates between 7.35% and 96.40%, and under out-of-vocabulary weights its confidence fails inversely on open-set evaluation, assigning higher confidence to unseen than to seen classes (AUROC 0.38), a failure we term silent hallucination. We retain the slot representation and replace the readout with RSAJEV, a training-free abstaining prototype memory that performs retrieval-based prototype matching: K reference images per class are cached as prototypes via their slot representations under the same prompt, and classification becomes nearest-neighbor retrieval over the prototypes, refined by a shrinkage-Mahalanobis geometric correction. This readout is mechanistically isomorphic to retrieval-augmented generation (RAG) but operates on discriminative decisions; it is the first retrieval-based readout for MLLM slot representations. The retrieval readout reaches 89.43% on the UCM cross-dataset gate with 16 references per class and zero gradient updates, and 96.90% on AID. Its ID-ness scores attain 0.91-0.99 AUROC on open-set evaluation over two datasets, exceeding symbolic confidence (0.38-0.96) in all four open-set settings, restoring a calibratable abstention channel and supporting class addition/removal at zero extra inference cost via offline rescoring of the cached matrix.

[79] SpikeOcc: Towards Resilient Multimodal 3D Occupancy Perception via Spiking Experts cs.CVPDF

Yiding Sun, Zijie Xu, Hongqiang Lin, Jiajun Cao, Dongxu Zhang

TL;DR: SpikeOcc 用脉冲专家分别处理 RGB、LiDAR 和联合融合，并通过渐进传感器可用性训练提高缺失传感器时的性能下限，同时保持全传感器精度有竞争力。

Details

Motivation: 多模态 3D 占用预测对自动驾驶场景理解重要，但以往主要追求更高融合精度，较少研究传感器缺失时的鲁棒性；SNN 的离散脉冲也使优化更困难。

Result: SpikeOcc 提出 Spiking Mixture-of-Experts（SMoE）和 Sensor Availability Learning（SAL）。SMoE 将 RGB、LiDAR 和联合融合组织为专门路径，保留完整宽度的单模态表示，并以有界残差保留原多模态特征。SAL 超越普通传感器 dropout，采用带显式路由监督的分阶段课程，逐步提高严重传感条件下的鲁棒性。作者还发现 SAL 可提升传统非脉冲模型。多项实验显示 SpikeOcc 在全传感器输入时保持有竞争力的准确率，同时在传感器缺失下显著抬高性能下限；摘要未给数值。

Insight: 分离专家保留各传感器专属路径，再按传感器可用性训练路由，可减轻缺模态时融合表征崩溃。SAL 对常规模型也有效，提示方法价值不局限于神经形态 SNN；摘要未量化能耗或鲁棒性增幅。

Abstract: Multimodal 3D occupancy prediction underpins safe autonomous driving through comprehensive scene understanding. However, existing methods have largely pursued higher accuracy through sophisticated fusion, leaving resilience to missing sensors underexplored. In this paper, we investigate spiking neural networks (SNNs) as a means of building more resilient multimodal perception. Still, the discrete nature of spiking neurons makes optimization challenging. To overcome this difficulty, SpikeOcc is introduced with two complementary components, a Spiking Mixture-of-Experts (SMoE) and Sensor Availability Learning (SAL). Specifically, SMoE organizes RGB, LiDAR, and joint fusion into specialized pathways, enabling full-width unimodal representations while retaining the original multimodal features via a bounded residual. Beyond sensor dropout, SAL employs a staged curriculum with explicit routing supervision, progressively improving robustness under severe sensing conditions. Remarkably, SAL also yields notable gains for conventional models, highlighting its broader potential for resilient multimodal learning. Extensive experiments demonstrate that SpikeOcc maintains competitive full-sensor accuracy while markedly raising the performance floor under sensor outages. SpikeOcc illustrates the promise of neuromorphic computing, pointing toward more reliable and energy-efficient autonomous driving.

Songze Li, Haoming Gu, Ziyang Gong, Zhitong Wang, Gengze Zhou

TL;DR: RSI-Nav 让同一模型交替生成导航指令、执行轨迹和筛选样本，按小批更新保持伪监督与当前策略同步；R2R-CE/RxR-CE 成功率分别为 68.6%/72.4%。

Details

Motivation: 视觉语言导航需要高质量语言指令监督，但人工标注昂贵。固定模型预先标注整批轨迹会使生成监督落后于不断更新的策略，分开训练导航与指令生成也难共享能力。

Result: RSI-Nav 将视觉语言到动作的指令跟随与视觉动作到语言的指令生成统一在同一模型中。监督微调建立基础能力后，系统逐小批处理未标注轨迹：当前模型为轨迹生成指令，执行后按目标完成度和路径效率筛选合格样例，并在处理下一批前立即更新自身，以使监督跟随策略演化。方法还加入空间理解数据。实验中 R2R-CE 成功率为 68.6%，RxR-CE 为 72.4%；递归进化后指令生成实用性更强，空间理解也有小幅提升。

Insight: 逐批生成、执行、筛选和更新可减少静态伪标签与当前策略间的偏移，并共享指令理解/生成能力。摘要未给基线差值、数据规模或消融数值。

Abstract: Vision-language navigation requires agents to follow natural-language instructions to reach target locations, making high-quality language supervision essential. Since human annotation is costly, prior work has explored scalable data synthesis methods. However, existing methods typically either use separate models for navigation and instruction generation or couple the two tasks through task-specific components; most also annotate the full trajectory corpus with a fixed model before updating, which can cause generated supervision to lag behind the evolving policy. Therefore, we introduce RSI-Nav, a data-centric recursive self-improvement framework for navigation via policy-synchronized language-action dual bootstrapping, which unifies vision-language-to-action instruction following and vision-action-to-language instruction generation within a shared model. Specifically, after supervised fine-tuning establishes fundamental capabilities, RSI-Nav progressively processes unlabeled trajectories in small shards. The current model generates instructions for these trajectories, executes them to evaluate goal completion and path efficiency, retains qualified samples, and immediately updates itself before processing the next shard. This shard-wise update keeps newly generated supervision aligned with the evolving policy. We also incorporate spatial understanding data, since stronger spatial reasoning benefits navigation. Extensive experiments show that RSI-Nav reaches 68.6% SR on R2R-CE and 72.4% SR on RxR-CE, while providing stronger instruction-generation utility and modestly improved spatial understanding after recursive evolution.

[81] Reference-Efficient Multimodal Bitstream-Corrupted Video Recovery via Depth-Conditioned Adaptation cs.CVPDF

Tianyi Liu, Jiaojiao Wang, Ye Lu, Kejun Wu, Chengtao Cai

TL;DR: MBSCVR 以深度几何条件适配单个非局部 RGB 参考来修复受损视频码流；BSCV 基准 oracle mask 下 YouTube-VOS/DAVIS PSNR 为 32.30/28.93 dB，接收端延迟和峰值内存降低 36.8%/40.8%。

Details

Motivation: 视频码流损坏后，局部残差和视频中的远程 RGB 参考都可能不可靠；额外处理更多参考还会提高接收端计算与内存负担。预传输深度几何可在外观损坏时帮助恢复。

Result: 多模态码流损坏视频恢复（MBSCVR）结合 RGB 和深度信息，针对每个恢复窗口只选择一个非局部 RGB 参考。可靠性多模态选择器在编码参考前结合外观有效性与深度兼容性；Depth-Conditioned Reference Adaptation（DCRA）将选定 RGB 证据用于特征补全和内容生成，帧对齐深度则条件化补全查询。在完整 BSCV 基准、oracle masks 条件下，单参考方案在 YouTube-VOS/DAVIS 分别达到 32.30/28.93 dB，比使用全量参考的 B2SCVR 高 0.27/0.33 dB，且两个子集 SSIM 更高。匹配接收端前向测试显示延迟降低 36.8%、峰值内存降低 40.8%。

Insight: 通过深度兼容性筛选一个最有用外观参考，可在提升质量时避免多参考开销；结果依赖 oracle mask 和深度条件。PSNR 及资源收益都在所述 BSCV 评测设置下报告。

Abstract: Recovering visually faithful content from damaged video bitstreams is important for reliable multimedia communication and storage. Existing methods exploit local residuals and non-local RGB references of the input video sequence, but both sources can become unreliable after decoding, while processing additional references increases receiver computation and memory. This paper studies multimodal bitstream-corrupted video recovery (MBSCVR) using RGB and depth information, where pre-transmission depth-based geometry aids reconstruction when the received appearance is corrupted. We propose a novel MBSCVR framework that selects and effectively utilizes a single selected non-local RGB reference per recovery window to achieve higher performance. It integrates a reliability-aware multimodal selector that combines valid appearance with depth compatibility before reference encoding. Depth-Conditioned Reference Adaptation (DCRA) introduces selected RGB evidence into feature completion and content generation, while frame-aligned depth conditions completion queries. On the complete BSCV benchmark with oracle masks, MBSCVR uses one selected reference to achieve 32.30 dB on YouTube-VOS and 28.93 dB on DAVIS, surpassing full-reference B2SCVR by 0.27 and 0.33 dB and improving SSIM on both subsets. Matched receiver-forward measurements show 36.8% lower latency and 40.8% lower peak memory. These results demonstrate how compact geometric guidance and adapted appearance processing improve the quality–resource trade-off of multimedia recovery.

[82] Totem: Object-Centric 3D Memory for World-Consistent Video Generation cs.CVPDF

Zichen Wan, Tianyu Huang, Fan Li, Kai Wang, Hongzhi Zhang

TL;DR: Totem 用共享世界坐标系下的持久对象级 3D 记忆约束视角变化下的视频生成，并结合场景几何做双流 3D 条件化；多个世界生成/对象中心基准上表现出更稳定的对象与空间布局。

Details

Motivation: 从统一世界生成连贯视频仍困难，尤其相机视角大幅改变时，现有方法难以稳定表示单个对象并维持空间关系。

Result: Totem 为选定对象建立对齐共享世界坐标系的持久 3D 记忆，并为每个对象维护独立状态。双流 3D 条件结构将对象记忆与场景几何结合用于目标视角合成；对象记忆还通过三维结构信息约束注意力，以保留物体表面和相对空间布局。多个世界生成及对象中心基准实验报告，在具有挑战性的相机轨迹下能生成对象一致的三维世界视频，并支持多种应用；摘要未提供具体指标。

Insight: 对象级显式状态与共享世界坐标系把跨视角一致性作为生成条件，而不是只依赖逐帧视频先验。结果为摘要定性结论，尚无增益、存储或计算成本数值。

Abstract: Generating coherent videos grounded in a unified underlying world remains a major challenge, particularly under large viewpoint changes, where existing methods often struggle to maintain stable representations of individual objects. To this end, we introduce Totem, an object-centric 3D memory framework for world-consistent video generation. Totem first represents each selected object as a persistent 3D memory aligned to a shared world frame, assigning it a dedicated state. Based on these memories, we design a dual-stream 3D conditioning architecture that combines object memories with scene geometry for target-view synthesis. Moreover, these object memories constrain the model’s attention with 3D structural information, thereby preserving object surfaces and relative spatial layouts. Extensive experiments on world-generation and object-centric benchmarks demonstrate that Totem generates videos depicting a more object-consistent 3D world under challenging camera trajectories, while supporting versatile applications.

[83] Adaptive Look, Flexible Thought: Learning How and How Long to Think in Multimodal Reasoning cs.CV | cs.AI | cs.CLPDF

Quanlin Chen, Xiuwei Chen, Kun Xiang, Zehua Ma, Chunjin Ye

TL;DR: L2T 把工具产生的视觉状态离散成 token，统一优化视觉观察形式和文字推理深度；在多项任务中保持与显式工具调用相当的表现，吞吐提高 1.7–3.6 倍。

Details

Motivation: 现有多模态快/慢推理主要调节文本 CoT 长度，将外部工具或视觉 token 带来的视觉计算视为辅助部分，难以联合优化视觉感知与文字推理。

Result: L2T（Adaptive Look, Flexible Thought）将工具生成的视觉状态离散为视觉 token，纳入与文本推理相同的自回归空间。方法引入 modality-aware routing reward 选择合适推理模式，并根据任务复杂度用 difficulty-aware fast-slow 策略调整推理深度。一般感知和高推理需求任务实验显示，L2T 有效；相较显式工具调用方法，表现保持竞争力，同时吞吐提升 1.7–3.6 倍。

Insight: 将视觉工具输出纳入统一序列，使模型能共同决定如何观察和推理，而不把视觉处理固定成外部步骤。吞吐增益相较显式工具调用，摘要未提供各任务准确率或成本分解。

Abstract: The introduction of long chain-of-thought (CoT) reasoning and external tools has substantially improved the performance of multimodal large language models (MLLMs) on general and reasoning tasks. However, existing fast-slow reasoning paradigms mainly regulate textual reasoning length, while visual computation introduced by external tools or visual tokens is often treated as an auxiliary component. This separation prevents MLLMs from jointly optimizing visual perception and textual reasoning. To address this issue, we propose L2T (adaptive Look, flexible Thought), a unified reasoning framework that jointly optimizes the form and depth of computation. L2T first discretizes tool-produced visual states into visual tokens, bringing visual perception and textual reasoning into a shared autoregressive space. Based on this unified representation, L2T introduces a modality-aware routing reward to select appropriate reasoning modes and a difficulty-aware fast-slow strategy to adapt reasoning depth according to task complexity. Experiments on general perception and reasoning-intensive tasks demonstrate the effectiveness of L2T. Compared with explicit tool-calling approaches, L2T maintains competitive performance while achieving 1.7–3.6 times higher throughput.

[84] ChangeT$^3$: Evaluating Vision Test-Time Training for Remote Sensing Change Detection cs.CVPDF

Sharukshan Niranjan, Oshadha Samarakoon, Nichula Wasalathilaka, Roshan Godaliyadda, Parakrama Ekanayake

TL;DR: ChangeT³ 在 Siamese ViT 编码器各图像流独立计算快速权重，再由多尺度解码器比较双时相特征；LEVIR-CD+ 三种子 F1 为 80.61±0.90，但按标准协议在测试集选 checkpoint，且跨数据集零样本迁移失败。

Details

Motivation: 遥感双时相变化检测需要同时保留空间细节和变化上下文。测试时训练可按图像适配小型内模型，但需检验图像条件快速更新能否在不加专门时间交互模块下有效。

Result: ChangeT³ 使用层级 ViT³-Tiny 编码器与多尺度 ChangeFormer 解码器组成 Siamese 结构。两路共享学习参数，但快速权重从各自图像的空间特征独立计算；编码器保留全局简化 SwiGLU 和局部自适应深度卷积分支，解码器在四个尺度学习融合特征并逐级聚合得到二值预测。在 LEVIR-CD+ 从头训练，控制解码器、配方和预算与 ChangeFormer 对照一致，ChangeT³ 三种子 F1 为 80.61±0.90，对照 78.84±0.21；关闭所有内更新后同编码器为 78.63±0.20。该数据集标准协议在测试集选 checkpoint。ChangeT³ 与对照均无法零样本迁移到 SYSU-CD、DSIFN-CD 或 WHU-CD；小输入下 ChangeT³ 还更慢，参数量和运算量相近。

Insight: 图像条件快速权重带来 LEVIR-CD+ 内的 F1 提升，但评估存在测试集选 checkpoint 的协议限制，且无跨数据集泛化证据。速度在小输入下反而较慢，不能只凭精度结论称其更高效。

Abstract: Bitemporal remote sensing change detection requires representations that preserve spatial detail while capturing contextual information relevant to changes between acquisitions. Visual test-time training offers an alternative to explicit token-to-token attention by adapting a compact inner model to each image. This letter describes ChangeT$^3$, a vanilla Siamese configuration combining the existing hierarchical ViT$^3$-Tiny encoder with the multiscale ChangeFormer decoder. The two image streams share learned encoder parameters, while their fast weights are computed independently from their respective spatial features. Each encoder block retains the inherited global simplified SwiGLU and local adaptive depthwise-convolution branches. Temporal comparison is performed by the decoder through learned fusion of paired features at four resolutions, followed by coarse-to-fine aggregation and binary prediction. This separation provides a direct setting for examining image-conditioned inner updates without introducing an additional temporal interaction module. Trained from scratch on LEVIR-CD+ with the same decoder, recipe, and budget as a ChangeFormer control, ChangeT$^3$ reaches an F1 of 80.61$\pm$0.90 against 78.84$\pm$0.21 for ChangeFormer (three seeds each); as in the standard protocol for this dataset, checkpoints are selected on the test split. Trained with both inner updates disabled, the same encoder reaches 78.63$\pm$0.20. Neither model transfers zero-shot to SYSU-CD, DSIFN-CD, or WHU-CD, and ChangeT$^3$ is the slower of the two at small inputs despite similar parameter and operation counts.

[85] Sharper Frames, Worse Rooms: Viewpoint Coverage Dominates Image Quality in Feed-Forward 3D Reconstruction cs.CVPDF

Alibek Dadajonov, Hikmatillo Umarhojiyev

TL;DR: 九个手持室内扫描中，挑最清晰 16 帧反而造成最高 48% 房间尺寸误差，而均匀时间采样仅 3%；因清晰帧常集中在小角度慢运动，视点覆盖比单帧清晰度更重要。

Details

Motivation: 多视图前馈 3D 重建有固定帧数预算，因此选帧是重要上游环节。常规认为应优先选择最清晰图像，但清晰度可能与相机运动和视角分布耦合。

Result: 研究在九个手持房间扫描中比较 200 个候选帧中选 16 帧的策略。选最清晰帧是所测最差方案，房间尺寸误差达 48%，而均匀时间采样为 3%。原因是清晰帧往往发生于慢运动，集中在狭窄角度范围；示例有 192° 房间未观察，最差单元超过 300°。2×2 因子实验独立操纵视角覆盖和单帧清晰度，发现覆盖主效应约为画质三倍；九次拍摄中有八次，清晰但聚集的选帧不如模糊但覆盖好的集合。帧预算增加、角覆盖饱和后收益减弱，与该机制相符。推理时重建模型多视图一致性分数随重建误差上升，而给模型相同帧的 VLM 自信度与自身空间误差无关。

Insight: 小视图预算下优先覆盖足够视角比挑单帧锐度更重要；清晰度与相机运动耦合导致错误选择。两种常见推理时不确定性信号都未能暴露上游覆盖缺失。结果来自九个房间扫描和固定视图预算。

Abstract: Grounded spatial reasoning about indoor scenes rests on a geometric reconstruction that the multimodal system does not itself produce. Feed-forward multi-view reconstruction models operate under a fixed view budget, making frame selection a critical upstream decision. We show that under this setting the conventional preference for sharp frames can be actively harmful. Across nine handheld room scans, selecting the sharpest 16 of 200 candidate frames is the worst strategy we test, reaching 48% room dimension error where uniform temporal sampling reaches 3%. We trace this failure to a coupling between image quality and camera motion: sharp frames tend to occur during slow motion and therefore concentrate the selection in a narrow angular region, leaving $192^\circ$ of the room unobserved in the case we illustrate and over $300^\circ$ in the worst cells. A $2\times2$ factorial experiment that independently manipulates viewpoint spread and per-frame sharpness shows that spread has roughly three times the main effect of quality, and that blurry but well spread selections outperform sharp but clustered selections in 8 of 9 captures. The advantage also diminishes as the frame budget increases and angular coverage saturates, consistent with this mechanism. Finally, we ask whether the failure is detectable at inference: the reconstruction model’s multi-view consistency score increases with reconstruction error, while a vision-language model given the same frames reports confidence unrelated to its own spatial estimation error. Under small view budgets, viewpoint coverage is a more important constraint than per-frame image quality, and neither uncertainty signal available at inference reveals this upstream failure.

[86] Mobility Helps Learning: Unsupervised Model Adaptation for Object Recognition via Movement cs.CV | cs.LGPDF

Yanan Ma, Yihang Tao, Zhengru Fang, Senkang Hu, Xianhao Chen

TL;DR: MoCaFe 利用同一物体随距离/视角变化的预测波动做无监督测试时适配，以跨视角逆方差融合构造伪标签；nuScenes、KITTI 和具身数据上最高准确率增益约 20%。

Details

Motivation: 自动驾驶和机器人需无标注地适配新环境。常见 TTA 把测试样本当独立快照，忽视移动智能体观察同一物体时不同距离和视角造成的预测质量变化，这种波动可作为额外监督信号。

Result: MoCaFe 是模型无关、对超参数不敏感的无监督适配范式。它根据跨视角预测方差做 mobility-calibrated pseudo-label filtering，以逆方差权重融合多视角结果，构建可靠伪标签，再用筛选结果挑选有学习价值的 hard samples。可选 p-MoCaFe 使用公开数据为样本加权，以得到偏差有界的 clean-loss 估计，进一步抑制伪标签污染。在 nuScenes、KITTI 及具身智能体数据集上的实验中，MoCaFe 显著超过先进 TTA 方法，准确率增益最高约 20%。

Insight: 运动产生的重复观察不只是冗余，可通过视角间预测方差评估伪标签可靠性。20%为最高准确率增益，摘要未列各数据集/模型具体结果，不能视为普遍提升幅度。

Abstract: Movable agents, such as autonomous vehicles and robots, should be able to autonomously adapt their pre-trained object recognition models to new environments without human annotations. While existing test-time adaptation (TTA) methods address label-free model adaptation, they typically treat test data as independent snapshots, without constructively exploiting a rich, unique source of supervision: the significant variance in prediction quality as an agent observes the same object from different distances and viewpoints. To capitalize on this, we propose \texttt{MoCaFe}, a model-agnostic and hyperparameter-insensitive paradigm that leverages motion-induced prediction variance for unsupervised adaptation. Specifically, we first develop mobility-calibrated filtering of pseudo-labels, which constructs reliable pseudo-labels by fusing cross-view predictions using inverse-variance weights. The filtered pseudo-labels then guide the selection of ``hard samples’’ for effective learning. To further suppress pseudo-label corruption, we introduce \texttt{p-MoCaFe}, an optional extension that uses a public dataset to assign weighting factors to data samples, yielding an estimator of the clean loss with bounded bias. Extensive experiments on autonomous driving (nuScenes, KITTI) and embodied-agent datasets show that \texttt{MoCaFe} significantly outperforms state-of-the-art TTA schemes, achieving up to approximately 20% accuracy gain.

[87] Gen Through the Lens: Camera-Controllable Video Generation via Cross-Coordinate Conditioning Fusion cs.CV | cs.AI | cs.GR | cs.LGPDF

Shuai Li, Jinxu Xiang, Xinyuan Li, Hanlei Guo, Xiyang Tan

TL;DR: Gen Through the Lens 将相机内参、场景几何和目标物体 3D 框三类坐标条件融合，控制 I2V 的相机运动、焦点和构图；实验定性报告真实感与可编辑性优于现有方法。

Details

Motivation: 图像到视频生成既需遵循参考图，也需产生真实、物理连贯的相机操作。既有方法把相机运动当抽象条件，或只使用单一坐标系，难以精确控制视角、焦点和取景。

Result: Gen Through the Lens 将控制信号分成三类坐标空间：用内参描述相机坐标，用空间几何提示表示世界坐标，用目标主体 3D 边界框表示物体坐标；通过跨坐标条件融合统一引导生成过程。方法目标是在控制相机运动、焦点和取景的同时保持时空连贯。实验报告在多种相机操作下生成结果可控且视觉一致，真实感与可编辑性均超过既有方法；摘要未列具体基准或分数。

Insight: 相机、世界和物体坐标各提供不同控制参照，显式融合可减少单一坐标条件的歧义。性能结论为摘要定性报告，无法量化不同操作类型的收益或比较条件。

Abstract: We present Gen Through the Lens, a camera-controllable image-to-video generation framework that synthesizes videos from a given reference image while enabling realistic and physically consistent camera operations. Unlike previous models that treat camera motion as abstract conditioning or rely on a single coordinate system, our approach explicitly decomposes and fuses control signals across three coordinate spaces: the camera coordinate (modeled by intrinsic parameters), the world coordinate (represented by spatial geometry hints), and the object coordinate (described by 3D bounding boxes of target subjects). Through cross-coordinate conditioning fusion, these heterogeneous cues are unified to guide the generative process, enabling precise control of camera motion, focus, and framing while preserving temporal and spatial coherence. Extensive experiments demonstrate that our method produces controllable and visually consistent results across diverse camera operations, outperforming existing approaches in both realism and editability.

[88] Learning Multimodal Sensor Fusion with JEPA Representations for Breathing Activity and Pose Recognition cs.CV | cs.AIPDF

Yunxiang Luo, Constantino Álvarez Casado, Xiaoting Wu, Nhi Nguyen, Le Nguyen

TL;DR: 该工作冻结 V-JEPA2 特征并融合 RGB、深度和雷达，覆盖七种非空传感器子集；录制级呼吸活动 macro-F1 为 75.23±0.60，全子集监督比匹配 pooling 对照高 1.923 个百分点。

Details

Motivation: 预测式视频表征本身没有说明多模态系统如何处理传感器缺失。研究考察将冻结 V-JEPA2 特征迁移到异构传感器后，融合结构、子集监督和池化选择的影响。

Result: 系统融合 RGB 视频、深度视频和雷达时序，使用模态 query、固定传感器槽和显式可用性来分类呼吸活动及姿态，覆盖全部七种非空传感器组合。研究在 39 名 OMuSense-23 受试者上采用五折受试者分组验证、三随机种子，控制子集监督、池化容量和表示迁移。最终模型录制级活动 macro-F1 跨七个子集平均 75.23±0.60。每批监督所有传感器子集，相较容量匹配的 query-pooling 对照提升 1.923 个百分点，受试者 bootstrap 95%区间为[0.647,3.209]；平衡监督下 query pooling 相较匹配 mean pooling 没有明显优势。雷达探针支持迁移超越汇总小波像素；64-token 推理近似保留平均活动表现，测试的 Consensus-JEPA loss 相较严格控制未带来可测收益。

Insight: 显式学习缺失传感器子集比选择复杂池化更有明确收益，部分设计（query pooling、额外 JEPA loss）在严格控制下并无明显改善。结果基于 39 人和七种组合，不能泛化到其他传感器/任务。

Abstract: Predictive video representations do not by themselves specify how a multimodal system should handle missing sensors. We transfer frozen V-JEPA2 features to a set of heterogeneous sensors comprised by RGB video, depth video, and radar time series. A shared fusion model uses modality queries, fixed sensor slots, and explicit availability to classify breathing activity and pose from all seven non-empty sensor subsets. The contribution is a controlled evaluation of subset supervision, pooling capacity, and representation transfer on 39 OMuSense-23 subjects with five subject-group folds and three seeds. With matched downstream parameter counts, the final model reaches recording-level activity macro-F1 of $75.23%\pm0.60$ averaged across seven subsets. Supervising all subsets in each batch improves the matched query-pooling control by 1.923 percentage points (conditional 95% subject-bootstrap interval [0.647, 3.209]), while query pooling shows no clear advantage over matched mean pooling under balanced supervision. Radar probes support transfer beyond pooled wavelet pixels, 64-token inference approximately preserves mean activity performance, and the tested Consensus-JEPA loss adds no measured benefit against its exact control.

[89] VIEWCRAFT: Evolving What Multimodal Agents See and How They Read It cs.CVPDF

Binxiao Xu, Ruichuan An, Bocheng Zou, Weihua Du, Xinyu Wei

TL;DR: ViewCraft 联合进化可执行视觉证据接口与读取流程，通过失败轨迹修订、回放和配对评测保留改进；八个基准平均任务分提高 6.1 分，跨 executor 转移均优于直接执行。

Details

Motivation: 可复用技能可在不更新模型权重的情况下增强多模态智能体，但技能效果同时取决于呈现哪些视觉证据及采用什么流程解读。

Result: ViewCraft 联合进化证据接口和读取程序：接口负责选择、变换和组织观察，Reader 解释生成的视图。失败轨迹推动修订，回放及配对评估保留有用更改；对被拒绝的联合方案，接口投影可恢复其中可用的视觉编辑。八个基准上，相较直接执行，任务分数平均提高 6.1 分，且每个基准都提升；在匹配输入 token 预算下优势仍保留。五项多来源任务中，学习到的接口比初始技能少用 8%–15% 输入 token。技能跨 executor 转移时，在所有评测中仍超过直接执行。

Insight: 证据构造和证据读取需要作为耦合、可复用技能一起优化；相同 token 预算的比较说明收益不只是输入更多。6.1 是任务分数的平均提升，摘要未说明各基准指标尺度。

Abstract: Reusable skills improve multimodal agents without updating model weights, but their effectiveness depends on both the evidence presented and the procedure used to read it. We introduce ViewCraft, which jointly evolves an executable evidence interface and its reading procedure. The interface selects, transforms, and organizes observations; the reader interprets the resulting views. Failure traces guide revisions, replay and paired evaluation retain useful changes, and interface projection recovers visual edits from eligible rejected joint proposals. Across eight benchmarks, ViewCraft improves over direct execution by an average of $6.1$ points in task score, with gains on every benchmark. The advantage persists at a matched input-token budget. On all five multi-source tasks, the learned interface uses 8%-15% fewer input tokens than Initial skill. The learned skills also transfer across executors and outperform direct execution in every evaluated transfer, supporting evidence construction and reading as a coupled, reusable skill.

[90] Frequency-Step Cache: Separating Synchronization and Repair for Efficient Autoregressive Video Diffusion cs.CV | cs.LGPDF

Kerry Xu, Param Misra, Zachary Stenglein, Athri Karunamurthy, Fanjiang Ye

TL;DR: FSC 将精确同步缓存的时间点与两同步间的修复计算分开调度，以频谱漂移和 token age 风险控制整步重放/选 token 重算；SkyReels-V2 上速度版 2.36× 且 SSIM/PSNR 高于 MotionCache。

Details

Motivation: 自回归视频扩散需要反复运行大型扩散 Transformer。训练免费缓存可重用中间计算，但现有控制器常把两个决定耦合：何时进行精确同步，以及同步之间应修复多少计算。

Result: Frequency-Step Cache（FSC）在稠密到稀疏轨迹上安排完整计算 anchor，并用频带分解漂移与 token 年龄构成的轻量潜空间风险，决定同步间采用整步重放、选择性 token 重算或完全修复。SkyReels-V2 上，整步复用贡献主要加速，选择性 token 计算主要用于保真修复；受控消融显示用同一反应信号同时决定同步和修复，会在相同延迟下降低保真度，FramePack 和 Matrix-Game 上也观察到分离优势。多种 AR 视频主干上折衷改善。SkyReels-V2 的 FSC-Speed 相较 MotionCache 为 2.36× 对 2.275×，SSIM 高 0.019、PSNR 高 0.5 dB；匹配 MotionCache 速度时，FSC-Quality 的 SSIM 高 0.031、PSNR 高 1.0 dB、LPIPS 低 0.024。无需重调迁移至 SkyReels-V2 14B 时加速 2.12×，MotionCache 为 1.79×。

Insight: 缓存治理包含长期同步频率和局部修复强度两个不同控制问题，统一信号会损害等延迟保真度。速度/质量两种配置对应不同目标，SkyReels-V2 14B 结果显示无需重调仍可迁移。

Abstract: Autoregressive (AR) video generation enables long-horizon synthesis but requires repeated forward passes through large diffusion transformers. Training-free caching reduces this cost by reusing intermediate computation, yet existing controllers often couple two decisions: when to synchronize the cache exactly and how much computation to repair between synchronizations. We introduce Frequency-Step Cache (FSC), which separates these roles. FSC schedules exact full-computation anchors on a dense-to-sparse trajectory, while a lightweight latent-space risk based on frequency-resolved drift and token age controls inter-anchor repair through whole-step replay, selective token recomputation, or full repair. On SkyReels-V2, whole-step reuse provides most of the acceleration, while selective token computation is most effective as fidelity repair. A controlled ablation shows that using one reactive signal for both synchronization and repair reduces fidelity at matched latency, with the same separation advantage observed on FramePack and Matrix-Game. Across multiple autoregressive video backbones, FSC improves speed-quality trade-offs. On SkyReels-V2, FSC-Speed reaches 2.36x speedup versus MotionCache’s 2.275x while improving SSIM by 0.019 and PSNR by 0.5 dB. At MotionCache-matched speed, FSC-Quality improves SSIM by 0.031 and PSNR by 1.0 dB and lowers LPIPS by 0.024. Without retuning, FSC-Quality transfers to SkyReels-V2 14B at 2.12x speedup versus 1.79x for MotionCache.

[91] When Trackers Fail: VLM-Guided Verification and Recovery for Robust Video Object Segmentation cs.CVPDF

Valay Bundele, Susmit Agrawal, Mehran Hosseinzadeh, The Nam Nguyen, Hendrik P. A. Lensch

TL;DR: Dual-Agent 为视频分割跟踪加入时空 VLM 核验和目标恢复智能体，检测身份漂移后再继续传播；LVOS-v2/SecVOS 上 SAM3 的 J&F 分别提升 1.2/2.7。

Details

Motivation: 基于 SAM 的跟踪器通过记忆传播掩码获得强性能，但遮挡、重现或干扰物可让错误预测进入记忆并持续传播为身份漂移。现有方法常改善记忆/时序建模，却缺少明确的失败识别和恢复机制。

Result: Dual-Agent 将传播与显式核验/恢复结合。视觉语言核验智能体利用时空上下文判断跟踪是否仍指向目标；发现失败后，校正智能体根据时间上下文恢复目标，再恢复跟踪。校正器基于真实 SAM3 失败训练，并加入针对性的身份切换和干扰物扰动，覆盖恢复歧义。在 LVOS-v2 和 SecVOS 上，相较 SAM3 分别提升 1.2 和 2.7 J&F。框架无需重训可迁移至 SAM2.1、SeC、Cutie，以及 LaSOT-ext、TNL2K、VOT-LT2022 长期跟踪基准。

Insight: 错误传播需要专门的检测与回退/恢复接口，单独改进初始掩码传播不够。增益是两项视频分割基准上的 J&F，跨模型/长期跟踪迁移为作者报告的无需重训结果。

Abstract: Recent trackers built on foundation models such as the Segment Anything Model (SAM) achieve strong performance by propagating object masks through memory. However, this strength is also a key source of failure: under occlusion, reappearance, or distractors, an incorrect prediction can enter the tracking loop and be propagated over time, resulting in persistent identity drift. Existing approaches seek to prevent such failures through improved memory and temporal modeling, but once a failure occurs, the tracker has no explicit mechanism to recognize or correct it. We address this gap with Dual-Agent, a failure-aware tracking framework that complements propagation with explicit verification and recovery. We address this gap with Dual-Agent, a failure-aware tracking framework that complements propagation with explicit verification and recovery. A vision-language verification agent first reasons over spatio-temporal context to determine whether the tracker is still following the target. When a failure is detected, a correction agent uses temporal context to recover the target before tracking continues. The correction agent is trained on real SAM3 failures, augmented with targeted identity-switch and distractor perturbations that expose it to challenging ambiguities during recovery. On LVOS-v2 and SecVOS, Dual-Agent improves SAM3 by 1.2 and 2.7 $J&F$, respectively. Importantly, the proposed framework transfers without retraining to SAM2.1, SeC, and Cutie, as well as to long-term video object tracking benchmarks including LaSOT-ext, TNL2K, and VOT-LT2022. Our results show that explicitly reasoning about when tracking fails and how to recover provides a complementary path to robust tracking beyond improving propagation alone. Project page: valaybundele.github.io/Dual-Agent.

[92] PIGEONS: Incremental Long-Video Captioning with Persistent Identity Gallery and Evolving Overarching Narrative Summary cs.CVPDF

Xinlong Chen, Weihong Lin, Yuanxing Zhang, Xinyu Liu, Yue Ding

TL;DR: PIGEONS 用持久身份图库和不断更新的全局叙事摘要支持逐片段长视频描述，缓解主体称谓不一致与长程叙事碎片化；作者报告长视频准确性、完整性和主体指代一致性优于基线。

Details

Motivation: 现有视频描述多针对短片段，扩展到长视频受模型上下文限制。按段独立描述再拼接虽简单，却容易让主体称谓不一致、长程叙事断裂。

Result: PIGEONS 在生成描述前先扫描全片，建立记录关键主体及视觉参考的身份图库，作为稳定指代锚点。随后逐片段生成描述，每步都以持久身份图库和更新后的全局叙事摘要为条件；摘要紧凑保留并更新长程上下文。训练采用整合监督微调与强化学习的三阶段策略。实验显示，PIGEONS 在长视频描述准确性、完整性和主体指代一致性方面优于现有方法，同时保持短视频描述竞争力；摘要未提供具体基准或分数。

Insight: 长视频分段处理需同时维护“谁是谁”的视觉锚点和“整体发生了什么”的演化摘要，单靠拼接局部描述无法保障连贯。方法报告短视频表现也具竞争力，但未列出量化结果。

Abstract: Detailed video captioning has become an essential semantic interface for multi-modal understanding and generation. While existing methods mainly target short videos, extending them to long videos remains challenging, primarily due to limited context windows of current video understanding models. A naive solution is to adopt segment-wise captioning followed by concatenation, yet this paradigm often suffers from inconsistent subject references and fragmented long-range narratives. To mitigate these issues, we introduce PIGEONS, an incremental long-video captioning model built upon a Persistent Identity Gallery and an Evolving Overarching Narrative Summary. Before caption generation, PIGEONS scans the entire video to construct an identity gallery that captures key subjects along with their visual references, thereby providing a stable referential anchor. The model then performs clip-by-clip captioning conditioned on the persistent gallery and an evolving overarching narrative summary, which compactly maintains and updates essential long-range contextual information after each captioning step. To enable these capabilities, we develop a three-stage training strategy that integrates supervised fine-tuning and reinforcement learning. Extensive experiments demonstrate that PIGEONS markedly improves long-video captioning performance, achieving superior accuracy, completeness, and subject referential consistency over existing methods, while maintaining competitive performance on short-video captioning.

[93] Meaningful Moments: A Corpus of VLM-Derived Temporal Importance Labels for Video Action Recognition cs.CVPDF

Cameron Keith, SouYoung Jin

TL;DR: Meaningful Moments 发布约 4.58M 个 VLM 生成片段重要性分数并固定动作分类器测试时序采样；Diving-48 上重要片段快进保留全视频识别，而反向控制下降约 30 个百分点。

Details

Motivation: 识别动作时并非每帧都同等重要，少数关键瞬间可能提供决定性证据；现有视频数据集很少提供大规模、按动作条件标注的短片段重要性。

Result: Meaningful Moments（MM）包含近 50 万段带类别视频、约 458 万个逐片段 VLM 重要性分数，覆盖 SSv2、Kinetics-400、Diving-48 共 622 类。实验冻结动作分类器，只改变时间采样；alpha 参数控制重要性选择与反向选择，以区分 VLM 信号收益和变密采样本身的影响。三个数据集上 VLM 重要性均含有效信号：在 hard-cut 端点，保留重要片段优于保留 VLM 标记的 filler。效用依赖类别结构：Diving-48 上重要性快进保持全视频识别，反向控制约跌 30 个百分点；K400 上 hard-cut 比均匀采样高 1.45 点，阈值选择高 2.05 点；SSv2 有信号但未恢复全视频表现。直接评分在噪声范围内匹配贪心删除，oracle 调用约少 25 倍；四个 VLM 交叉 oracle 的下游趋势一致。

Insight: VLM 重要性分数可指导减少帧数，但收益因动作数据结构而异；SSv2 上仍未达到全视频性能。alpha 对照和反向重要性控制帮助分离语义重要性与采样密度效应。

Abstract: Not every moment in a video is equally useful for recognizing an action. A few decisive moments often carry the evidence: the instant an object moves, the peak of a dive, or one object placed on another. Existing video datasets rarely provide large-scale, action-conditioned labels for which short segments carry that evidence. This work asks whether vision-language models can identify meaningful moments at scale and whether those labels improve action recognition.We introduce Meaningful Moments (MM), a corpus of about 4.58 million VLM-generated per-segment importance scores for nearly 500,000 class-labeled videos spanning Something-Something v2 (SSv2), Kinetics-400 (K400), and Diving-48 across 622 action classes. We freeze the action classifier and vary only temporal sampling. An alpha-parameterized comparison of importance-led selection against an importance-inverted control separates gains from the VLM signal from gains caused by variable-density sampling alone. Across all three datasets, VLM-derived importance is informative: at the hard-cut endpoint, keeping VLM-important segments outperforms keeping VLM-designated filler. Usefulness depends on class structure. On Diving-48, importance-led fast-forwarding preserves full-video recognition while the inverted control collapses by roughly 30 percentage points. On K400, hard-cut selection beats uniform sampling by 1.45 percentage points, rising to 2.05 under threshold selection. On SSv2, the labels carry signal but do not recover full-video performance. Direct scoring matches greedy removal within its noise floor at roughly 25 times fewer oracle calls, and a four-VLM cross-oracle study shows consistent downstream trends. We release the importance labels and evaluation protocol under CC-BY-4.0; SSv2 class strings remain under CC-BY-NC-4.0.

[94] Geometry-Guided Semantic Reconstruction for 3D Instance Segmentation cs.CVPDF

Yuanhao Su, Shaofeng Zhang

TL;DR: 该方法以掩码语义重建将 2D 预训练先验蒸馏到 3D 实例分割编码器，并用几何调制控制语义注入；ScanNetV2/ScanNet200 全量与 few-shot 设定均达到 SOTA。

Details

Motivation: 纯 3D 编码器擅长空间结构，但外观语义弱；直接拼接 2D 先验则虽带来语义，却不保证与 3D 几何对齐。

Result: 方法将 2D 到 3D 语义迁移重构为掩码语义重建：随机遮住部分背景 lifted 2D 先验，训练 3D 编码器恢复缺失语义，显式将 2D 语义知识蒸馏至 3D 主干。Geometry-Modulated Semantic Injection 在稀疏编码前用原始 3D 几何调制提升的 2D 先验，让几何控制语义注入，减小 2D 语义特征与 3D 几何特征的分布差异。两个基线框架实验显示，在 ScanNetV2 和 ScanNet200 的全量数据及 few-shot 设置下均达到所比较方法中的 SOTA；摘要未给具体分数。

Insight: 语义迁移需要显式几何对齐监督，而非简单特征拼接；掩码恢复提供蒸馏目标，几何调制控制注入。结果覆盖两个框架、两个数据集及全量/few-shot 设置，量化收益未见摘要。

Abstract: 3D instance segmentation requires representations that are both geometrically structured and semantically discriminative. Pure 3D encoders model spatial structure well but often lack strong appearance semantics, while directly lifting 2D priors into 3D by concatenation injects rich semantics without ensuring that they align well with the underlying 3D structure. In this paper, we address this problem by formulating 2D-to-3D semantic transfer as a \textbf{masked semantic reconstruction} task. Specifically, we mask a subset of background lifted 2D priors and train the 3D encoder to recover the missing semantic priors, providing explicit supervision for distilling rich 2D semantics into the 3D backbone. We further propose \textbf{geometry-modulated semantic injection}, which uses raw 3D geometry to modulate lifted 2D priors before sparse encoding, so that geometry explicitly controls semantic injection and alleviates the distribution gap between 2D semantic features and 3D geometric features. Experiments on two baseline frameworks demonstrate that the proposed method achieves state-of-the-art results in both full-data and few-shot settings on ScanNetV2 and ScanNet200. Code is publicly available at https://github.com/yharoldsu0627/Geometry_Guided_Semantic_Reconstruction

[95] Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding cs.CV | cs.LG | cs.MMPDF

Shuo Gao, Chenhao Zheng, Zhongzheng Ren

TL;DR: FramePack-U 将任意长视频按提示相关性与视觉多样性选帧，并以不同分辨率打包进固定长度上下文，实现常量视觉 token 内存；Qwen3.5-35B-A3B 上三项长视频基准提高 12.9/10.6/9.7%。

Details

Motivation: 逐帧编码使 VLM 输入 token 数随视频长度线性增长，长视频造成显存和推理延迟瓶颈。token 剪枝或关键帧选择虽缩短序列，仍保留随时长 O(n) 增长。

Result: FramePack-U 是免训练、无需改模型的常量内存帧打包框架。它先按帧与文本提示的语义相似度评分，并以多样性驱动排序，再将选中帧按不同分辨率装入固定长度视觉上下文，令预算集中于指令相关且视觉多样的帧。作者指出理论上可打包任意帧数，并给出跨设置较稳健的经验实例。LongVideoBench、MLVU、Video-MME 上，方法在不同模型家族、规模和帧预算下优于全分辨率基线、既有 token 压缩及关键帧方法。Qwen3.5-35B-A3B、128 帧预算时，相对官方模型三基准分别提升 12.9%、10.6%、9.7%。

Insight: 常量上下文预算下联合控制帧相关性、多样性与分辨率，可从根本上避免随视频长度线性增加视觉 token。具体提升对应 Qwen3.5-35B-A3B 的 128 帧设置；摘要未说明百分比口径细节。

Abstract: Hour-long videos, streaming feeds, and movie-length content are pushing vision-language models (VLMs) past their breaking point – the standard recipe of encoding each frame into visual tokens yields sequences that grow linearly with video length, causing prohibitive memory consumption and inference latency. Existing approaches, such as token reduction and keyframe selection, shorten the token sequence but leave the O(n) scaling intact, failing to resolve the fundamental bottleneck. We propose FramePack-U, a strikingly simple, training-free frame packing framework that compresses arbitrarily long videos into a fixed size visual representation at constant memory cost. FramePack-U requires no fine-tuning or architectural modification – it plugs directly into off-the-shelf VLMs at inference time. Inspired by FramePack from the generation literature, we propose a content-aware compression strategy for visual understanding under a fixed token budget. Specifically, FramePack-U (1) scores each frame by its semantic similarity to the text prompt and ranks frames in a diversity-driven manner, and (2) packs the selected frames at varying resolutions into a fixed-length context window. This concentrates the memory budget on frames that are both instruction-relevant and visually diverse. While FramePack-U can in theory pack arbitrarily many frames into the fixed budget, we identify an empirical instantiation that generalizes robustly across settings. Extensive experiments on LongVideoBench, MLVU, and Video-MME show that FramePack-U consistently outperforms full-resolution baselines and prior token reduction and keyframe selection methods across diverse model families, parameter scales, and frame budgets. Notably, on Qwen3.5-35B-A3B with a 128-frame budget, FramePack-U yields gains of +12.9%, +10.6%, and +9.7% over the official model on the three benchmarks, respectively.

[96] A Concept-informed Interpretable Artificial Intelligence Framework for Breast Ultrasound Diagnosis cs.CVPDF

Shengzhou Zhong, Ziying Wang, Liang Li, Lihua Yan, Shuochun Chen

TL;DR: IBUS 收集 9,941 名患者、跨 20 个数据集的超声图像和结构化 BI-RADS 描述，并在 14 个测试数据集验证跨中心表现；八名放射科医生参与研究显示辅助系统提高诊断敏感度。

Details

Motivation: 乳腺超声 AI 多依据图像直接预测，较少纳入 BI-RADS 知识及放射科医生中间评估过程，因此可解释性和临床流程对齐有限。

Result: 作者构建多中心乳腺超声图像与结构化 BI-RADS 描述数据集，含 9,941 名患者、20 个独立数据集。IBUS 在 14 个测试数据集上表现出跨临床中心和成像条件的稳定泛化。BI-RADS 描述支持细粒度风险分层；将预测 BI-RADS 分级与恶性概率交叉核对，可发现内部不一致、可能需额外临床复核的预测。八名放射科医生读片研究中，IBUS 辅助提高初级和资深医生的诊断敏感度，并可能帮助有针对性地修订风险分层。作者将其定位为补充放射科医生、提供额外风险估计与可解释证据的第二阶段决策支持系统；摘要未给敏感度提升数值。

Insight: 结构化 BI-RADS 表征既用于风险分层，也可作为模型自身预测的核验维度，读片研究提示辅助潜力。摘要未报告各中心指标或读者研究具体效果，不能据此断言可替代临床判断。

Abstract: Most existing artificial intelligence systems for breast ultrasound rely primarily on image-driven prediction, with limited incorporation of the knowledge from the Breast Imaging Reporting and Data System (BI-RADS) and the intermediate assessment processes underlying radiologists’ workflow, thereby constraining their interpretability and alignment with clinical practice. We curated, to our knowledge, the largest multicenter collection of breast ultrasound images paired with structured BI-RADS descriptors, comprising 9,941 patients across 20 independent datasets. Across 14 testing datasets, IBUS demonstrated consistent generalization performance across different clinical centers and imaging conditions. Incorporating BI-RADS descriptors enabled fine-grained risk stratification, while cross-checking predicted BI-RADS grades against malignancy probabilities identified internally discordant predictions that may warrant additional clinical review. In a reader study involving eight radiologists, IBUS assistance increased diagnostic sensitivity among both junior and senior readers, with the potential of guiding radiologists to make targeted revision in risk stratification. Overall, IBUS exhibits its potential as a second-stage decision-support system that complements radiologists by providing additional risk estimates and interpretable evidence for diagnostically challenging cases. The demo video of IBUS platform is publicly available at https://github.com/zsz1997/IBUS.

[97] Unravel: Recovering an Interactive 3D Scene from a Single Image cs.CV | cs.AIPDF

Hui Shan, Keyang Luo, Binjia Zhou, Ming Li, Kai Li

TL;DR: Unravel 用迭代物体剥离、遮挡区编辑重建和逆序场景组装，从单图生成可独立移动物体的交互 3D 场景；作者报告物体恢复、位置忠实度和物理合理性均优于基线。

Details

Motivation: 单图生成器常输出纠缠在一起、不能单独移动的几何；一次前馈预测所有物体时，错误猜测又难以发现或修复。单视角图像对遮挡物信息有限。

Result: Unravel 是无需训练的单图到交互 3D 场景框架。智能体按层迭代剥离输入：VLM 规划哪些物体完全可见，图像编辑器删除这些物体并补全被遮区域，每次编辑在纳入重建前先核验，使遮挡物逐步暴露为直接观察。随后按剥离顺序反向组装场景，使支撑物先于其上物体放置。单视角点云被视为物体尺寸下界：资产可扩展以覆盖已观察点，但不能缩小到遗漏这些点。最后进行物理稳定化以移除悬浮和互穿，再由智能体审查剩余摆放错误。室内场景实验报告相较现有方法找回更多物体、摆放更忠实且物理合理性更好；摘要未列指标。

Insight: 迭代编辑把原先被遮挡部分变成可观察证据，再从底层支撑关系反向组装，减少一次生成时的猜测错误。结果为多样室内场景上的定性比较，没有具体分数或人工评价方法。

Abstract: We present Unravel, a training-free framework that turns a single image into an interactive 3D scene, where every object is an independent and movable asset. Existing generators either produce one entangled geometry in which no object can be moved, or predict all objects in one feed-forward pass, where a wrong guess cannot be detected or repaired. Unravel instead takes a different approach: it first takes the scene apart, and then puts it back together. Starting from the input image, an agentic harness that iteratively peels the image layer by layer, and adopts point cloud as major modality for spatial relationship recovery. Specifically, a vision-language model plans which objects are fully visible, an image editor removes them and restores the occluded regions. Every edit is verified before being incorporated into the reconstruction. This process gradually turns occluded parts of the scene into direct observation, allowing objects to be reconstructed when they are best exposed rather than hallucinated from the original image alone. The scene is then rebuilt in reverse removal order, so that the supporting objects are naturally placed before the objects they support. Since a single-view image only reveals part of an object, we treat the observed points as a lower bound on its size: an asset may grow to cover its points, but can not shrink away from them. Finally, we physically settle the reconstructed scene to remove floating objects and interpenetrations, followed by an agentic audit that corrects remaining placement errors. Extensive experiments on diverse indoor scenes show that Unravel recovers more objects, places them more faithfully, and produces scenes that are more physically-plausible than existing methods.

[98] Esports World Model Dataset cs.CVPDF

Weiyu Ma, Zeyuan Gong, Xuhui Liu, Wenxuan Zhang, Jian Ding

TL;DR: EWM 收集 StarCraft II、Dota 2、CS2 的 3,053 场比赛/地图和约 4,880 玩家视角小时，并将视频、行动、状态和事件按 replay 时钟对齐，可研究多玩家共享世界与私有观察。

Details

Motivation: 交互世界模型需学习多个参与者如何共同改变共享环境，同时每位玩家观察到的世界不同。战争迷雾、团队视野和相机位置为研究不同可见性与共同状态提供具象数据。

Result: Esports World Model Dataset（EWM）包含 StarCraft II、Dota 2、Counter-Strike 2 真人对局，合计 3,053 场游戏/地图、约 4,880 玩家视角小时（CS2 仍在扩展，含估计时长）。数据包括 4,628 个完成的 StarCraft II 视角、303 场 Dota 2 的全部十个视角，以及 CS2 每回合最多十个视角。回放时钟把画面与同一比赛的游戏原生动作、结构化状态和事件对齐。三款游戏分别提供私有战争迷雾、共享团队视野和第一人称视角。StarCraft II 另有玩家特定实体记录，区分当前观测与记忆快照并连接对立视角共同观察实体，帮助分辨合法信息差异与对共享状态的认知错误。

Insight: 多视角同步状态记录允许研究者区分真实部分可观测与模型对同一世界状态判断不一致。EWM 面向游戏环境，CS2 时长含估计数据，应用于现实机器人/一般世界需额外迁移。

Abstract: Interactive world models must capture how multiple participants jointly change a shared environment while observing different parts of it. Competitive games make this challenge concrete: players cooperate and compete, while fog of war, team vision and camera placement determine what each can see. Modeling these interactions requires data that connect individual observations and actions to shared state and events. We introduce the Esports World Model Dataset (EWM), a collection of human gameplay spanning StarCraft II, Dota 2 and Counter-Strike 2, with 3,053 games and maps and approximately 4,880 player-view hours, including estimated hours for the ongoing CS2 extension. It includes 4,628 completed StarCraft II views, all ten perspectives for 303 Dota 2 matches, and up to ten views per round in CS2. Across these views, replay-clock alignment connects rendered video with game-native actions, structured state and events from the same match. The three games expose complementary observation conditions: private fog of war, shared team vision and first-person viewpoints. In StarCraft II, player-specific entity records further distinguish live observations from remembered snapshots and link co-observed entities across opposing views. Together with temporal alignment, these records allow researchers to distinguish legitimate information differences from disagreement about shared state. By connecting human decisions, shared dynamics and player-specific information, EWM supports research toward jointly controlled multiplayer world models and interactive environments populated by teammates and opponents learned from human play.

[99] All-St3R SLAM: Real-time RGB Dynamic SLAM with Pixel-wise Motion Masking cs.CVPDF

Yewon Kim, Paul Hongsuck Seo, Sang Chul Ahn

TL;DR: All-St3R 将 DUSt3R 预测的自运动像素与 All-Tracker 点跟踪运动比较，只遮罩实际移动像素；RGB-only 在约 10 FPS 运行，比密集动态 SLAM 的 0.5–3 FPS 更快。

Details

Motivation: 动态环境下 SLAM 必须同时定位和建图，但已有动态 SLAM 常过度删除静态部分或计算昂贵，依赖固定语义类别也不适合所有环境。

Result: All-St3R 提出实时 RGB 动态 SLAM，不使用预定义标签或分割模型。系统用 DUSt3R 根据自运动预测像素运动，并用 All-Tracker 跟踪点，两类运动来源的差异用于定位真正运动的像素并仅移除它们。与整对象屏蔽相比，当前静止的物体部分仍可用于跟踪。系统仅使用 RGB，约 10 FPS；对标准 30 FPS 视频每三帧采样一次，从而达到实时处理。作者报告相比现有密集动态 SLAM（0.5–3 FPS）有明显速度优势。

Insight: 逐像素比较几何自运动预期与实测点运动，可减少动态对象整体屏蔽造成的信息浪费。10 FPS 依赖每三帧取样，实际处理速率与输入 30 FPS 不同，速度比较应注意这一设置。

Abstract: SLAM (Simultaneous Localization and Mapping) is a core technology in autonomous systems that enables them to estimate their own position and build a map of unknown environments simultaneously. SLAM systems in dynamic environments are of great practical importance. However, existing dynamic SLAM approaches still have several limitations like over-removal and high computational costs. To address these limitations, we propose a real-time RGB Dynamic SLAM framework that precisely removes moving objects at the pixel level without relying on predefined labels or segmentation models. Our system combines pixel motion predicted by ego-motion using DUSt3R with point tracking from All-Tracker, removing only actually moving pixels by analyzing discrepancies between the two motion sources. Unlike object-level masking methods, our approach eliminates only currently moving regions, allowing temporarily static parts to be utilized for tracking. Operating at approximately 10 FPS, the system achieves real-time performance on standard 30 FPS video streams by subsampling every 3 frames. Relying solely on RGB inputs, this significantly improves upon state-of-the-art dense dynamic SLAM approaches (0.5-3 FPS), presenting a highly practical solution.

[100] Geometric Parrots: Rule Absence vs. Rule Fragility in Spatial and Physical Reasoning for Vision-Language Models cs.CVPDF

Bilal Jawaid, Hamza Hanif, Muhammad Maaz Uddin, Anas Zafar, Inad Alqurashi

TL;DR: GRIP 用程序生成的场景和双环评测，区分视觉语言模型在几何、空间与物理推理中的规则缺失，以及受到错误挑战后放弃正确答案的规则脆弱性。

Details

Motivation: 单轮准确率无法判断模型改答案是纠正错误还是屈从于错误提示；现有评测也难以定位模型究竟在哪一层空间或物理推理上失效。

Result: 作者构建 GRIP，覆盖 34 个程序生成推理域、10 万张图像和 60 万个问题，答案由场景状态直接确定。双环评测先用五步组合问题梯定位错误，再通过同伴模型答案或无依据的用户断言进行受控挑战，并将改答分为错转对的纠正和对转错的屈从。摘要未报告具体模型分数或实验比较数值。

Insight: 确定性场景真值使评测能分开考察能力缺失与交互中的抗压稳定性，也避免把所有改答一概视为错误。不过这是程序生成基准，其覆盖范围和挑战形式能否代表真实对话中的多样情境，仍需结合后续验证。

Abstract: Vision-language models (VLMs) can answer many geometric and physical reasoning questions correctly, yet still behave inconsistently in interaction, for example, abandoning a correct answer when confronted with a confident but incorrect user assertion. Standard single-turn accuracy cannot distinguish such capitulations from beneficial corrections because it scores each response in isolation. We introduce GRIP, a controlled benchmark with 34 procedurally generated reasoning domains, 100,000 images, and 600,000 questions spanning geometric, spatial, visual, and physical reasoning. Each scene is generated by deterministic code, and every answer is derived directly from the underlying scene state, providing exact ground truth without human annotation. On GRIP, we propose Dual-Loop Evaluation. A closed loop probes each image with a five-step compositional question ladder to localize where reasoning fails. An open loop then re-elicits the model’s answer under controlled challenges, including peer-model responses and unsupported user assertions. Because ground truth is known exactly, answer changes can be signed as corrections (wrong-to-right) or capitulations (right-to-wrong), separating two failure modes: rule absence and rule fragility. The dataset is available at https://huggingface.co/datasets/bilaljawaid980/GRIP-Benchmark, with further details at https://bilaljawaid980.github.io/grip-benchmark/.

[101] VGGA: Geometry-Grounded Autoencoder for Video Generation cs.CVPDF

Chaofan Gan, Yian Zhao, Shijie Li, Tieyuan Chen, Zicheng Zhao

TL;DR: VGGA 将深度、运动对应和投影变形等几何约束加入视频自编码器，意在让潜变量表示动态场景结构，并改善后续视频生成。

Details

Motivation: 视频生成常用自编码器主要按像素重建优化，缺少结构化几何监督。作者主张视频是动态三维世界的二维投影，单纯重建信号不足以促使潜表示编码场景结构与运动。

Result: VGGA 将结构深度、运动对应和投影变形三个目标整合进视频几何解码器，通过共享时空上采样主干联合优化几何学习与 RGB 重建。作者报告其在 RealEstate10K 上 gFVD 降低 28.1%，在 Taichi-HD 上降低 37.7%，同时重建质量与基线相当。摘要未列出绝对 gFVD 数值或更多生成指标。

Insight: 几何监督能在维持像素重建质量的情况下改善生成指标，支持把潜空间学习目标从外观重建扩展到结构与运动。当前结论来自摘要报告的两个数据集和 gFVD，相对降幅不等同于所有视频类型上的普遍提升。

Abstract: Autoencoders are the core component of latent video generative models, compressing high-dimensional videos into compact latent spaces for downstream generation. However, existing video autoencoders are primarily optimized for pixel-level reconstruction, providing limited supervision for learning structured latent representations. We argue that videos should not be treated merely as signals to be reconstructed, but rather as 2D projections of dynamic 3D worlds that exhibit explicit and well-structured geometric properties. Building on this insight, we propose \textbf{V}ideo \textbf{G}eometry \textbf{G}rounded \textbf{A}utoencoder (\textbf{VGGA}), a geometry-aware video autoencoder that explicitly incorporates geometric constraints into the reconstruction process. Specifically, VGGA grounds the learned representations along three complementary geometric dimensions: \emph{Structural Depth}, \emph{Motion Correspondence}, and \emph{Projective Warping}. These objectives are unified within a Video Geometry Decoder, which employs a shared spatio-temporal upsampling trunk to jointly optimize geometric learning and RGB reconstruction. Through these geometry-grounded objectives, VGGA encourages the latent representations to capture the underlying 3D structure and motion of dynamic scenes. VGGA consistently improves video generation across both geometry-centric and human-motion domains, reducing gFVD by 28.1% on RealEstate10K and 37.7% on Taichi-HD. Notably, these substantial generation improvements are achieved while maintaining comparable reconstruction quality, demonstrating that geometry-grounded representation learning yields a more effective latent space for video generation.

[102] DIGS: Distribution-Informed Gaussian Splatting for Training-Free Open-Vocabulary 3D Segmentation cs.CVPDF

Zhifan Zeng, Wenbo Xiao, Arcot Sowmya, Changming Sun

TL;DR: DIGS 以训练免除的多标签分布表示和实例级视觉—文本融合，改善开放词汇 3D 高斯分割中边界高斯被迫单标签化造成的实例边界不稳。

Details

Motivation: 常见 3D 高斯语义分割要么逐场景梯度训练、耗时约 1–4 小时，要么闭式蒸馏；两者通常让每个高斯绑定单一特征，使跨越多个 2D 实例的边界高斯难以正确检索和渲染。

Result: DIGS 分两阶段处理：先进行类别无关映射，在每个高斯上用温度化 Dirichlet 更新保留 top-M 标签分布，并通过双向状态机汇总为持久实例本体；再融合实例独立 3D 足迹的 CLIP 视觉嵌入与 LLM 属性描述的 CLIP 文本嵌入。单张 A100 上，在 LERF-OVS 每场景映射和编码需 1.0–3.2 分钟（取决于是否启用 LLM 融合），原生 3D 查询耗时 5.4 毫秒；LERF-OVS 平均 mIoU 为 71.95%，超过文中比较的免训练 SFS 和训练式 LaGa。作者还报告其在 LERF-Mask、3D-OVS 上平均 mIoU 最高，并在各自评测协议下超过 Replica 上的 Gaga 和 ObjectGS。

Insight: 边界处保留多峰类别分布，延迟到渲染时再决策，是缓解单高斯硬标签导致边界破碎的关键；实例级图文融合则帮助区分外观相似对象。运行时间是摘要报告的单 A100、按场景指标，跨数据集比较应遵循各自协议；摘要未列出后两个基准的具体数值。

Abstract: Open-vocabulary 3D Gaussian segmentation typically binds semantic features to the per-Gaussian state via per-scene gradient training (1-4 h) or the more recent closed-form distillation. Both commit each primitive to a single feature; therefore, boundary Gaussians spanning multiple 2D instances are forced into a hard decision that degrades retrieval accuracy and yields unstable mask boundaries at render time. We present DIGS, a training-free framework with two stages, class-agnostic mapping and multi-modal encoding, that decouples instance discovery from linguistic naming. At the Gaussian level, our distribution-informed representation maintains an explicit top-$M$ label distribution per primitive, updated by a tempered Dirichlet rule; preservation of multi-peaked label distributions at boundary primitives allows a late-argmax renderer to recover clean instance boundaries. A bidirectional state machine aggregates these distributions into a persistent class-agnostic instance ontology. The instance-level encoding stage then fuses the CLIP visual embedding of each instance’s isolated 3D footprint with the CLIP-text embedding of an LLM-generated attribute description, disambiguating visually similar instances. On a single A100, DIGS completes mapping and instance encoding on LERF-OVS in 1.0-3.2 min per scene, depending on whether LLM fusion is enabled, and serves native 3D queries in 5.4 ms. It reaches 71.95% mean mIoU on LERF-OVS, outperforming both the training-free SFS and training-based LaGa baselines. DIGS also achieves the highest mean mIoU on LERF-Mask and 3D-OVS and surpasses Gaga and ObjectGS on Replica under their respective evaluation protocols. Code is available at https://github.com/ZephyrZos/DIGS.

[103] EventLens: Event-Structure Reinforcement Learning for Video Understanding cs.CV | cs.AI | cs.MMPDF

Guowen Zhang, Boshen Xu, Zihao Yue, Ziheng Wang, Xiaokun Liu

TL;DR: EventLens 将视频理解训练目标从复述字幕转向恢复事件结构，以可验证的事件分割、进展判别和跨事件关系任务进行强化学习。

Details

Motivation: 视频问答模型虽然能回答不少问题，却常合并相邻事件、忽略单个事件的变化过程，或推断不合理的事件关系。作者认为字幕式监督容易让模型依赖语言线索，而非基于视觉证据理解时间结构。

Result: EventLens 定义三级时间本体，并在时间粒度变化、进展反转、事件顺序打乱等受控扰动下训练三类任务：事件分割、进展判别和多事件关系推理。任务可确定性评分，因此以 GRPO 训练专门策略，无需人工偏好标签或学习式奖励模型；再用多教师在线策略蒸馏（MT-OPD）合并为统一模型。在 Qwen3-VL 骨干上的实验报告称，时间定位和事件敏感推理基准持续改善，同时保持一般视频问答表现；学习曲线显示，相较直接混合强化学习，达到相近下游迁移效果所需训练成本显著降低。摘要未提供具体分数或成本比例。

Insight: 核心做法是直接监督事件结构，并用确定性任务奖励减少对偏好标注和奖励模型的依赖。结果目前只能按摘要中的定性结论表述；具体增益、成本口径及跨骨干的普适性需查看实验表格确认。

Abstract: Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent. They may merge neighboring events, miss single-event progression, or infer implausible relations across events. We study structured temporal understanding as the ability to recover event structure from visual evidence, including event boundaries, single-event progression, and multi-event relations. We argue that this gap arises from a mismatch between current post-training objectives and temporal reasoning, where caption-style supervision allows models to rely on language shortcuts rather than visual temporal grounding. To address this, we propose EventLens, a vision-centric reinforcement learning framework that learns video understanding through event-structure recovery. Instead of reconstructing captions, EventLens trains models to recover temporal structure under controlled perturbations, such as altered temporal granularity, reversed progression, and shuffled event order. We instantiate this framework with a three-level temporal ontology and derive three verifiable task families: event segmentation, progression discrimination, and multi-event relation reasoning. These tasks admit deterministic rewards, enabling task-specialized GRPO training without human preference labels or learned reward models. We further introduce multi-teacher on-policy distillation (MT-OPD) to consolidate specialized policies into a unified model. Experiments on Qwen3-VL backbones show consistent improvements on temporal grounding and event-sensitive reasoning benchmarks, while preserving general video-QA performance. Learning curves further demonstrate that EventLens achieves comparable downstream transfer with significantly reduced training cost compared to direct mixed RL optimization.

[104] Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation cs.CV | cs.AIPDF

Zixu Wang, Xiaoran Zhao, Yahui Liu, Feng-ao Wang, Junhao Xiao

TL;DR: GenRevisit 让图像生成智能体同时选择从哪张历史候选图继续编辑、以及采取什么编辑动作，并用 NACA 将轨迹奖励分配到具体动作。

Details

Motivation: 迭代编辑并非每一步都变好，但既有智能体通常只能从最新图像继续，可能在质量下降后仍沿着退化结果优化。只依据最终结果给整条轨迹同一优势，也难以区分各步动作的贡献。

Result: GenRevisit 通过监督微调后进行智能体强化学习，联合决策继续状态和编辑动作。作者提出归一化动作信用分配（NACA），依据各动作对减少剩余目标的贡献，重新分配 GRPO 的正向优势。在 GenEval2 上，使用 Qwen-Image-2512 和 Z-Image-Turbo 时 GM 分别为 84.3、74.4；相对使用同一生成器的最强基线分别高 12.7、10.9 分。摘要还称在 GenEval、GenEval++、GenEval2 上优于所比较基线。

Insight: 显式回看并选择较好的历史状态，可缓解只从最新编辑结果前进造成的非单调退化；NACA 则试图改善稀疏终局奖励的动作级归因。报告优势是在指定生成器及基准设置下，摘要没有交代分数统计细节或更广泛的生成质量与成本权衡。

Abstract: Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and editing, adapting their actions based on interaction history. However, they typically fix the continuation state (i.e., the image from which refinement proceeds) to the latest image. Since iterative editing is not necessarily monotonic, the agent may continue refining from the latest image even when it has already degraded. In this work, we introduce GenRevisit, an agentic framework that extends action selection to joint continuation-state and action selection. The agent can not only use interaction history to adjust editing instructions but also select an image candidate to edit, jointly deciding where to continue from and how to refine. We train GenRevisit with SFT followed by agentic reinforcement learning. To address the coarse credit assignment of final outcome reward in GRPO, which assigns the same trajectory-level advantage to all actions, we further propose Normalized Action Credit Assignment (NACA) to redistribute positive advantage according to each action’s contribution to reducing the remaining objective. On GenEval2, GenRevisit achieves GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, surpassing the strongest baselines using the same generators by 12.7 and 10.9 points respectively. GenRevisit achieves state-of-the-art performance on GenEval, GenEval++, and GenEval2 among baselines.

[105] Seg2Click: Learning Pixel-Level Action Masks for GUI Visual Grounding cs.CVPDF

Jingchao Wang, Dingjiang Huang

TL;DR: Seg2Click 将 GUI 指令定位改写为可执行点击区域掩码预测，再以确定性规则选出点击点，减少单点监督和视觉 token 网格对细小目标定位的限制。

Details

Motivation: 拥挤界面中，一个控件往往有多个有效点击位置，单点坐标监督把区域压缩为一个点；基于视觉 patch 的定位又受 token 覆盖范围限制，可能混入邻近控件或背景。

Result: Seg2Click 使用边界框监督预测可执行区域掩码，并通过确定性规则从中生成最终点击点，不需要验证器或二次放大。作者在两个模型家族的五种骨干配置上评测；Seg2Click-9B 在 ScreenSpot、ScreenSpot-v2、ScreenSpot-Pro 上分别为 92.06%、94.10%、55.22%。匹配骨干对比显示其优于坐标生成和 patch 定位基线；摘要未给出各匹配比较的具体差值。

Insight: 预测一片合法点击区域比强制拟合唯一坐标更符合 GUI 操作的空间容差，也能突破粗粒度视觉 token 覆盖限制。ScreenSpot-Pro 分数明显低于另两个基准，说明复杂界面仍具挑战；摘要未说明这些百分比是否为统一准确率定义。

Abstract: Visual grounding is a central challenge for GUI agents, requiring an instruction to be mapped to a target within a crowded screenshot. Coordinate generation with single-point supervision reduces a spatially extended target to one position, despite the presence of many valid click locations. Region-based patch scoring alleviates this ambiguity, but its spatial decisions remain tied to visual-token footprints that may span the target, neighboring controls, and background. To address these limitations, we propose Seg2Click, which reformulates GUI grounding as executable-region prediction: identifying where a click can validly execute the given instruction. Using bounding-box supervision, Seg2Click predicts an executable-region mask and computes the final click point through a deterministic rule. This formulation captures the spatial extent of valid click locations and enables finer localization beyond the visual-token grid, making it well suited to small, densely arranged controls without requiring a verifier or zoom-in pass. Experiments across five backbone configurations from two model families demonstrate the effectiveness of Seg2Click. Seg2Click-9B achieves 92.06%, 94.10%, and 55.22% on ScreenSpot, ScreenSpot-v2, and ScreenSpot-Pro, respectively, while matched-backbone comparisons show improvements over both coordinate generation and patch-based grounding. Code will be available at https://github.com/jcwang0602/Seg2Click.

[106] Learning Temporal Evidence Selection for Video Anomaly Understanding cs.CVPDF

Hangil Park, Seungheon Kim, Tae-Kyun Kim

TL;DR: 该方法把视频异常理解中的时间采样作为面向解释生成的证据选择问题，提出用视频对象语义条件化的可学习采样器。

Details

Motivation: 以异常分数或固定规则挑帧虽能集中观察可疑时刻，却可能漏掉上下文、对象交互和分散在时间轴上的证据，影响对异常事件的完整解释。

Result: 模型提出语义条件化 Temporal Sampling Q-Former，根据帧级视觉表征计算软时间选择权重，并合并重复选帧。视频中自动提取的对象标签经 CLIP 编码后仅用于条件化采样，不拼入 LLM 的生成提示。在 HIVAU-70K 上，作者称其优于所比较的微调模型；消融显示可学习采样优于固定异常导向采样，对象标签条件也优于通用提示条件。摘要未报告具体指标数值。

Insight: 证据选择直接由异常解释目标训练，有助于从“哪里异常”转向“哪些片段能支撑解释”。摘要提供的是定性对比，具体增益、标签抽取误差影响及其对不同异常场景的适用性未说明。

Abstract: Video anomaly understanding (VAU) aims to generate language-level interpretations of anomalous events in long videos, requiring visual evidence that supports explanation beyond anomaly detection alone. Existing VAU pipelines commonly rely on anomaly-centric or predefined temporal sampling strategies, which efficiently emphasize suspicious moments but overlook contextual cues, object interactions, and temporally distributed evidence needed for reasoning. Motivated by recent progress in learnable temporal selection for long-video question answering, we revisit VAU temporal sampling as a downstream-aligned evidence selection problem. We introduce the semantic-conditioned Temporal Sampling Q-Former that learns to select temporal visual evidence through the VAU generation objective. Since generic VAU prompts provide weak video-specific guidance, we condition the sampler on CLIP-embedded object tags automatically extracted from the input video. These tags are used only for temporal evidence selection and are not appended to the LLM generation prompt, separating sampler conditioning from prompt augmentation. The proposed sampler operates in visual-token space, computes soft temporal selection weights from frame-level representations, and merges duplicated selections before response generation. Experiments on HIVAU-70K show that our model achieves the best performance among the compared fine-tuned models, while ablations demonstrate that learnable temporal sampling improves over fixed anomaly-focused sampling and that CLIP-embedded object-tag conditioning further improves over generic-prompt conditioning. Our results suggest that temporal sampling in VAU should be aligned with explanation-oriented evidence utility rather than anomaly saliency alone.

[107] 3D Gaussian Splatting for Fine-grained Urban Scene Reconstruction via Semantic-guided Density Control and Loss Design cs.CVPDF

Yueling Liao, San Jiang, Xing Zhang, Zhongwen Hu

TL;DR: SG-GS 将 SAM 语义区域用于指导 3D 高斯的密度控制与分阶段损失设计，以改善复杂城市场景中细小结构的重建。

Details

Motivation: 标准 3D Gaussian Splatting 的密度控制较粗，损失函数也不利用语义信息，容易在复杂城市场景的线状结构和区域边界处产生伪影或模糊。

Result: 方法以 SAM 推导的语义分割作为约束：按场景语义区域动态调整梯度阈值，并对边界采用增强致密化，使新增高斯更集中于结构关键处；同时设计多特征引导损失，并通过渐进训练逐步加入细粒度监督。在三个数据集上，作者报告重建精度持续提升；相较原始 3DGS，PSNR 约提高 2.7%–5.2%，线状结构伪影和边界模糊得到缓解。摘要未给出数据集名称、绝对分数或额外消融结果。

Insight: 该工作把语义信息同时注入高斯增密位置和优化目标，针对结构边界的表示能力，而非仅替换渲染器。摘要结论限于三个未指名数据集及相对 PSNR 提升，复现和跨场景比较还需查看正文中的数据与协议。

Abstract: 3D Gaussian Splatting (3DGS) has been becoming a promising 3D reconstruction technique because of its advantage of high-fidelity rendering effect with high efficiency. However, it still suffers from coarse density control and semantically blind loss functions when reconstructing the fine-grained structures in complex urban scenes. Thus, this study proposes SG-GS, a semantic-guided 3DGS method for fine-grained structure reconstruction. It takes the semantic segmentation deduced by the Segment Anything Model (SAM) as a constraint and systematically integrates it into both Gaussian density control and loss function design. On the one hand, a semantic-constrained adaptive density control mechanism is constructed, which dynamically allocates gradient thresholds to different semantic regions of the scene and introduces an additional boundary-enhanced densification strategy, so that new Gaussians are directed to structurally critical regions. On the other hand, a multi-feature guided loss function is designed, and fine-grained supervision signals are gradually introduced through a phased progressive training strategy, alleviating the artifacts caused by the lack of semantic constraints in the basic loss. Experiments on three datasets show that the proposed method consistently improves reconstruction accuracy. Compared with the original 3DGS, the PSNR increases by approximately 2.7% to 5.2%, while artifacts on linear structures and boundary blurring are effectively mitigated. The method can serve the fine-grained 3D reconstruction of urban scenes.

[108] Anomaly-Guided Observation and Two-Level Semantic Feedback for Training-Free Video Anomaly Analysis cs.CVPDF

Hangil Park, Seungheon Kim, Tae-Kyun Kim

TL;DR: 该免训练异常分析框架先用异常分数引导主动选帧并生成描述，再把描述作为视频级与局部语义反馈用于修正时间异常分数。

Details

Motivation: 冻结视觉语言模型和大语言模型的异常检测方法虽会利用语义上下文，但通用上下文未必突出异常，也缺少明确机制将视频级描述转化为局部异常分数调整。

Result: 方法从上游异常分数曲线出发，以高分片段初始化观察，并结合异常分数与视觉语言模型的相关性判断继续探索时间片段。冻结模型基于选中帧生成一次开放式异常描述，再以两层反馈修正检测：视频级决定抑制、保留或锐化分数分布并选择性复查；局部层检索相关时间窗，经目标局部视觉验证后才提升分数。模型权重全程冻结，反馈不重新生成描述。在 UCF-Crime 与 XD-Violence 的 HIVAU 视频摘要评测中，异常引导观察相较匹配的 CLIP 引导版本改善 GPT 评估的合理性、细节和一致性。检测方面，UCF-Crime AUC 从 84.36% 提至 86.25%；XD-Violence AP 从 68.07% 提至 72.58%，摘要另报最终 AUC 为 91.74%。

Insight: 语义反馈作用于上游分数曲线，并要求局部证据核验，避免仅凭整段描述泛化到所有时间片。摘要中的 GPT 评价是模型评审指标；XD-Violence 最终 AUC 91.74% 所对应的比较基线表述不够清楚，不宜与前述 AP 提升混为一谈。

Abstract: Training-free language-based video anomaly detection (VAD) uses frozen vision–language models (VLMs) and large language models (LLMs) to localize abnormal events without task-specific model-weight optimization. Existing approaches increasingly exploit broader temporal and semantic context, yet generic context does not necessarily foreground abnormal events or specify how video-level semantics should modify local anomaly scores. We introduce a training-free semantic feedback framework that uses VAD-guided active observation for video anomaly understanding (VAU) and returns the resulting anomaly-focused description to refine temporal detection. Starting from an upstream anomaly-score curve, we adapt reasoning-based iterative frame selection: high-score regions seed the observations, and anomaly scores together with VLM relevance assessments guide subsequent temporal exploration. A frozen video-language model generates a single open-ended anomaly description from the selected frames. The description is returned through two complementary feedback levels: video-level feedback determines whether the score distribution should be suppressed, preserved, or sharpened, with selective visual rechecking, while local feedback retrieves event-related windows and promotes scores only after target-local visual verification. All model weights remain frozen, and feedback is applied once without regenerating the description. On the HIVAU video-summary evaluations of UCF-Crime and XD-Violence, VAD-guided observation improves GPT-rated reasonability, detail, and consistency over the matched CLIP-guided variant. For VAD, semantic feedback improves the re-evaluated upstream detector from 84.36% to 86.25% AUC on UCF-Crime and from 68.07% to 72.58% AP on XD-Violence, where the final AUC is 91.74%.

[109] BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates cs.CV | cs.AIPDF

Lijian Li, Yuanpeng He, Yudian Zheng, Chung-ju Huang, Linyu Li

TL;DR: BAT3R 提出免训练的贝叶斯状态更新框架，通过约束状态漂移并按观测可靠性和空间相关性自适应融合历史与新帧，面向长序列在线 3D 重建。

Details

Motivation: 整段视频输入的前馈 Transformer 具有二次复杂度，难以扩展到长序列；流式重建虽可保持常量内存，但循环状态若不建模观测可信度与历史置信度，会随时间累积误差。

Result: BAT3R 包含状态创新门控 SIG 和贝叶斯引导自适应状态更新 BASU。SIG 将状态演化锚定到几何初始化以抑制表征漂移；BASU 在 Kalman 滤波框架下递归更新状态，结合信息论增益表示观测的时间可靠性，并通过交叉注意力刻画空间相关性，从而保护已稳定区域、积极吸收新场景内容。作者在 7-Scenes、TUM、ScanNet、KITTI 等标准基准上报告其持续优于所比较的在线 3D 重建方法，长序列和复杂场景增益最明显。摘要未给出具体误差或定量差值。

Insight: 方法把长期记忆置信度与新观测可靠性结合，针对在线递归更新的漂移和过度覆盖问题；“训练免除”可能便于迁移，但摘要未说明对不同前馈骨干的依赖、计算开销或各基准的定量结果。

Abstract: Feed-forward transformer models have driven remarkable progress in 3D vision,yet their quadratic complexity renders them impractical for long video sequences. Streaming 3D reconstruction addresses this bottleneck by processing frames sequen?tially with constant memory. However, existing recurrent architectures inevitably suffer from progressive degradation over long sequences, as they fail to account for the intrinsic reliability of observations and the accumulated confidence of the historical state, leading to severe error accumulation. To address these issues, we introduce BAT3R, a training-free Bayesian adaptation framework comprising two synergistic modules: State Innovation Gating (SIG) and Bayesian-guided Adaptive State Updating (BASU). SIG acts as a distributional regularizer that anchors state evolution to its geometric initialization, effectively arresting representational drift during long-term inference. Building on this, BASU reformulates state update as a recursive Bayesian inference process within a Kalman filtering framework. By jointly modeling the temporal reliability of observations via information-theoretic gain and spatial relevance via cross-attention maps, BASU achieves an asymmetric spatio-temporal state fusion: well-consolidated historical regions are protected from redundant updates, while genuinely novel scene content is aggressively incor?porated. Extensive experiments on standard benchmarks (7-Scenes, TUM, ScanNet, KITTI, etc.) demonstrate that BAT3R consistently outperforms state-of-the-art online 3D reconstruction methods, with the most significant gains observed on long sequences and complex scenes.

[110] RAR: Rate-Adaptive Rollout for Time-Consistent Interactive World Models cs.CVPDF

Yaoyang Liu, Dongyu Yan, Yihua Du, Shengju Qian, Zeke Xie

TL;DR: RAR 用共享物理时间坐标和多步长训练，使交互式视频世界模型能在推理时改变时间速率、切换生成速率并延长物理时间跨度。

Details

Motivation: 固定时间速率造成 temporal-rate coupling：运行时吞吐与模型学习速率不匹配时，世界演化会显得过快或过慢；单一速率还使固定上下文下的运动细节与长时跨度相互制约。

Result: RAR 的 Stride-RoPE 将不同速率采样的潜帧映射到共同物理时间坐标，以支持多速率一致生成、未见速率泛化及在线切换且保留历史。多步长视频采样则联合训练细粒度动力学与长时间演化。作者在动作控制和相机控制环境中报告了时间一致的多速率 rollout 与在线切换，并称在保持高速率生成质量的同时延长物理时间范围；也将方法迁移到预训练世界模型且维持感知质量。摘要没有给出数据集、定量分数或具体延长比例。

Insight: 把时间速率显式绑定到物理时间坐标，目标是解耦模型帧率与世界演化速度，并缓和长短时间尺度的折衷。当前摘要结论以定性描述为主，实际可控范围、切换开销和长期稳定性需结合正文结果判断。

Abstract: Video-based interactive world models aim to enable real-time, user-controllable environments. Yet most existing systems are trained and rolled out at a fixed temporal rate, limiting both runtime-rate adaptation and physical-horizon coverage – a limitation we term temporal-rate coupling. Specifically, when runtime generation throughput differs from the learned temporal rate, the generated world evolves unintentionally more slowly or quickly. Moreover, using a single temporal rate imposes an inherent trade-off between temporal granularity and physical horizon: higher rates capture finer motion but cover a shorter duration under a fixed context window. To address both limitations, we introduce Rate-Adaptive Rollout (RAR), a framework that enables interactive world models to roll out at controllable temporal rates. At its core, Stride-RoPE aligns latent frames sampled at different rates on a shared physical-time coordinate, enabling consistent multi-rate rollout, unseen-rate generalization, and online rate switching without discarding history. Meanwhile, RAR uses multi-stride video sampling to enable joint learning of fine-grained dynamics and long-horizon evolution. Across action- and camera-controlled environments, RAR achieves temporally consistent multi-rate rollout and online rate switching while preserving high-rate generation quality and extending physical horizons under matched training budgets. RAR also transfers to pretrained world models while preserving perceptual quality.

[111] From Sound to Sense: An Anatomy of Semantic Supervision for Speech-Driven Visual Grounding in Remote Sensing Imagery cs.CVPDF

Zhongyu Pang

TL;DR: SemEcho 将冻结 Whisper 语音通路接入视觉定位模型，分析语音到遥感图像定位中不同语义监督方式，并显示直接音频定位是否优于 ASR 级联系统取决于识别器质量。

Details

Motivation: 直接从音频预测目标框可绕开语音识别，但其是否真正理解语言语义、何时优于先转写再定位，尚未得到测量；而该任务的语音训练数据由文本表达合成，提供了可比较的文本上界。

Result: 作者构建共享解码器的 SemEcho，并考察特征模仿蒸馏、词级 CTC 监督和两阶段 ASR 预训练等语义注入途径，使用线性探针分析语义表征。还发布 Audio-DIOR-R（80,973 个训练片段，10 种音色、K=3）和控制声学/语义属性的 Audio-DIOR-Comp。摘要可见的结果显示：面对较弱的领域内 ASR（CTC 0.53），直接音频方法在 Acc@0.5 上高 9.0 分，转写错误集中于空间关系词；面对较强 ASR（CTC 0.029），级联方法达到文本上界（80.8 对 81.0），并比最强直接模型高 8.8 分。摘要提到有四项发现，但当前可见文本仅包含第一项。

Insight: 是否绕开 ASR 有效，关键取决于识别器是否能准确转写领域空间关系词；“端到端音频必然更好”的结论不成立。由于现有摘要内容在第一项发现后结束，其他语义监督比较结论无法据此补写；报告的数据集构造细节也表明结论与合成语音条件有关。

Abstract: Speech-driven visual grounding in remote sensing (RSVG) removes the typing bottleneck of text-based interaction and is intended to preserve prosodic intent, enabling hands-free target acquisition in time-critical missions. The direct-audio paradigm, mapping raw speech straight to a bounding box and bypassing automatic speech recognition (ASR), was recently adopted by RSTalker, but what such grounders actually learn about language remains unmeasured. SemEcho grafts a frozen Whisper speech pathway onto a VLTVG grounding host so that speech and text share one decoder, making the paired text, free by construction since speech corpora in this task are synthesized from DIOR-RSVG expressions, a training-time teacher and a fair, same-host text upper bound. On this instrument we dissect three semantic injection channels (feature-mimicry distillation, word-level CTC supervision, and two-stage ASR pretraining) with linear semantic probes, and release Audio-DIOR-R (80,973 training clips; 10 timbres x K=3; Qwen3-TTS synthesis with online sox noise lambda in [0, 0.4]) and Audio-DIOR-Comp, a re-split holding out 17 attribute-category combinations under a 2x2 acoustic-semantic protocol. Our anatomy yields four findings. (i) Whether direct alignment beats a domain-trained ASR cascade is decided by recognizer quality: against a weak in-domain cascade (CTC 0.53) direct wins by 9.0 Acc@0.5 points with transcription errors concentrated on spatial-relation words, but a strong in-domain cascade (CTC 0.029) reaches the text upper bound (80.8 vs. 81.0) and beats even the strongest direct model by 8.8 points; the anti-ASR premise holds only against weak recognizers.

[112] VideoSlate: Stateful Commit Control for Temporally Grounded Agentic Video Reasoning cs.CVPDF

Shuo Zheng, Yuanjie Dang, Peng Chen, Jianjun Xiang, Ronghua Liang

TL;DR: VideoSlate 用共享证据黑板与显式提交控制管理视频智能体的候选片段修订，只有候选证据支持修订时才替换当前片段。

Details

Motivation: 多阶段视频推理会不断获得新候选和观察，重新定位既可能纠正错误，也可能丢弃当前片段中已有的有效证据；系统需要将证据积累和状态更新分开管理。

Result: VideoSlate 为候选片段、来源、构造过程和验证观察建立类型化证据黑板，由相对于当前片段的认证器综合候选、现有片段和相关提案来判断修订支持度。控制器据认证结果批准更新或保留当前片段，并把决定及依据记录到黑板。作者在视频定位和基于定位的视频问答基准上进行评估，报告选择性修订可改善时间定位；摘要未提供数据集名称、具体指标或提升幅度。

Insight: 显式提交门控可以减少新证据到来时的无依据替换，并让每次状态变更保留可追溯依据。现有摘要只定性报告定位改善，尚不足以判断证据黑板和认证器的单独贡献，以及额外记忆或推理开销。

Abstract: Agentic video reasoning systems gather and assess video evidence across stages, selecting relevant spans as visual evidence for answering questions. New candidates and observations prompt reassessment of the current span, but revisions can correct localization errors or discard useful evidence already present in the adopted span. Systems therefore need to connect evidence organization across stages with temporal state control, allowing accumulated information to jointly inform revision assessment while explicit adoption decisions govern how assessment outcomes affect downstream visual evidence. We introduce VideoSlate, a stateful commit-control framework built around shared evidence. A typed evidence blackboard links candidate spans, provenance, construction, and verifier observations, enabling an incumbent-relative certifier to assess revision support from the candidate, current span, and related proposals. Based on certification, the controller authorizes a revision or retains the current span, separating evidence accumulation from state updates and recording each decision and its evidence on the blackboard. Extensive evaluations across video grounding and grounded question-answering benchmarks show that VideoSlate can improve temporal localization through selective revision.

[113] SyncReward: Human-Aligned Synchronization Reward for Audio-Visual Generation cs.CV | cs.SDPDF

Kaisi Guan, Ying Ba, Wenjing Wang, Xihua Wang, Chen Li

TL;DR: SyncReward 构建带专业人员同步评分的数据与基准，并训练与人类判断对齐的音视频同步奖励模型，用于评测和生成时筛选或强化优化。

Details

Motivation: 联合音视频生成仍常出现语音、声音与画面事件时间错位；研究受限于缺少人工标注数据、可靠基准及能评估生成同步质量的模型。

Result: 作者发布 SyncReward-Data-80K，包含对 8 种音视频生成器片段的 8 万条专业同步评分，按语音、起始时刻、乐器和环境声四类构建；另建 1,000 个片段的留出基准，每段由多位专业人员独立评分。对比表明，专门同步指标和全模态模型与人类判断存在明显偏差。SyncReward 以真实片段的语义—时间对比预训练，并用生成片段的人类监督微调；作者称其在人类评分一致性上优于所评方法，并可通过推理时 best-of-N 选择和强化学习后训练改善同步。摘要未报告一致性系数或提升幅度。

Insight: 该工作同时提供监督数据、独立基准和面向人类判断的奖励模型，能用于质量评测和生成优化两个环节。摘要未给出生成同步改进的量化幅度或不同声音类别的误差分布；人类标注覆盖的生成器也只有 8 种。

Abstract: Recent years have witnessed rapid progress in joint audio–visual generation. However, audio–visual synchronization remains a challenge, as generated speech and sound often temporally misalign with visual events. A key bottleneck is the lack of human-annotated data, reliable evaluation benchmarks, and synchronization-aware evaluation models for generated audio–visual outputs. To address these limitations, we conduct a comprehensive study spanning data, benchmarking, and reward modeling. We first introduce SyncReward-Data-80K, a dataset containing 80K professional synchronization ratings for clips from eight audio–visual generators. A taxonomy-guided construction pipeline covers four sound-event families: Speech, Onset, Instrument, and Ambient. We further construct \textbf{SyncReward-Bench}, a held-out benchmark of 1K clips, each independently rated by multiple professionals. Using this benchmark, we systematically evaluate specialized synchronization metrics and omni-modal models, revealing their substantial misalignment with human judgments. Finally, we develop SyncReward, a reward model that combines semantic–temporal contrastive pretraining on real clips with human supervision on generated clips. As an evaluation metric, SyncReward achieves the strongest agreement with human ratings among the evaluated methods. As a reward, it improves synchronization through inference-time best-of-$N$ selection and reinforcement learning based post-training. We will release the data, benchmark, model, and code at https://github.com/WeChatCV/SyncReward.

[114] Throwing4: Phase-Aware Pose Modeling for Throw-Distance Prediction in Elite Athletics cs.CVPDF

Bikash Kumar Badatya, Kartike Tiwari, Jyotirmoy Amin, Ravi Hegde

TL;DR: Throwing4 提供四类精英投掷项目的广播视频姿态、动作阶段和官方距离标注；实验显示显式阶段对齐有助于姿态序列的投掷距离回归。

Details

Motivation: 将整段投掷动作视作无结构序列可能忽略助跑、发力等技术阶段差异。研究考察按生物力学阶段切分并对齐姿态变化，能否在真实广播视频条件下提升成绩预测稳定性。

Result: 数据集包含 500 段标枪、铁饼、链球和铅球比赛视频，配有二维姿态序列、逐帧阶段标注及官方距离。作者将距离预测建模为回归，对比完整序列建模与将动作划分为四阶段、融合阶段运动表征的方法；在多种图结构姿态编码器上，阶段条件方法均降低预测误差。摘要未提供具体误差数值、数据拆分细节或编码器名称。

Insight: 动作阶段可作为姿态运动建模的有效结构先验，并且该数据同时支持细粒度技术分析。500 段的规模和广播视频来源限制了可得结论；具体提升幅度及跨项目泛化能力需查阅正文实验。

Abstract: Elite athletics throwing provides a structured biomechanical action sequence together with an objective continuous outcome, namely throw distance, making it a suitable setting for performance-oriented sports analytics. In this work, we investigate whether explicit phase alignment of pose dynamics improves throw-distance prediction under realistic broadcast conditions. To support this study, we curate Throwing4, a dataset of 500 elite throwing clips spanning javelin, discus, hammer, and shot put, with corresponding two-dimensional pose sequences, frame-level phase annotations, and official distance labels. We formulate throw-distance prediction from broadcast-derived pose as a regression problem and compare full-sequence modeling against a phase-conditioned approach that partitions each attempt into four technique stages and fuses phase-specific motion representations. Experiments across multiple graph-based pose encoders show a consistent reduction in prediction error when phase structure is incorporated, indicating that phase-aligned modeling provides a more stable representation than treating the entire motion as an unstructured sequence. These findings suggest that explicit biomechanical phase information is a useful inductive prior for pose-based performance modeling in elite athletics throwing. Beyond score prediction, the dataset and protocol provide a basis for future work on fine-grained technique analysis, interpretable sports analytics, and robust modeling from broadcast video.

[115] Tikhonov-Stabilized Bezier Representation Forecasting for Training-free Diffusion Acceleration cs.CVPDF

Lei Zhu, Mujie Lin, Ruochong Zheng, Guangyi Wang, Hao Li

TL;DR: BeziCast 用带 Tikhonov 稳定拟合的低阶 Bézier 轨迹预测扩散模型靠近输出端的去噪表示，提供无需重训的特征缓存加速。

Details

Motivation: 扩散模型推理需要反复调用去噪器；缓存并外推中间特征虽便于部署，但基于有限差分的高阶 Taylor 外推在较长缓存间隔下易不稳定，局部模块预测误差也可能沿去噪网络累积。

Result: BeziCast 以 Tikhonov 稳定化拟合 Bézier 控制点，不显式估计高阶导数，并使轨迹容量与缓存查询密度解耦；另将预测权重改写为凸包启发的风险保护，对高风险预测向保守的单纯形投影估计软回拉。在多种图像和视频扩散模型上，作者称质量优于加速基线；FLUX.1 最高加速 4.79 倍，HunyuanVideo 最高 4.11 倍。摘要未提供对应速度档位下的具体质量指标。

Insight: 稳定的低阶轨迹拟合和外推保护旨在同时减少缓存预测抖动与误差累积。最高加速比不能单独说明各加速档位的质量折损；需要结合正文中的采样设置和指标曲线比较实际质量—速度权衡。

Abstract: Diffusion models, particularly Diffusion Transformers, achieve strong image and video generation quality but remain expensive at inference time due to repeated denoiser evaluations. Feature caching offers a deployment-friendly acceleration strategy by reusing or predicting intermediate representations without retraining the generator or modifying the sampler. However, existing cache-then-forecast methods often rely on Taylor-style finite-difference extrapolation, which can become unstable over longer cache intervals, and typically predict local module outputs whose errors may accumulate through subsequent denoiser blocks. We propose \textbf{BeziCast}, a training-free diffusion acceleration framework that forecasts output-proximal denoising representations using low-order B’ezier trajectories. BeziCast estimates B’ezier control points via Tikhonov-stabilized fitting, providing a smooth, capacity-controlled temporal parameterization that avoids explicit high-order derivative estimation and decouples trajectory capacity from cache-query density. To moderate aggressive extrapolation, we further derive equivalent forecasting weights and introduce a convex-hull-inspired guardrail that detects high-risk predictions and softly pulls them toward a conservative simplex-projected estimate. Extensive evaluations across advanced image and video diffusion models demonstrate the effectiveness of BeziCast. In particular, BeziCast accelerates FLUX.1 by up to (4.79\times) and HunyuanVideo by up to (4.11\times), while preserving substantially better generation quality than competing acceleration baselines.

[116] Pixels or Code? Benchmarking Instruction-Guided Visual Artifact Operation cs.CVPDF

Shun Guo, Shuizhou Shi, Boao Wang, Wei Xiong, Siyu Sun

TL;DR: VAO-Bench 系统评测图表、技术图、网页和文档等视觉制品的指令编辑，覆盖直接修改、依赖更新与内容保留，并比较像素编辑、代码生成和工具智能体。

Details

Motivation: 编辑图表、网页或工程图不仅要改指定区域，还需同步更新依赖内容、保留不相关部分并满足领域约束；以往基准只覆盖部分需求，例如图表级联更新。

Result: VAO-Bench 包含 9 类视觉制品、3 个难度级别、共 1,573 个经人工筛选的任务。可执行制品与确定性渲染用于追踪参考变换，并标注直接编辑、依赖更新和保留约束。所有系统在仅图像输入协议下接收相同源图和指令，评估直接编辑准确性、依赖感知更新、内容保留、领域有效性和视觉质量。实验显示，代码可执行系统更擅长依赖更新；像素编辑执行更快且视觉质量较强。明确写出依赖关系可同时改善两类方法，提示隐式依赖推断是主要瓶颈。摘要未给出各系统的具体分数。

Insight: 该基准把视觉编辑从“改得像不像”扩展到依赖一致性与内容保留，并将多类执行范式放在同一图像输入协议下比较。不同方法体现质量、速度与结构更新能力的取舍；摘要中的发现为定性结论，未提供量化幅度。

Abstract: Visual artifacts such as charts, technical diagrams, webpages, and document pages underpin scientific, commercial, and engineering workflows. Editing these artifacts requires direct modification, dependency-aware updates, preservation of unaffected content, and compliance with domain constraints. We define this capability as Visual Artifact Operation (VAO). Prior benchmarks test parts of this problem, including cascading updates in charts. We study direct edits, dependent updates, and preservation across visual domains and pixel, code, and agent paradigms under one image-only protocol. We introduce Visual Artifact Operation Bench (VAO-Bench), comprising 1,573 manually screened tasks across nine artifact categories and three difficulty levels. VAO-Bench uses executable artifacts and deterministic rendering to trace reference transformations and annotate direct edits, dependent updates, and preservation constraints. All systems receive the same source image and instruction; we evaluate pixel-space editors, code-generating multimodal systems, and tool-using agents on direct-edit accuracy, dependency-aware update, content preservation, domain validity, and visual quality. Experiments show that executable-code systems better support dependent updates, while pixel-space editors offer faster execution and strong visual quality. Explicitly stating dependencies improves both paradigms, identifying implicit-update inference as a central bottleneck. VAO-Bench provides a reproducible foundation for dependable visual artifact editing across generation paradigms.

[117] VisExMEM: Visual Evidence-Based Explainable Multimodal Evaluation Metric cs.CV | cs.AI | cs.CLPDF

Subhadeep Roy, Christoph Leiter, Steffen Eger

TL;DR: VisExMEM 将图像描述拆成原子主张、定位视觉证据并逐项核验，再汇总为全局评分和可追溯解释；作者另提出 AGITA-Bench 评测全局及主张级能力。

Details

Motivation: 图文评价常用单一整体分数，难以指出描述中哪些语义细节得到图像支持；即使指标提供解释，解释依赖的视觉证据仍不易核查。

Result: VisExMEM 是无需参考文本的评估器，按原子主张定位、核验并平均支持度。基于 Qwen3.5 的 4B 和 9B 版本在 10 万样本上以多任务监督微调、反事实精炼和证据感知 GDPO 训练。8 个基准上，VisExMEM-9B 在所比较开源评估器中描述综合分最高，比 VQAScore 高 11.7 分，并与 GPT-5.6 Sol 具有竞争力。AGITA-Bench 包含 600 对人工标注图文和 8,700 个原子主张。对未调优的 Qwen3.5、GPT-5.6 Sol，分解式提示相较整体式提示提高同一综合分 3.6–5.1 分；人工评估中，VisExMEM 解释优于既有可解释指标并接近 GPT-5.6 Sol。

Insight: 把评分绑定到可定位、可逐项核对的证据，有望让图文指标更可解释，也能在 AGITA-Bench 上单独检查主张级质量。摘要未说明综合分的组成和跨模型校准方式；与 GPT 的“接近”也未给出具体差距，结果应限于所列评测集和对比系统。

Abstract: Current image-text evaluation often relies on a global score, which gives limited insight into which semantic details are supported by the image. Explainable metrics provide rationales, but the visual evidence supporting those rationales can still be difficult to verify. We introduce VisExMEM, a reference-free evaluator that decomposes captions into atomic claims, localizes visual evidence, verifies each claim, and averages claim-level support into a global score. Explanations use the same evidence and decisions. We train Qwen3.5-based 4B and 9B models on 100K samples with multi-task SFT, counterfactual refinement, and evidence-aware GDPO. Across eight benchmarks, VisExMEM-9B achieves the highest descriptive composite among open-source evaluators in our comparison, exceeding VQAScore by 11.7 points while remaining competitive with GPT-5.6 Sol. We also introduce AGITA-Bench, containing 600 human-annotated image-text pairs and 8.7K atomic claims for global and claim-level evaluation. For untuned Qwen3.5 and GPT-5.6 Sol controls, decomposed prompting raises the same composite by 3.6-5.1 points relative to holistic prompting. Human evaluation shows that VisExMEM explanations outperform prior explainable metrics and approach GPT-5.6 Sol.

[118] GHOST: Geometry-Constrained Hybrid Diffusion for Open-Vocabulary 3D Scene Graph Generation cs.CV | cs.MMPDF

Changqun Feng, Wangxiandi Yin, Xin Hu, Ruiting Dai, Feiyang Yang

TL;DR: GHOST 通过把谓词文本映射为可微几何约束，并联合扩散生成场景图节点几何与关系边谓词，提升开放词汇 3D 场景图的空间关系一致性。

Details

Motivation: 开放词汇场景图方法常以语言相似度预测空间关系，但语义共现不能替代明确几何判断，因而易出现与物体布局不符的空间谓词。

Result: GHOST 采用几何到约束再到谓词的去噪流程，联合建模连续节点几何和离散关系谓词。语言条件约束归纳将谓词嵌入转为可微几何约束，以指导布局细化和谓词重加权；几何—拓扑协同去噪器通过噪声自适应交互和方向感知建模共同更新节点与边。在 3DSSG 上，作者报告其在封闭集和开放词汇设置均达到当时最佳结果，空间谓词指标较 RelationField 高 17 分以上。摘要未说明该分值对应的具体指标。

Insight: 把自然语言谓词转化为布局约束，让空间关系预测接受几何证据约束，而不只依赖语义共现。报告的 17+ 分提升针对 3DSSG 空间谓词对比；具体度量、开放词汇词表与跨场景泛化范围需看正文。

Abstract: 3D scene graph generation requires understanding both object semantics and metric spatial relations. However, existing open-vocabulary methods often struggle with spatial predicates because language similarity reflects semantic co-occurrence rather than explicit geometry. We propose GHOST, a Geometry-constrained Hybrid diffusion framework for Open-vocabulary 3D Scene graph generaTion. GHOST jointly models continuous node geometry and discrete edge predicates, following a geometry-to-constraint-to-predicate denoising process to ensure geometrically consistent relation prediction. It introduces language-grounded constraint induction, which maps predicate embeddings to differentiable geometric constraints that guide both layout refinement and predicate reweighting. A Geometry-Topology Collaborative Denoiser further couples node and edge reasoning through noise-adaptive interaction and orientation-aware modeling. Experiments on 3DSSG demonstrate state-of-the-art performance in both closed-set and open-vocabulary settings, including over 17-point gains over RelationField on spatial predicates.

[119] Not Every Cue Is Evidence: Cue Supportability for Forensic Distillation in Face Forgery Detection cs.CVPDF

Binjia Zhou, Dawei Luo, Shuai Chen, Feng Xu, Haoyuan Li

TL;DR: 该方法在面部伪造检测的蒸馏前审查教师提供的取证线索，要求线索既能由剩余特权上下文预测，也能与当前图像中的高归因视觉区域建立因果联系。

Details

Motivation: 伪造生成过程中的细粒度线索可能超出输入图像分辨率和质量的可见范围。将这些不可观测线索作为特权监督教给学生，可能使其输出模板化或泛化解释，即出现判断正确但证据错误（CAWE）。

Result: 作者提出双重可支持性审计：候选线索需同时通过基于剩余特权上下文的可预测性检验，并与当前图像高归因区域相关联；仅对通过两项检验的线索赋予较大蒸馏权重。在 8 个基准上，作者报告该方法维持或提高检测准确率，同时显著提升解释忠实度。摘要未给出基准名称、数值或统计显著性。

Insight: 将“教师知道某条线索”与“学生能从图像观察到该证据”分开，可缓解解释监督超出输入可见信息的问题。摘要结果是定性报告，审计本身能否稳定识别不可见线索，以及归因区域的选择对结论的影响，需要正文实验确认。

Abstract: Explainable face forgery detection requires not only correct judgments but also forensic evidence supported by the input image. However, training cues are typically derived from the forgery generation pipeline. Since images available to the model are of limited quality and resolution, it is difficult to observe the fine-grained traces these cues refer to. Training methods that further supply such cues as privileged supervision exacerbate the problem, as the student is taught to articulate evidence it cannot perceive. The student consequently resorts to templated or over-generalized expressions, giving rise to the correct-answer, wrong-evidence (CAWE) failure. We propose a dual supportability audit that filters privileged cues before distillation. Each candidate cue must be predictable from the remaining privileged context and causally tied to high-attribution visual regions in the current image. Only cues satisfying both conditions receive substantial distillation weight, preventing unsupported details from leaking into the student. Experiments across eight benchmarks show that our method maintains or improves detection accuracy while significantly raising explanation faithfulness compared to existing paradigms.

[120] Teaching Vision-Language Models to Decline: Support Completion Against Mirage cs.CVPDF

Alfaj Uddin Ahmed, Md. Shaown Miah, Md Arid Hasan, Syed Ishtiaque Ahmed, Sayeed Shafayet Chowdhury

TL;DR: TIDE 通过让视觉语言模型同时面对空白图像、同领域错误图像、域外图像和有效图像，并为无效证据训练拒答，抑制模型在缺少图像支持时产生“幻视式”答案。

Details

Motivation: 模型可能在没看到相关图像或图像与问题不匹配时仍给出碰巧正确的答案，这种答案无法由旁观者核验。作者指出常规交叉熵并不奖励拒答，且只凭正确答案无法区分视觉证据推理与忽略图像。

Result: 作者形式化 support completion，理论上证明 TIDE 下图像盲预测器每样本至少次优 1.04 nats；最优策略只在反事实无效图像条件下拒答，而保留有效图像行为。有限模型的答/拒决策错误下界与表征区分有效和无效证据的能力相关。Qwen2.5-VL-7B 的冻结留出评测中，报告的是 TIDE 的三条件前身：无效证据 mirage 率从 0.759 降至 0.097，空白图像率从 0.700 降至 0.003，误拒成本为 0.126。无标签 grounding 使 mirage 率升至 0.859，似然优势对比仅降至 0.718；在 support completion 之后再做 grounding 又会重新引入 mirage。四条件完整协议在三种基础模型上的实验仍在进行，摘要未报告。

Insight: 拒答需要由无效证据条件明确监督，单做 grounding 可能不能阻止模型在无依据时作答，甚至会重新打开 mirage。当前数字来自三条件前身而非摘要所称的四条件完整 TIDE；误拒与拒答率也需和具体判定协议一起解读。

Abstract: Vision-language models answer questions about images they were never shown, a failure called mirage. Such an answer can be correct without being justified by the evidence, and the difference matters wherever a model’s output is checked by someone who did not see its input. We locate the failure in the training objective: cross-entropy separates a model that reads the image from one that ignores it by only the conditional mutual information between answer and image, and no training example rewards declining. We formalise support completion, TIDE, which also shows each training question with a blank image, a same-domain wrong image and an out-of-domain image, with targets that decline, and prove that under TIDE every image-blind predictor is suboptimal by at least 1.04 nats per example, that the optimum declines exactly on the counterfactual kinds while keeping its behaviour on valid images, and that for any finite model the error of its answer-or-decline decision, declining on valid input plus answering on invalid input, is bounded below by how well its representation separates valid from invalid evidence. On Qwen2.5-VL-7B under a frozen held-out protocol with two judges, the three-condition predecessor of TIDE reduced the invalid-evidence mirage rate from 0.759 to 0.097 and the blank-image rate from 0.700 to 0.003 at a false-abstention cost of 0.126; label-free grounding raised the rate to 0.859, a likelihood-advantage contrast moved it only to 0.718 because its empty-image term carries no gradient, and grounding applied after support completion re-opened the mirage. The four-condition protocol on three base models, with seeds and component ablations, is in progress and not reported here.

[121] Auditing Observation-Quality Shortcuts in the Evaluation of Person-Centered Motion for Video Anomaly Detection cs.CVPDF

Chaehyun Kim, Changwon Lim

TL;DR: 该研究审计视频异常检测是否依赖人物检测、姿态估计和跟踪质量等观测捷径，并在边际质量变量匹配后检验运动特征是否仍有额外区分能力。

Details

Motivation: 异常偏离正常行为不一定代表安全关键事件；基于人体运动的异常识别可能反映的是检测或跟踪可靠性差异，而非任务相关运动本身。

Result: 作者在 UCF-Crime 的以人为中心类别上提出结合质量诊断、质量平衡、运动正交化和鲁棒性分析的评测框架。原始队列中，仅用观测质量特征就可达到或超过多线索运动特征的区分能力，显示存在强质量捷径。基数匹配后，11 个测量质量变量均低于预设边际平衡阈值，多线索运动在 Logistic Regression、Random Forest、RBF-SVM 上均优于质量特征，ROC-AUC 增益为 0.046–0.141。但匹配后非线性质量模型仍高于随机水平，因此作者将结果解释为“超出已测边际质量变量的额外区分信息”，不称其为无质量影响或因果运动理解。

Insight: 高 VAD 分数不能直接作为模型学到异常运动的证据；需要审计人物观测管线质量并控制相关差异。即便如此，边际匹配未排除质量变量的非线性组合或未测混杂，所以作者谨慎地不作质量无关或因果解释。

Abstract: Video anomaly detection (VAD) identifies deviations from normal patterns, but such deviations do not necessarily correspond to safety-critical events. We focus on human-centered emergencies and violence and investigate whether motion-based anomaly discrimination reflects task-relevant motion information or observation-quality differences arising from person detection, pose estimation, and tracking. We introduce an observation-quality-aware evaluation framework that combines quality diagnostics, quality balancing, motion orthogonalization, and robustness analyses. On human-centered categories of UCF-Crime, quality-only features match or exceed the discrimination of multi-cue motion in the original cohort, revealing a strong observation-quality shortcut. After cardinality matching brings all 11 measured quality variables below a prespecified marginal balance threshold, multi-cue motion consistently outperforms quality-only features across Logistic Regression, Random Forest, and RBF-SVM, yielding ROC-AUC gains of 0.046-0.141. Because nonlinear quality-only prediction remains above chance after matching, we interpret this result as additional discrimination beyond measured marginal quality rather than as evidence of quality-free or causal motion understanding. These findings show that high VAD performance should not be interpreted as evidence of abnormal-motion modeling without auditing the reliability of the person-observation pipeline. Our study provides a systematic protocol for evaluating observation-quality shortcuts and for assessing whether motion-based discrimination survives explicit control of detection, pose-estimation, and tracking reliability.

[122] KARAMEL: Knowledge-Aware Ranking for Multimodal Entity Linking in the Arts via Heterogeneous Graphs cs.CV | cs.LGPDF

Raffaele Scaringi, Alejandro Sierra-Múnera, Gennaro Vessio, Giovanna Castellano, Ralf Krestel

TL;DR: KARAMEL 将视觉语言编码与异构知识图神经网络结合，用社会历史和关系上下文解决艺术文化遗产中的多模态实体链接，并发布 WikiMuSA 数据集。

Details

Motivation: 多模态发现中的实体线索可能隐含在图像、文本和背景关系中；只做视觉—语言表面匹配，难以区分含糊或描述不充分的艺术实体。

Result: KARAMEL 用视觉语言编码器表示图文线索，并与异构图神经网络融合，使模型能利用实体间的结构化上下文。作者构建 WikiMuSA，将艺术品与 Wikidata 实体通过图像、文本描述和结构化背景知识关联。实验报告 KARAMEL 超过强基线，尤其在隐式或抽象提及上；消融强调领域结构化上下文对表征学习的重要性。摘要未提供数据集规模、具体指标或提升幅度，并称数据集与代码公开。

Insight: 领域背景知识可以补足视觉—语言相似度对隐含指称的不足；该方法在艺术文化遗产中实例化，但是否迁移到其他科学发现领域取决于相应知识图谱的覆盖与质量。摘要只有定性比较，无法量化不同模块的实际增益。

Abstract: Discovering and disambiguating entities from multimodal and heterogeneous data is a key challenge in discovery science, where relevant knowledge is often implicit and distributed across modalities. We present KARAMEL, a framework for multimodal entity linking (MEL) - the task of aligning visual and textual mentions to entities in a knowledge base - in the cultural heritage domain. Unlike prior vision-language models that primarily rely on surface-level multimodal alignment, our approach incorporates sociohistorical and contextual knowledge through a knowledge-aware architecture that fuses a vision-language encoder with a heterogeneous graph neural network. This design enables contextual reasoning over complex relational structures and supports robust entity disambiguation under ambiguous or underspecified evidence. To benchmark multimodal discovery in this setting, we introduce WikiMuSA (Wikidata-based Multimodal Semantic data for Art), a new dataset that links artworks to Wikidata entities via images, textual descriptions, and structured contextual knowledge. Experimental results show that KARAMEL surpasses strong baselines, particularly on implicit or abstract mentions, while ablation studies highlight the critical role of domain-specific, structured context in learning expressive representations. While instantiated in the cultural heritage domain, KARAMEL addresses a general knowledge discovery problem and applies to other scientific settings involving multimodal observations and heterogeneous background knowledge. Both the dataset and source code are made publicly available.

[123] MiMaAlign: Hierarchical Micro-Macro Alignment for Physically Plausible Video Generation cs.CVPDF

Yansong Liu, Litong Gong, Weijie Li, Tiezheng Ge, Ming Zeng

TL;DR: MiMaAlign 将连续运动动力学与事件常识分别对齐到视频扩散模型的浅层和深层特征，以改善生成视频的物理合理性。

Details

Motivation: 文本生成视频虽能产生高保真画面，仍常出现运动不连续和事件交互不合理。以往表征对齐多迁移通用视频特征或短程运动线索，将物理合理性视为单层表示问题。

Result: MiMaAlign 使用训练阶段辅助目标进行微—宏层级对齐：微观层用 Neural ODE 学到的连续时间运动表示约束扩散特征；宏观层借助指令提示视觉语言模型提供物体交互与事件语义，再将两种教师监督分别对齐到生成器浅层和深层。CogVideoX-2B 在 VideoPhy 上的 Physical Commonsense（PC）分数相对基线最高提升 21.0%；在更难的 VideoPhy2 上也有明显提升，同时维持 VBench 一般视频质量，并由盲测用户比较支持。

Insight: 不同抽象层级分别学习连续动力学与事件常识，试图避免单一对齐信号覆盖不了物理合理性的多尺度因素。摘要未提供 VideoPhy2 的数值、用户盲测规模或训练成本变化；最高 21% 是相对提升，不能与绝对分数提升混淆。

Abstract: Text-to-video (T2V) diffusion models have made remarkable progress in generating high-fidelity visuals, yet their outputs still often exhibit physically implausible phenomena, including discontinuous dynamics and implausible event interactions. Existing representation alignment methods mainly transfer generic video features or short-range motion cues, treating physical plausibility as a single-level representation problem. To address this issue, we propose MiMaAlign, a hierarchical micro-macro alignment framework that transfers complementary dynamics-oriented and event-oriented knowledge into different abstraction depths of a pretrained video diffusion model. Specifically, micro-level continuous dynamics alignment imposes constraints by aligning diffusion features with a continuous-time motion representation learned by a Neural Ordinary Differential Equation (Neural ODE), while macro-level event-commonsense alignment provides representations of object interactions and event semantics through an instruction-prompted vision-language model (VLM). With training-only auxiliary objectives, we align the complementary supervision from the two teachers with shallower and deeper generator features, respectively, enabling the generator to internalize the corresponding physical priors. Empirical evaluations show that, on CogVideoX-2B, MiMaAlign achieves a relative improvement of up to 21.0% in the Physical Commonsense (PC) score on the VideoPhy benchmark over the baseline, and also delivers substantial gains on the highly challenging VideoPhy2 benchmark, while maintaining general video quality on VBench and being supported by blind user comparisons.

[124] Deep Learning-based Mechanical Characterization of Synthetically Generated Plant Fibers under Transverse Compression cs.CVPDF

Nikolaos Kegkeroglou, Bilal Sabillah, Myrto Inglezou, Vincent Placet, Céndric Clevy

TL;DR: 该研究用递归 U-Net 预测植物纤维横向压缩中的逐步变形与应力场，再由 CNN 回归横向杨氏模量，以替代耗时的有限元计算。

Details

Motivation: 植物纤维微结构差异会造成横向力学行为显著波动；有限元仿真每根纤维需数分钟，难以对真实纤维群体进行高通量表征。

Result: 模型针对来自真实大麻纤维截面的合成几何（含或不含髓腔）递归预测完整应力演化，而非固定构型上的单次场预测。训练数据为 1,000 种纤维几何及有限元真值。应力序列预测 SSIM 为 0.9574±0.0074、RMSE 为 0.0345±0.0079；横向杨氏模量估计 R² 为 0.938±0.024、MAE 为 24.31±1.23 MPa、MAPE 为 5.25±0.91%。每根纤维推理少于 1 秒，对比有限元仿真约 5 分钟。

Insight: 递归预测可将力学场变化和最终参数估计整合为快速替代流程，速度提升显著；误差是在合成纤维几何及有限元生成的标签上测得，摘要未报告真实实验测量验证，因此不能直接等同于真实材料测量精度。

Abstract: Plant fibers are increasingly used as sustainable reinforcements in composites, but the intrinsic variability of their microstructure causes substantial scatter in transverse mechanical behavior, complicating composite design. Characterizing this behavior relies on finite element (FE) simulations requiring several minutes of high-performance computing per fiber, making high-throughput characterization across realistic fiber populations impractical. In this study, we propose a deep learning framework replacing this FE bottleneck with a fast pipeline. It includes two coupled recursive U-Nets predicting the step-by-step evolution of fiber deformation and internal stress fields under transverse compression, followed by a regression CNN estimating the transverse Young’s modulus from the final stress field. Unlike prior approaches predicting stress fields for fixed configurations or idealized geometries, our framework recursively resolves the full temporal stress evolution for synthetic fibers derived from real hemp cross-sections, with or without a lumen. Trained on 1,000 synthetic fiber geometries with FE ground truth, it reproduces the stress field sequence with a Structural Similarity Index Measure (SSIM) of $0.9574 \pm 0.0074$ and a Root Mean Squared Error (RMSE) of $0.0345 \pm 0.0079$. It estimates transverse Young’s modulus with an $R^2$ of $0.938 \pm 0.024$, a Mean Absolute Error (MAE) of $24.31 \pm 1.23$ MPa, and a Mean Absolute Percentage Error (MAPE) of $5.25 \pm 0.91%$. Inference requires under one second per fiber versus five minutes per FE simulation, drastically reducing computational cost. This offers a fast, reliable tool for mechanical characterization of synthetic plant fibers, with promising applications in designing natural fiber-reinforced composites.

[125] NoduleFlow: Automatic 3D Lung Nodule Detection and Segmentation cs.CVPDF

Giulia Romoli, Filippo Ruffini, Francesco Di Feola, Valerio Guarrasi, Matteo Tortora

TL;DR: NoduleFlow 从连续 CT 体积序列自动进行肺结节定位、分类和分割，并研究借助随访扫描间配准传播候选区域是否优于逐次独立分析。

Details

Motivation: 低剂量 CT 肺癌筛查中的结节人工识别耗时且存在观察者差异；同一患者随访影像中的细微变化也难以稳定量化。既有方法可能只做粗定位、依赖先验病灶标注或人工初始化，或将检测与分割割裂。

Result: NoduleFlow 由四个相互连接的模块构成，推理时无需先验病灶标注、手工初始化或预选 ROI。研究还比较通过跨扫描配准场传播候选区域与每次扫描独立分析，并在四个独立数据集上系统评估泛化。摘要称结果显示自动化肺结节分析流程具有筛查与诊断应用潜力，但未报告数据集名称、指标或性能数字。

Insight: 方法将检测、分类、分割和随访信息传播纳入自动流程，针对既有方法依赖人工定位与忽略时间连续性的问题。摘要没有给出定量指标或跨中心数据细节，因此不能据此评估临床性能、假阳性风险或部署成熟度。

Abstract: Low dose Computed Tomography (CT) is the standard imaging modality for lung cancer screening. On CT scans, lung cancer typically manifests as pulmonary nodules, whose manual identification is time-consuming and prone to observer variability, image quality limitations, and suboptimal reading conditions. Moreover, subtle temporal changes are often difficult to quantify reliably in CT scans from the same patient across follow-up examinations. Deep learning-based approaches for the automatic detection and segmentation of lung nodules have demonstrated improved accuracy and efficiency. However, most existing approaches provide only coarse region-of-interest localisation, rely on prior lesion annotations or human supervision, or treat detection and segmentation as separate tasks. In this work, we present NoduleFlow, a framework for lung nodule detection, classification, and segmentation directly from series of CT volumes. Our proposal consists of four interconnected modules and operates in an automated manner at inference time, without requiring prior lesion annotations, manual initialisation, or region-of-interest pre-selection. We further investigate whether nodule localization benefits from the propagation of candidate regions across follow-up scans through inter-scan registration fields, as opposed to conventional independent per-scan analysis. The generalisability of the framework is systematically evaluated across four independent datasets. The results underscore the potential of automated pipelines for lung nodule analysis in screening and diagnostic settings.

[126] Heuristics are Not Enough: Calibrated and Corrected Caching for Video Diffusion Acceleration cs.CV | cs.AIPDF

Jiyue Jiang, Zhiyu Huo, Yuhan Chen, Yanruisheng Shao, Gong Zhang

TL;DR: AlignCache 用分阶段预算路由、输出端反事实误差校准和带动量的残差预测，构成闭环视频扩散缓存加速框架。

Details

Motivation: 现有自适应缓存常用与最终生成质量不完全对应的启发式差异度量，采用全程单一阈值，并不验证预算估计与真实近似误差是否一致，导致偏差持续累积。

Result: AlignCache 将去噪过程划分为结构不同的阶段，分别设置阈值及过时保护；在 classifier-free guidance 输出处测量反事实误差，并以 EMA 在线校准缓存决策边界；草稿合成结合方向与幅度门控的残差外推，以及锚点引导的无条件校正。作者称该闭环方法在多种先进视频生成模型上取得更好的加速—质量折衷。摘要未给出具体模型、加速率或生成质量数值。

Insight: 闭环测量最终输出误差，目标是让缓存调度直接对齐生成影响，而不只依赖特征变化启发式。摘要没有定量结果，因此无法比较其相对收益、测量开销及与其他加速技术组合时的表现。

Abstract: Feature caching has become a mainstream training-free strategy for accelerating video diffusion models by reusing intermediate outputs across denoising steps. However, existing adaptive methods rely on heuristic difference measures misaligned with actual generation quality, employ a single global threshold that ignores heterogeneous phase dynamics, and never validate their budget estimates against true approximation error, allowing systematic bias to propagate uncorrected. We propose \textbf{AlignCache}, a closed-loop caching framework that directly aligns scheduling decisions with their true influence on the final generated video through three tightly coupled components: (1) phase-adaptive budget routing that partitions the denoising trajectory into structurally distinct phases, each governed by an independent threshold and a staleness-aware guard; (2) closed-loop threshold calibration that exploits counterfactual error measurement at the classifier-free guided output level to adjust decision boundaries online via an EMA feedback loop; and (3) momentum-augmented draft synthesis combining direction-magnitude gated residual extrapolation with anchor-guided unconditional correction. By transforming scheduling from a static risk accumulator into a feedback-driven, self-correcting controller, AlignCache consistently achieves superior acceleration-quality trade-offs across state-of-the-art video generation models. Our code is available at https://github.com/jiangjyjy/AlignCache.

[127] VLA-GAUGE: Systematically Evaluating Generalization and Robustness in Vision-Language-Action Models cs.CV | cs.AI | cs.ROPDF

Jun Xiao, Rui Ye, Dong Cao, Jinyu Gu, Ying Wang

TL;DR: VLA-GAUGE 用统一可执行任务表示联合评估视觉语言动作模型的零样本泛化与组合扰动鲁棒性，并把诊断结果用于定向微调。

Details

Motivation: 熟悉操作任务上的高成功率未必反映真实零样本泛化或对多因素扰动的鲁棒性；现有评测缺少能力级诊断、组合扰动覆盖，以及从诊断到模型改进的闭环。

Result: 框架包含六个能力组的分布内/外配对任务、单因素敏感度分析和组合扰动筛查，并进行诊断驱动的数据增强与微调。统一评测 12 个不同规模和动作建模范式的 VLA：平均零样本分数仅 22.3%，SimpleVLA-RL 为 8.0%，在所评模型中倒数第二；三种代表性组合扰动配置的平均 R-Score 为 9.3。针对选定扰动进行诊断式微调后，π0.5 在留出场景的零样本泛化和鲁棒性分别提高 72.4、31.8 个计分点；后者超过 LIBERO-Plus 单因素微调的 10.9 分提升，且使用更少成功训练轨迹。

Insight: 熟悉任务的成功率可能掩盖明显的分布外与组合扰动弱点，诊断驱动的定向数据可比通用单因素微调更有效地修复指定问题。评测提升针对选定扰动和留出场景，摘要未说明分数换算尺度，也不能外推为所有扰动下的普遍改善。

Abstract: Vision-Language-Action models (VLAs) have achieved high success rates on familiar manipulation tasks, yet it remains unclear whether such scores reflect genuine zero-shot generalization and perturbation robustness. Existing evaluation frameworks provide limited capability-level diagnosis and combinatorial perturbation coverage, often without connecting diagnosis to targeted model refinement. We introduce VLA-GAUGE, a diagnostic framework that unifies both evaluations under a shared executable task representation. It combines (1) paired in-distribution and out-of-distribution tasks across six capability suites, (2) single-factor sensitivity analysis with combinatorial perturbation screening and evaluation, and (3) diagnosis-driven refinement through targeted data augmentation and model fine-tuning. We evaluate 12 VLAs spanning different model sizes and action-modeling paradigms under a unified protocol. We find that (1) The mean Zero-Shot Score on VLA-GAUGE is only 22.3%, with SimpleVLA-RL scoring 8.0% and ranking second to last among the evaluated models. (2) Under three representative combinatorial perturbation configurations, the mean R-Score is 9.3. (3) Diagnosis-driven fine-tuning improves $π_{0.5}$’s zero-shot generalization and robustness under a selected perturbation combination on held-out scenarios by 72.4 and 31.8 scoring points, respectively; the latter exceeds the 10.9-point gain from LIBERO-Plus single-factor fine-tuning despite fewer successful training trajectories. These results show that high reported success rates on familiar manipulation tasks do not guarantee strong zero-shot generalization to unseen tasks or robustness to combinatorial perturbations, while diagnostic evaluation provides actionable signals for targeted model improvement.

[128] CineRecon: Benchmarking Frontier Models for Agentic Cinematic Reconstruction cs.CVPDF

Jialei Wang, Zehan Wang, XiaoYang Wu, Ziang Zhang, Siyu Chen

TL;DR: CineRecon 是评测智能体能否把参考视频重建为可编辑动态 3D 场景的基准，并比较不同几何与外观细节要求下的重建能力。

Details

Motivation: 可编辑动态场景可为可控视频生成提供布局、动作和镜头代理，但现有评测尚不能系统衡量智能体从视频构建、动画化、渲染并迭代完善场景的能力。

Result: 基准包含 50 段电影、真实拍摄和游戏参考视频，以及 360 个标注评测片段；设定分为 Proxy、Simplified、Detailed，逐步提高几何与外观要求，同时保持空间关系、镜头行为和主体动态要求不变。评测清单把可观察属性和事件对应到实体与时间，由视觉语言评审比较时间对齐的参考/渲染帧，覆盖镜头、静态和动态一致性 11 类。对 12 个模型及相应智能体框架的评测发现，三类一致性均有失败；即使在 Proxy 设置中，主体动作与交互仍难重建。摘要未报告各模型具体分数。

Insight: 分层设置可区分事件复现与高细节建模能力，时间对齐评审也让失败定位更具体；不过当前评测中各系统仍存在明显动作短板。评分依赖视觉语言评审和清单设计，摘要未提供其与人工判断的一致性或模型排名。

Abstract: Reconstructing reference videos as editable dynamic 3D scenes is an important path toward controllable video generation for filmmaking: these scenes provide modifiable, renderable visual proxies that enable explicit control over scene layout, subject actions, and camera behavior. To evaluate agents’ ability to reconstruct editable dynamic 3D scenes from dynamic reference videos, we introduce CineRecon, a benchmark comprising 50 reference videos and 360 annotated evaluation clips. The reference videos span film, real-world footage, and games. Given a complete video, an agent uses tools to construct, animate, render, and refine its reconstruction. We evaluate this task under three representation settings: Proxy, Simplified, and Detailed. These settings increase geometric and appearance requirements while retaining the same requirements for spatial relations, camera behavior, and subject dynamics. This design tests whether agents can reproduce the events in a video when detailed appearance is unnecessary, and how performance changes as representation requirements increase. For evaluation, we construct checklists that associate observable properties and events with specific entities and times. A vision-language judge compares timestamp-aligned reference and rendered frames across 11 categories, grouped into camera, static, and dynamic consistency dimensions. We evaluate 12 models using their corresponding agent harnesses and find failures across all three dimensions. Subject motions and interactions remain difficult to reproduce even under the Proxy setting.

[129] Dense Endoluminal Mapping and Localisation from Monocular Bronchoscopy cs.CVPDF

Salvatore Esposito, Matias Mattamala, Daniel Rebain, Liang Zhao, Kev Dhaliwal

TL;DR: 该单目支气管镜 SLAM 系统仅用 RGB 视频进行气道内定位与表面重建，以微调深度估计提供米制尺度，并结合稠密匹配与自重建 ICP 稳定轨迹。

Details

Motivation: 支气管镜定位通常依赖电磁跟踪器和术前 CT；研究探索能否只用单目摄像流定位镜头并重建周围气道表面。

Result: Depth Anything 3 在超过百万张逼真气道渲染图上微调，输出米制逐像素深度，是系统唯一尺度来源；DINOv3 特征经交叉注意力匹配器提供黏膜上的亚像素对应，并与针对系统自身累积重建的投影 ICP 稠密约束共同进入因子图。在 6 条由患者 CT 渲染的气道上，平均轨迹绝对误差为 26.1 mm，平均表面 F-score 为 77.5%；平均表面精度超过 5 个近期单目 SLAM/前馈基线。允许用真值自由拟合尺度的 MASt3R-SLAM 轨迹误差更低，为 22.0 mm；ORB-SLAM3 只完成 6 条中的 2 条，而本文方法完成全部 6 条，并重建了 2 个离体标本。

Insight: 方法不依赖外部跟踪器，依靠学习深度和运行中的稠密几何约束弥补黏膜低纹理及部分方向弱可观测问题。评测中的 6 条气道由 CT 渲染，离体标本仅报告成功重建；尚不能据此推断真实患者在线临床定位的误差、安全性或鲁棒性。

Abstract: Bronchoscopy inserts a camera through the airways to diagnose endoluminal disease. Locating the camera along the airway usually requires additional equipment, such as an electromagnetic tracker and a preoperative CT scan. In this paper, we investigate the use of the monocular camera stream to localise the bronchoscope without requiring additional sensing. We introduce a monocular SLAM system that localises the tip of a bronchoscope and reconstructs the airway surface around it from RGB video alone. A Depth Anything 3 backbone, fine-tuned on over a million photorealistic airway renderings, returns metric per-pixel depth and is the system’s only source of scale. DINOv3 features routed through a custom cross-attention matcher supply sub-pixel correspondences on mucosa where classical descriptors fail. Both enter a factor graph alongside a dense constraint from projective iterative closest point (ICP) registration against the system’s own running reconstruction, so surface evidence stabilises the directions the visual factors leave weakly observable. On six airways rendered from patient CT scans, the system reaches 26.1 mm mean absolute trajectory error and 77.5% mean surface F-score. Its mean surface accuracy exceeds that of five recent monocular SLAM and feed-forward baselines. MASt3R-SLAM, which fits a free scale to the ground truth, has a lower mean trajectory error (22.0 mm). ORB-SLAM3 completes only two of the six airways, whereas our system completes all six and also reconstructs two ex vivo specimens.

[130] Code-Interactive Visual Reasoning for Financial Candlestick Understanding cs.CV | cs.LGPDF

Chenxi Zhang, Youwei Pang, Xiaoqi Zhao, Ziliang Gan, Rongjunchen Zhang

TL;DR: Fin-Thinker 让视觉语言模型生成并执行代码，从蜡烛图中提取证据、对齐跨日期和面板信息并完成金融计算，针对仅答对但缺乏图表依据的问题进行训练与评测。

Details

Motivation: 蜡烛图理解需要识别图形元素、跨时间和面板对齐，并结合数值及金融推理；只统计答案准确率无法判断结论是否有图表证据支持。

Result: 基于 Qwen3.5-9B，Fin-Thinker 使用 32,184 条经回放验证的代码—观察轨迹（FinKLine-32K），轨迹由教师引导学生探索得到，含成功解法与错误恢复。三阶段训练包括视觉—代码基础微调、金融过程监督，以及执行反馈强化学习。FinKLineBench 有 1,200 道问题、138 类任务，分别评估答案正确性、代码执行和证据 grounding。在该基准上，严格准确率比所评最佳直接回答基线高 13.1 个百分点，比带沙箱访问的基座模型高 33.2 个百分点；消融显示过程监督和执行反馈 RL 依次改善准确率及 GPT-5.6 评估的 grounded accuracy。

Insight: 执行代码获取可核验证据，使数值推理过程不只依赖模型对图表的口头描述。严格准确率增益是在作者特定基准和基线下得到；GPT 评审的 grounded accuracy 属于模型评审指标，摘要未提供人类验证结果。

Abstract: Financial candlestick understanding requires locating visual elements, aligning them across trading dates and panels, and combining observations through numerical and financial reasoning. Answer accuracy alone does not reveal whether predictions are supported by chart evidence. We introduce Code-Interactive Financial Candlestick Reasoning and present Fin-Thinker, a model based on Qwen3.5-9B that generates and executes code to acquire evidence and perform computations. The model uses the returned observations to guide subsequent reasoning. To train these capabilities, we construct FinKLine-32K with 32,184 replay-verified code-observation trajectories collected through teacher-guided student exploration. These include successful solutions with and without error recovery. Our three-stage training framework combines foundational visual-code fine-tuning, financial process supervision of successful reasoning and corrective continuations, and execution-feedback reinforcement learning. For evaluation, we introduce FinKLineBench with 1,200 questions across 138 financial task types to assess answer correctness, code execution, and evidence grounding separately. On FinKLineBench, Fin-Thinker improves strict accuracy by 13.1 and 33.2 percentage points over the best-performing direct-answer baseline among those evaluated and its base model with sandbox access, respectively. Training-stage ablations further show that financial process supervision and execution-feedback reinforcement learning yield successive improvements in both answer accuracy and GPT-5.6-judged grounded accuracy.

[131] JtT: Rethinking Point Trajectory Diffusion for Diverse and Rigid Motion from an Image cs.CV | cs.LGPDF

Gabrijel Boduljak, Yushi Lan, Christian Rupprecht, Andrea Vedaldi

TL;DR: JtT 用坐标空间扩散 Transformer 从单张图像生成任意稠密查询点的未来轨迹，发现逐帧速度参数化、合适时间步和查询点 RGB 有助于刚体运动预测。

Details

Motivation: 现有轨迹生成器即便扩大模型或使用真实数据，在简单合成场景中也难同时生成多样且物理合理的刚体运动；刚性结构又常由架构限制或深度等几何输入提供。

Result: JtT 不使用轨迹 VAE 或固定查询网格，而直接在坐标空间建模任意稠密点轨迹。系统实验发现逐帧速度扩散是捕捉刚体性的有效参数化，需配合合适时间步安排；在查询点加入全分辨率 RGB 可缓解边界失败。作者还发现训练中物理合理性与多样性并不共同提升：较大模型可能记住每张训练图像唯一真值未来，刚性更强但覆盖未来更少。严格的检查点平均评测下，约 5,000 万参数的 JtT 在分布与刚性指标上优于此前方法并接近真值刚性。摘要未列出具体指标值。

Insight: 单一真值监督可能让运动模型通过记忆换取刚性，从而牺牲多样性；模型排名也会随规模、指标和停止时点变化，因此采用检查点平均协议很重要。JtT 的优势基于该严格评测设置，摘要未说明未来多样性与刚性间的全部定量权衡。

Abstract: We study forecasting rigid-body motion from a single image: generating diverse and physically plausible future trajectories for arbitrary points in a scene, without building rigidity into the architecture or relying on geometric inputs such as depth. Like video generators, recent trajectory generators fail at this task, even in simple synthetic Kubric scenes, despite scaling up model size and training on real data. We show how to solve this with the right architecture and parameterization. Our model, JtT, is a simple diffusion transformer that operates directly in coordinate space and decodes arbitrary and dense query patterns, without the trajectory VAEs or fixed grids assumed in prior work. Through a systematic study of the design space of point trajectory diffusion, we find that diffusing per-frame velocities is the only parameterization that captures rigidity, that it must be paired with the right timestep schedule, and that adding full-resolution RGB at the query points resolves the boundary failures of prior methods. However, evaluating such 2D models rigorously is itself non-trivial. To our knowledge, we are the first to observe that physical plausibility and diversity do not improve together during training: larger models, as used in prior work, memorize the single ground-truth future per training image, becoming more rigid while covering fewer possible futures. Yet prior work has paid little attention to this, so we show how model rankings depend on model size, metric, and stopping time, and adopt a stricter, checkpoint-averaged evaluation protocol. Under this protocol, JtT improves over prior methods on both distributional and rigidity metrics and approaches ground-truth rigidity, with only about 50M parameters.

[132] Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models cs.CVPDF

Zhifang Zhang, Qi Wei, Weitong Chen, Olaf Maennel, Wei Emma Zhang

TL;DR: InertiaGuard 用视觉信念惯性度量多模态推理链中可疑的低熵答案翻转，并在测试时检测视觉触发的后门 CoT 劫持。

Details

Motivation: 多模态推理模型的后门可在触发图像出现时把原本忠实的推理链导向攻击者指定答案。面向直接作答模型的防御难以检测链内劫持，而劫持过程又可能伪装成正常自我纠错。

Result: 作者观察到，诚实推理通常在高熵“顿悟”时修改答案，后门则可能在近乎确定的推理步骤中推翻高置信答案。信念惯性定义为一段推理的累积 next-token 熵减去其引发的答案信念修订量；再只统计依赖图像的片段，得到 VBI。InertiaGuard 用少量干净轨迹校准阈值，低于阈值即报警，无需训练。跨多种攻击和基准，作者报告其检测性能优于既有基线，并可迁移至不同模型架构和触发器类型；摘要未提供具体检测率或误报率。

Insight: 检测关注“高度确定时为何突然翻转”，而不是仅搜索可疑文本表面模式，因此针对 CoT 内的后门行为设计。阈值需用干净轨迹校准，实际安全效果仍取决于攻击分布和误报代价；摘要未报告数值，不能据此比较部署可靠性。

Abstract: Chain-of-thought (CoT) reasoning enables multimodal large reasoning models (MLRMs) to solve complex visual problems, but also exposes them to CoT hijacking: an adversary implants a backdoor through poisoned fine-tuning so that a visual trigger induces a hijacking turn, redirecting an otherwise faithful reasoning chain to an attacker-chosen answer. However, since the hijack acts within the reasoning chain, existing multimodal backdoor defenses designed for MLLMs that answer directly struggle to detect it. Moreover, inspecting the chain itself does not reliably expose the hijack, since the hijacking turn can resemble honest self-correction in form. In this paper, we find that honest reasoning tends to revise its answer at high-entropy “aha moments”, whereas a hijack can overturn a confident answer through a near-deterministic turn. To quantify this contrast, we formalize this phenomenon as belief inertia, the cumulative next-token entropy of a reasoning span minus the answer-belief revision it induces, with markedly lower values for hijacked spans than for honest reasoning spans. Since hijacking tokens are also image-dependent, we further improve detection by restricting the measure to image-dependent spans, yielding visual belief inertia (VBI). Building on this, we propose InertiaGuard, a training-free, test-time detector that flags a trace when its VBI falls below a threshold calibrated on a small set of clean traces. Across various attacks and benchmarks, InertiaGuard achieves higher detection performance than existing baselines, while generalizing to different model architectures and trigger types.

[133] Answer It or Not? Towards Prompt-Adapted Adversarial Images on Vision Language Models cs.CV | cs.AIPDF

Shunyuan Bai, Yuchen Ren, Le Yang, Chenhao Lin, Chao Shen

TL;DR: ProAda 优化与问题提示相适配的对抗图像，使其对有害视觉隐私问题拒答，同时尽量保留对合法问题的正常回答。

Details

Motivation: 视觉隐私保护型对抗图像常与提示无关，导致模型对所有问题一概拒答，也会阻断合法使用；现有方法较少显式处理 VLM 对不同文本提示的反应。

Result: ProAda 在迭代优化图像扰动时同步更新文本提示，并用简单分类器判断提示是否仍属有害，以形成更精准的攻击方向。在视觉隐私相关数据集上，作者报告可拒绝超过 90% 的有害提示，同时允许超过 90% 的合法提示正常回答；整体表现优于提示无关的 VLM 对抗方法。摘要还称其具有有竞争力的跨模型迁移能力，并能更好抵抗反制措施，但未给出对应数值或具体防御设置。

Insight: 将有害/合法提示区分纳入扰动目标，可减少一味拒答带来的可用性损失。超过 90% 的比例来自作者的隐私数据集及其有害提示定义；跨模型和抗反制结论缺乏摘要中的定量细节。

Abstract: With the advancement in language modeling, research on adversarial images has recently shifted from image classifiers to vision-language models (VLMs). Despite the new, text-prompt-driven nature of VLMs, we find that existing VLM adversaries have largely overlooked the dynamics of text prompts. In a typical use case, adversarial images can protect visual privacy by rejecting VLMs from answering questions about the image. However, existing protections are prompt-agnostic, i.e., indiscriminately rejecting any question, including those for potential legitimate use of the image. To address this limitation, we propose Prompt-Adapted (ProAda) adversarial images, which are optimized to balance the rejection of harmful prompts and the correct response of legitimate prompts. In particular, to achieve a more accurate adversarial direction, the text prompt is also iteratively updated with a simple classifier judging whether the harmfulness of the prompt remains. Extensive experiments on visual-privacy-related datasets validate that ProAda can reject over 90% of harmful prompts while allowing over 90% of legitimate ones, demonstrating superior overall performance compared to existing prompt-agnostic VLM adversaries. In addition, our ProAda achieves competitive overall cross-model transferability and improved robustness against countermeasures.

[134] REEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory for Any Long-Horizon Multi-Shot Video Generation cs.CV | cs.AIPDF

Hang Xu, Guohui Zhang, Xiaoxiao Ma, Hu Yu, Jie Huang

TL;DR: REEL 学习跨镜头检索、生成和更新视觉记忆，并通过适配器将同一记忆策略接入不同黑盒视频生成器，以改善长视频的身份与状态连续性。

Details

Motivation: 长视频生成需要兼顾用户控制和跨镜头一致性，但现有系统常将记忆绑定到特定生成器接口或预设工作流，缺少为跨镜头目标独立训练的通用记忆策略。

Result: REEL 用统一工具接口检索参考、创建缺失视觉证据并更新持久记忆；特定后端适配器再将证据计划转成生成器所需条件输入。训练包括监督微调、单镜头 GRPO 和多镜头 GRPO，后者在连续镜头间携带策略生成的记忆，使后续结果反向指导前序参考选择与记忆更新。REEL-Bench 测试文本请求及定时上传参考图像条件下的身份保持和状态变化。在 JoyAI-Echo 1.5、MiniMax-H3 上，REEL 相比无记忆基线，ST-Bench 跨镜头一致性分别提高 15.1%、20.7%；自建 REEL-Bench 的 Core-100 分数分别提高 2.41、8.32 分。

Insight: 把记忆策略与生成器后端解耦，能让单一策略为不同黑盒生成模型规划证据；多镜头训练也显式学习记忆跨镜头传递。结果比较的是无记忆基线，摘要未给出与专用记忆方法的直接比较或记忆策略的额外推理成本。

Abstract: Memory-based long-video generation aims to combine user control with long-range visual consistency. We organize this process into three functional components: a planner, an adapter, and a generator. Representative multi-shot systems couple memory to generator-specific interfaces or use predefined agent workflows, rather than independently training a memory policy for cross-shot objectives. We introduce REEL, a general-purpose visual-memory planner. Through a unified tool interface, REEL learns to retrieve references, create missing visual evidence, and update persistent memory. Backend-specific adapters translate its evidence plans into supported conditioning inputs, allowing the same policy to guide different black-box video generators. To train REEL, we construct two high-quality datasets, REEL-SFT-4k and REEL-RL-1k, covering diverse interaction patterns and prompt types. Training proceeds in three stages: supervised fine-tuning, single-shot Group Relative Policy Optimization (GRPO), and multi-shot GRPO. Multi-shot GRPO carries policy-generated memory across consecutive shots, so later outcomes guide earlier reference selection and memory updates. REEL-Bench tests identity preservation and intended state changes under text-only requests and timed reference-image uploads. On JoyAI-Echo~1.5 and MiniMax-H3, REEL improves cross-shot consistency on ST-Bench by 15.1% and 20.7%, and Core-100 on our proposed REEL-Bench by 2.41 and 8.32 points, respectively, over the no-memory baseline.

[135] Revisiting Natural Language Prompts from Large Language Models for Open-Vocabulary Video Recognition cs.CVPDF

Haichen He, Weibin Liu, Weiwei Xing

TL;DR: P2CLIP 从类别、视频和帧三个层面设计开放词汇视频识别提示，利用判别式类别语义和知识引导时间聚合改善图文对齐。

Details

Motivation: 直接在视频识别数据上微调 CLIP 容易因文本信息少而过拟合；用 LLM 扩写提示虽常见，但简单指令可能让类别语义过度泛化、边界模糊，且未充分利用视频帧级与整体语境。

Result: P2CLIP 在类别层面用 LLM 生成兼具丰富语义和区分性的 Generic-Discriminative Prompt；视频及帧层面由 VLM/LLM 生成对应描述，再用知识引导时间聚合让 LLM 专家知识参与视频时序建模及联合视觉文本空间构造。作者在多个基准的零样本与少样本设置中报告方法有效。摘要未给出基准名称、具体分数或相对增益。

Insight: 提示质量不只取决于信息丰富程度，还需要保留类别间判别边界；分层使用类别语义和时间描述，尝试同时处理静态语义与动态信息。当前摘要未量化效果，无法判断各提示组件的独立贡献或对不同类别规模的适用性。

Abstract: Vision-language models (VLMs) like CLIP have shown great success in image tasks, however, applying CLIP to open-vocabulary video recognition poses challenges. Due to the limited text information, directly fine-tuning CLIP on video recognition datasets may lead to overfitting and hinder its generalizability. A prevalent method is to enrich text information by generating prompts using Large Language Models (LLMs). However, limited efforts have been made to explore how to generate higher-quality prompts that can contribute to classify unseen categories, and how to further enhance vision-text alignment based on these prompts. In this paper, we propose P2CLIP (Prompt-guided CLIP), which aims to bridge this gap by exploring the impact of prompts across three aspects: category, video, and frame. At the category level, we argue that LLM-generated prompts under simple instructions may suffer from semantic overgeneralization, which can blur the category boundaries and hinder the performance of the model. To address this issue, we design Generic-Discriminative Prompt using LLMs to obtain discriminative category embeddings with rich semantic information. At the video and frame levels, we first use VLM and LLM to generate frame and video descriptions for enhancing vision-text alignment. Then we propose a Knowledge-Guided Temporal Aggregation mechanism, which leverages the expert knowledge derived from LLMs to guide the temporal modeling of videos and encourage a better construction of the joint visual-textual feature space. Extensive experiments across various benchmarks, including zero-shot and few-shot settings, demonstrate the effectiveness of our method.

[136] EvoSpatial: Experience-Driven Skill and Tool Evolution for Spatial Reasoning Agents cs.CVPDF

Bin Yu, Shijie Lian, Zhaolong Shen, Xiaopeng Lin, Yao Zhang

TL;DR: EvoSpatial 将成功与失败的空间推理轨迹转化为可复用技能和计算工具，并通过任务结果反馈持续改进技能库。

Details

Motivation: 视觉语言智能体利用工具执行空间测量时，还需跟踪几何假设并检查中间证据；只保存答案或工具调用，难以保留计算步骤及其适用条件。

Result: EvoSpatial 在推理时检索技能、复用工具，并可生成可复用辅助函数；提交答案后，再用监督结果反馈优化技能和整合工具。作者在 20 个基准上报告其平均得分居所比较空间智能体首位，比 SpatialClaw 高 4.0 个百分点。其他实验显示技能可跨基准迁移，并可通过技能引导蒸馏提升推理，即使评测时不提供外部技能。摘要未说明平均分口径及各基准的单项差异。

Insight: 把计算程序与几何假设、何时信任结果的经验一起沉淀，使技能既服务推理时复用，也可用于训练。平均领先幅度不能说明各任务一致受益；跨基准迁移和无外部技能蒸馏的具体协议需结合正文核实。

Abstract: Tool-augmented vision-language agents compose perception and measurement to reason about spatial relationships, but effective execution also requires tracking geometric assumptions and checking intermediate evidence. Preserving this experience across tasks requires retaining procedural knowledge together with the computations it governs. We introduce EvoSpatial, an experience-driven framework that evolves a library of linked skills and tools. Our key insight is that successful and failed spatial trajectories reveal both how to compute and when to trust the result. EvoSpatial retrieves skills, reuses tools, and synthesizes reusable computational functions (generated helpers) during reasoning. After answer submission, supervised outcome feedback guides skill refinement and tool consolidation for subsequent tasks. Across 20 benchmarks, EvoSpatial attains the highest average score among the compared spatial agents, exceeding SpatialClaw by 4.0 percentage points. Further experiments show cross-benchmark transfer and improved reasoning through skill-guided distillation without external skills at evaluation. These results support accumulating spatial experience for both inference-time reuse and trajectory-based training.

[137] 3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification cs.CVPDF

William Grolleau, Astrid Sabourin, Guillaume Lapouge, Catherine Achard

TL;DR: 3D-VITAL 从单视角图像重建可渲染 3D 网格以合成新视角，并用跨视角一致性筛除错误几何、用可见性图减少幻觉纹理对 ReID 的干扰。

Details

Motivation: 单视角航空—地面行人重识别需要跨大视角变化匹配身份。新视角合成可扩充训练数据，但重建几何可能错误；即使几何正确，源图不可见区域的渲染纹理仍可能是幻觉。

Result: 作者利用重建与渲染过程中得到的逐像素源视角可见性图，构建两阶段方法：CVCF 在可见区域比较源图与合成视角的一致性，剔除几何错误网格；VAAS 在训练时引导模型关注合成视图中可靠区域。3D-VITAL 在三个单视角航空—地面 ReID 基准上达到新的最佳结果。摘要未列出基准名称、指标值、提升幅度或消融结果。

Insight: 合成视角不应被当作同等可信的完整观测；区分真实可见区域和重建补全区域，能过滤结构错误并降低伪纹理影响。当前摘要仅定性报告三个基准上的领先，无法评估绝对性能和合成数据质量控制的各自贡献。

Abstract: We address the Cross-View Object Re-Identification (ReID) task, aiming to match identities across drastically different viewpoints. We focus on Single-View Aerial-Ground ReID (SV AG-ReID), where models train on one viewpoint and must generalize to another at inference. To bridge this viewpoint gap, related work employs generative models to synthesize novel views from source images as additional training data. We leverage recent advances in large-scale single-view 3D mesh reconstruction to lift source images into textured meshes and to render them from new viewpoints. However, this pipeline produces unreliable data in two distinct ways. First, the reconstruction may fail and produce geometrically incorrect meshes, resulting in severe corruption in rendered novel views. Second, even when the geometry is correct, the rendered novel views contain hallucinated textures in regions not visible in the source image. The reconstructed mesh carries a per-pixel partition between regions visible from the source view and regions that the reconstruction model had to hallucinate, available as a deterministic byproduct of rendering. We introduce this per-pixel visibility map and exploit it in a two-stage design. Cross-View Consistency Filtering (CVCF) discards geometrically incorrect meshes by ensuring multi-view consistency between source and novel views over visible patches. Visibility-Aware Attention Supervision (VAAS) leverages the visibility map to guide the model’s attention toward reliable regions of the rendered view during training. Combined, these two components enable our method, 3D-VITAL, to better leverage the synthetic signal from single-view 3D mesh reconstruction. 3D-VITAL sets a new state-of-the-art on three SV AG-ReID benchmarks. Code and data will be released at https://github.com/TurtleSmoke/3D-VITAL.

[138] VAttnPO: Visual Attention-Guided Policy Optimization cs.CV | cs.AIPDF

Lihao Zheng, Zhenwei Shao, Zhou Yu, Yan Yang, Haoyi Sun

TL;DR: VAttnPO 将正确答案条件下的视觉注意分布与 rollout token 的注意分布比较，生成零均值、按位置调整的 token 级信用修正，同时保留验证器的轨迹级奖励。

Details

Motivation: 带可验证奖励的多模态强化学习常把同一个序列优势分配给整条回答，无法显式反映生成过程中各 token 对视觉证据的利用变化。

Result: 冻结旧策略在 teacher-forcing 正确答案条件下为每个 rollout token 构造视觉注意参考；VAttnPO 用 Jensen–Shannon 散度比较注意分布，而非使用教师输出 token 概率或另设注意力匹配目标。轨迹中心化、相对位置加权和信用平衡将比较结果转为序列优势上的零均值位置感知修正。在 Qwen3-VL-8B-Instruct 上，六任务图像 AVE 从 GRPO 的 70.97 提至 73.50，且所有报告的结构化、视频和 MME 指标均高于 GRPO；Qwen2.5-VL-7B-Instruct、InternVL3-8B-Instruct 上也报告了相对 GRPO 的增益。

Insight: 该方法用注意力一致性提供细粒度信用信号，但保留原有轨迹平均优势；诊断实验还试图验证其与模型行为、视觉对应及 token 位置的关系。摘要未报告各指标的数值、统计显著性及额外计算开销；AVE 的计分范围也未交代。

Abstract: Multimodal reinforcement learning with verifiable rewards commonly assigns one sequence advantage to an entire response. This outcome-level signal distinguishes overall response quality but leaves changes in visual-evidence use across generated tokens implicit. We introduce VAttnPO (Visual Attention-Guided Policy Optimization), which uses answer-conditioned visual attention to refine token-level credit while preserving the verifier-derived trajectory-mean advantage. A frozen old policy constructs a visual reference by teacher forcing the known correct answer and produces a visual distribution for each rollout token. VAttnPO measures their agreement in visual-attention space using Jensen–Shannon divergence, rather than reading guidance from teacher output-token distributions or introducing an attention-matching objective. Trajectory centering, relative-position weighting, and credit balancing turn this agreement into a zero-mean, position-aware correction to the GRPO sequence advantage. Token intervention, visual-index permutation, and trajectory analysis provide diagnostic evidence linking this signal to model behavior, visual correspondence, and generation position. On Qwen3-VL-8B-Instruct, VAttnPO raises six-task image AVE from 70.97 with GRPO to 73.50 and scores higher on every reported structured, video, and MME metric. Gains over GRPO also hold on Qwen2.5-VL-7B-Instruct and InternVL3-8B-Instruct.

[139] InfoVis: Unified Slide Image Generation and Editing cs.CV | cs.AI | cs.LGPDF

Jiahao Wang, Xiyan Jiang, Chenyuan Wu, Haoqiang Zhang, Xin Luo

TL;DR: InfoVis 提供 9B 统一幻灯片生成与编辑模型，并配套百万级持续预训练/监督数据和多维评测基准，覆盖文本呈现、指令遵循、视觉定位与编辑。

Details

Motivation: 幻灯片生成和编辑需要在同一画布中同时处理文字准确性、指令执行、视觉 grounding、审美设计和精确修改；现有工作通常只覆盖部分能力，也缺少规模化 slide 原生监督和诊断式评估。

Result: 作者提出四类幻灯片任务，构建超过 1,300 万条持续预训练记录、190 万条监督微调记录，以及含 1,527 个人工筛选实例的 VisSlideBench。基准自动排名与盲法分维度人工排序一致。InfoVis 在 11 项指标上均领先所比较开源模型，总分 7.94，接近 Nano-Banana-2 的 7.99；从 FLUX.2-klein-base-9B 开始的幻灯片专用后训练使总分提高 1.6 分，内容忠实度提升最大，而与专有模型的差距在审美方面最大。文档转幻灯片流程在外部 SlidesGen-Bench 的审美分排名第二。

Insight: 工作同时建设数据、基准和统一模型，揭示专用后训练可改善内容忠实度，但审美仍是相对短板。总分差距较小不代表每一维持平；模型评测、主观审美排名和外部工作流结果应分别解读。

Abstract: Generating and editing presentation slides requires models to combine accurate text rendering, instruction following, visual grounding, aesthetic design, and precise editing within one canvas. Existing work covers only subsets of these capabilities and lacks large-scale slide-native supervision and diagnostic evaluation. We formulate four slide-centric tasks and build a scalable data engine with over 13M continual-pretraining records and 1.9M supervised fine-tuning records. We also introduce VisSlideBench, a benchmark of 1,527 manually screened instances that covers all four tasks with task-specific, multidimensional evaluation. Its automatic rankings agree with blinded, dimension-wise human rankings. Using these resources, we train InfoVis, a unified 9B slide generation and editing model. On VisSlideBench, InfoVis leads every evaluated open-source model on all eleven metrics and scores 7.94 overall against 7.99 for Nano-Banana-2. Starting from FLUX.2-klein-base-9B, slide-specific post-training adds 1.6 points overall, with the largest generation gains in content faithfulness; the remaining gap to proprietary models is smallest in content faithfulness and largest in aesthetic quality. Built on InfoVis, our document-to-slide workflow achieves the second-highest Aesthetics score on the external SlidesGen-Bench. We will release the model, code, data, and benchmark.

[140] FoveaTree: Information-Density Trees for Query-Guided Visual Token Pruning in Video Large Language Models cs.CVPDF

Zihan Meng, Weidong Chen, Dexiang Hong, Xinyan Liu, Zhendong Mao

TL;DR: FoveaTree 先估计区域合并的信息损失建立四叉树密度图，再利用存活 token 的树结构调整文本到视觉注意，训练免除地剪枝视频模型视觉 token。

Details

Motivation: 视频大模型每帧会产生数百视觉 token，推理成本高。相似度合并或注意力筛选可能忽略区域可压缩性而丢失关键信息，位置效应和注意力汇又可能让文本—视觉注意偏离问题相关区域。

Result: 第一阶段 Merge-Gain Density Estimation 衡量区域合并带来的特征失真并构建四叉树密度图，据此决定丢弃、合并或保留 token；第二阶段 Attention Tree Pruning 在中间 LLM 层利用树拓扑修正文本—视觉注意，抑制低相关内容并保留相关区域。作者在 InternVL3.5-8B、Qwen3-VL-8B 上测试五个视频问答基准和一个视频描述基准，视觉 token 保留率为 50%、30%、10%；在所评设置中报告整体性能达到当时最佳并具有较好的准确率—效率折衷。摘要未提供各设置的具体分数。

Insight: 区域级合并损失和树结构注意共同考虑空间层级，避免只按 token 相似度或原始注意力分数剪枝。所谓整体领先限定于所列骨干、六个基准与三档保留率；摘要未量化延迟、显存节省或最激进压缩下的质量下降。

Abstract: Video large language models (Video LLMs) encode each frame into hundreds of visual tokens, making video inference expensive. Existing methods reduce this cost through similarity-based merging or attention-based selection but face two challenges: Hierarchy-blind Pruning, which overlooks regional compressibility and can discard critical evidence, and Attention Drift, where positional effects and attention sinks shift text-to-vision attention away from task relevance. We propose FoveaTree, a two-stage, training-free visual-token pruning framework. First, Merge-Gain Density Estimation measures the feature distortion caused by regional merging and constructs a quadtree-derived density map to guide token dropping, merging, and retention. Second, Attention Tree Pruning uses the surviving units and inherited tree topology to refine text-to-vision attention in intermediate LLM layers, preserving relevant regional evidence while removing low-relevance content. We evaluate FoveaTree on InternVL3.5-8B and Qwen3-VL-8B across five video question-answering benchmarks and one captioning benchmark at 50%, 30%, and 10% visual-token retention. FoveaTree consistently achieves overall state-of-the-art performance across the evaluated settings, offering a strong accuracy-efficiency trade-off.

[141] Towards Complementary Keypoint Detection via Mixture of Detectors cs.CVPDF

Xiaoyong Lu, Guobao Xiao, Dong Liang, Songlin Du

TL;DR: MoD 将异构关键点检测器蒸馏到统一模型，再按图像局部区域和关键点预算动态选择检测器，以匹配下游几何匹配需求。

Details

Motivation: 不同关键点检测器对场景结构和点数预算各有优势，依赖单一检测器可能产生分布不佳或任务表现欠优的关键点集合。

Result: MoD 先将多个检测器蒸馏到统一模型，使其一次前向就能预测多种检测响应；之后由轻量路由器按 patch 和预算选择最适合该区域的检测器。路由器用下游匹配质量产生的任务奖励训练，以同时考虑关键点数量和精度。在 MegaDepth、ScanNet、HPatches 上，作者报告该方法在多种关键点预算下持续超过强基线，并改善关键点分布和下游几何表现。摘要未提供各预算下的指标或提升数值。

Insight: 用下游匹配任务来监督局部检测器选择，使关键点组合能适应实际几何任务，而非优化单一检测器自身指标。多基准领先仅有定性描述；蒸馏多检测器的额外复杂度、路由收益和不同场景的选择偏好需看详细消融。

Abstract: Keypoint detection plays a central role in geometric vision pipelines. Existing detectors exhibit complementary strengths across different scene structures and keypoint budgets, and reliance on a single detector often results in suboptimal keypoint distributions. Motivated by this observation, we propose \textbf{Mixture of Detectors (MoD)} for dense matching pipelines, a unified framework for complementary keypoint detection through the adaptive composition of heterogeneous detectors. MoD first distills heterogeneous detectors into a unified model, allowing diverse detection maps to be predicted efficiently within a single forward pass. Detector selection is subsequently formulated as a patch-wise, budget-conditioned decision process, in which a lightweight router assigns the most suitable detector to each local region. To avoid heuristic supervision, the router is trained with task-driven rewards derived from downstream matching quality, jointly capturing both keypoint quantity and precision. Extensive experiments on MegaDepth, ScanNet, and HPatches demonstrate that MoD consistently outperforms strong baselines across a wide range of keypoint budgets, demonstrating its effectiveness in improving keypoint distribution and downstream geometric performance.

[142] ViewStruct: View-Grounded Articulation and Connectivity Inference cs.CVPDF

Wenjing Bian, Ranran Huang, Andreas Geiger

TL;DR: ViewStruct 根据 RGB 图像及对齐的部件分割和深度联合预测物体刚体分组、部件连接与关节参数，并把结果锚定到实际观察视角。

Details

Motivation: 与陌生物体互动需要知道部件如何连接、如何运动；依赖类别规范朝向或观察到的关节运动会限制对新物体、新视角和真实场景的结构推断。

Result: ViewStruct 还可选用额外图像或视频，输出刚体分组、附件关系及关节参数。视角对齐表示将结构预测对应到图像区域，并在重力对齐的视角相对坐标系中表达关节几何，无需类别特定的标准朝向。基于预训练视觉语言模型的适配可在数秒内预测，不需要观测关节运动或逐实例优化。即使仅用合成任务数据训练，作者仍报告在结构预测基线上提升，并能迁移到未见类别、资产来源和真实人机交互视频；模拟操作实验显示预测关节可用于生成机器人运动轨迹。摘要未提供基准、分数或提升数值。

Insight: 把关节预测关联到可见区域，并使用视角相对几何，有助于减少对类别先验和运动观测的依赖。合成训练到真实视频的迁移结论目前只有摘要定性支持；机器人轨迹结果来自模拟，尚非实体机器人实验。

Abstract: Interacting with unfamiliar objects requires understanding how their parts connect and move. We introduce ViewStruct, a framework for jointly predicting connectivity and articulation in objects and scenes. Given an RGB image with aligned part segmentation and depth, optionally supplemented by images or video, ViewStruct predicts rigid-body groupings, attachments, and joint parameters. Its view-grounded representation links predictions to observed regions and expresses joint geometry in a gravity-aligned, view-relative frame, without requiring category-specific canonical orientations. A pretrained vision–language model is adapted to predict structure in seconds, without requiring observed articulation motion or per-instance optimisation. Experiments demonstrate improved structural prediction over evaluated baselines and transfer to unseen categories, asset sources, and real-world human-interaction videos, despite task-specific training exclusively on synthetic data. Simulated manipulation experiments further demonstrate the utility of predicted joints for generating robot motion trajectories.

[143] Carve VFM: Carving Specialists out of Generalists, Automatic Task-Specific Compression of Vision Foundation Models cs.CV | cs.LGPDF

Lotfi Abdelkrim Mecharbat, Abdelrahman Elsayed, Ibrahim Almakky, Martin Takáč, Mohammad Yaqub

TL;DR: Carve VFM 根据任务敏感度和剩余权重对被剪除行为的可恢复性，在视觉基础模型中直接搜索并恢复任务专用的稠密子网络。

Details

Motivation: 视觉基础模型可迁移但推理成本高；微调不缩小模型，蒸馏需要昂贵教师训练，结构化剪枝则常把层间容量选择与层内行为恢复分开，激进压缩时容易失败。

Result: 给定基础模型、校准集和参数预算，Carve VFM 搜索任务专用的层间容量分配，以恢复感知适应度引导搜索，并通过迭代闭式低秩修正将被删行为合并回保留权重。单 GPU 上数分钟即可生成稠密编码器，无需重训骨干。在 12 个基准、DINOv3 和 MetaCLIP-2 的 4 个模型上，它在 8 种骨干—预算配置中的 7 种取得被评估稠密结构剪枝方法最高平均准确率。进一步微调后准确率为 88.1%，超过所评最强边缘模型的 85.8%；裁剪模型将计算与内存降低，GPU 吞吐为未压缩模型的 1.3–2.0 倍。

Insight: 剪枝选择和行为恢复需共同优化：层容量按任务敏感度分配，删除影响则取决于剩余权重能否补偿。比较范围是所评四个模型、预算和稠密剪枝基线；微调后的准确率和未经微调的效率结果不应混为同一部署配置。

Abstract: Vision foundation models (VFMs) provide transferable visual representations, but their inference cost limits deployment in many settings. Fine-tuning specializes a VFM without shrinking it; distillation shrinks it, but through expensive teacher-guided training. Structured pruning offers a more direct route by extracting a compact subnetwork from the VFM. However, as compression gets aggressive, existing methods fail because they treat subnetwork selection and recovery as separate axes. Across layers, capacity must follow task sensitivity rather than spread uniformly. Within a layer, a removal cannot be judged by importance alone; its impact also depends on how well the lost contribution can be recovered from what remains. Selection and recovery must therefore be decided jointly. We introduce Carve VFM, which resolves this interdependence. Given a VFM, a calibration set, and a parameter budget, it (i) searches for task-specific capacity allocations across layers; (ii) guides the search via an adaptive recovery-aware fitness criterion; and (iii) reconstructs removed behavior through iterative closed-form low-rank corrections merged into the surviving weights. The result is a dense encoder produced in minutes on a single GPU without backbone retraining. Across twelve benchmarks and four models from DINOv3 and MetaCLIP-2, Carve VFM achieves the highest reported mean accuracy in seven of eight backbone-budget configurations among the evaluated dense structured-pruning methods. Further fine-tuning surpasses the strongest evaluated edge model (88.1% vs. 85.8%), and carved models cut compute and memory while delivering 1.3-2.0x higher GPU throughput than the uncompressed one. These results show that the compact model you need is inside the VFM; it only has to be carved out.

[144] X-Ragent: Tool-Augmented Reinforcement Learning for Evidence-Grounded Chest X-Ray Reasoning cs.CVPDF

Yi Wang, Zhenlong Yuan, Yuchan Guo, Maolei Huang, Zijian Song

TL;DR: X-Ragent 通过选择性寻找辅助证据、对定量测量作有效性门控，并以准确率门控的 GRPO 训练，提升胸片推理的证据支撑。

Details

Motivation: 胸部 X 光诊断需将视觉征象与任务相关证据结合。既有工具增强医学智能体通常缺少学习何时调用辅助证据的机制，也可能将不适用的定量测量带入诊断。

Result: 作者构建含 7,700 条样本的 Rad-CoT，融合按因果顺序组织的工具增强轨迹与报告式诊断推理，并用智能体监督微调初始化 Qwen3-VL-8B。随后以 GRPO 优化策略，使用准确率门控奖励：只有诊断正确的轨迹才获得测量一致性和辅助工具成本等过程项，避免“过程看似合理”补偿错误诊断。在三个胸片基准上，X-Ragent 报告取得最强结果；完全留出的 ChestAgentBench 上提升 8.6 分。摘要未说明该分数具体指标及其他基准的增益。

Insight: 准确率门控让工具使用的过程奖励不能掩盖最终诊断错误，并将测量有效性与是否调用工具分开处理。该结果来自胸片基准，不能直接外推到其他影像模态或临床实践；摘要也未说明工具调用成本的具体变化。

Abstract: Chest X-ray (CXR) interpretation often requires integrating visual findings with task-specific evidence rather than relying on direct visual recognition alone. While tool-augmented medical agents can supply such evidence via external modules, they often lack explicit mechanisms for learning when auxiliary evidence is needed and for preventing inapplicable quantitative evidence from entering diagnostic reasoning. To address this, we propose X-Ragent, a post-training framework for tool-augmented reinforcement learning that decouples learned auxiliary-evidence seeking from validity-gated quantitative measurement acquisition for evidence-grounded CXR reasoning. We construct Rad-CoT, a 7.7K-sample corpus combining causally ordered tool-augmented trajectories with report-style diagnostic reasoning, and use it to initialize Qwen3-VL-8B via agentic supervised fine-tuning (SFT). We then optimize the policy with Group Relative Policy Optimization (GRPO) using an accuracy-gated reward that applies process-level terms for measurement consistency and auxiliary-tool cost only to correct trajectories, preventing plausible process behavior from compensating for an incorrect diagnosis. Across three CXR benchmarks, X-Ragent achieves the strongest results, including an 8.6-point improvement on the fully held-out ChestAgentBench. By seeking evidence selectively and admitting only valid measurements, X-Ragent supports more faithful and evidence-grounded CXR reasoning.

[145] Growing to See: Efficient Pretraining of Self-Supervised Vision Transformers cs.CV | cs.AI | cs.LGPDF

Amir Mohammad Karimi Mamaghan, Ferdinand Kapl, Andrea Dittadi, Karl Henrik Johansson, Kaitlin Maile

TL;DR: 该研究系统考察自监督视觉 Transformer 的渐进深度增长，发现边训练边插入层可同时降低预训练计算并改善全局与密集下游任务表现。

Details

Motivation: 渐进深度增长已在大型语言模型训练中受到关注，但在 MAE 和 DINOv3 等自监督视觉 Transformer 中，它如何影响表征质量、内部计算模式和计算效率尚不清楚。

Result: 作者对 MAE 与 DINOv3 研究逐步增长网络深度对下游性能、中间层表示和计算过程的影响。全局和密集任务上，增长训练 FLOPs 降低 13.9%–24.8%，ImageNet 线性探针准确率最高提高 3.36 个百分点。中间层下游表现也改善；内部模式方面，MAE 出现全局与局部注意力交替，DINOv3 的局部到全局注意力过渡更早。

Insight: 渐进增长不仅节省训练计算，也会改变不同层级表征和注意力计算的组织方式，因此收益不只是更快完成预训练。报告涵盖 MAE、DINOv3 及多类任务，摘要未提供各增长策略的详细开销或跨架构适用边界。

Abstract: Progressive depth growth builds large networks by starting with shallow models and inserting layers during training. Recent advances in large language models have renewed interest in this longstanding approach, demonstrating benefits for training efficiency and reasoning alongside changes in internal computations. However, how growing affects representation quality and computational patterns in self-supervised vision transformers remains unclear. In this work, we systematically investigate progressive depth growth in two popular self-supervised families, namely masked autoencoders (MAE) and DINOv3, and analyze how growing affects downstream performance, intermediate representations, and internal computations. Across global and dense tasks, we show that growing improves downstream performance while reducing training FLOPs by 13.9–24.8%, with ImageNet linear-probe accuracy gains of up to 3.36 percentage points. Growing also improves downstream performance at intermediate layers and reshapes internal computational patterns, introducing recurring alternations between global and localized attention in MAE and earlier local-to-global attention progression in self-distilled DINOv3. Together, these findings establish progressive depth growth as an effective and efficient pretraining strategy for improving downstream performance across global and dense visual tasks with less computation, and advance our understanding of how growing shapes visual computation.

[146] Towards Cost-Efficient Video Dataset Construction with Large Language Models cs.CV | cs.AIPDF

Mohammed Shabot, Jonathan Adams, Jan Van den Stock, Joost Vennekens

TL;DR: 该研究把 LLM 元数据预筛加入视频数据集构建流程，在下载候选视频前判定相关性，以降低存储、时间和标注成本。

Details

Motivation: 现有视频数据筛选通常要先下载候选片段才能判断是否相关，造成不必要的下载和存储开销，也增加人工标注负担。

Result: 作者提出基于元数据的 LLM 预筛，并与两种不使用 LLM 的关键词启发式基线比较。人工参考标签由四名独立标注者的严格多数票形成，使用 Gwet’s AC1 检查标注者一致性；通过留一标注者（LOO）和加入一名标注者（AORI）机制评估 LLM 作为虚拟标注者的表现，并报告相对既有流程的成本对比。摘要未给出 LLM 名称、筛选准确率、一致性值或实际成本节省。

Insight: 在下载前利用元数据筛选有机会把视频获取成本前移为较廉价的相关性判断；多标注者和 LOO/AORI 设计也尝试控制单一标注者基准的不确定性。由于摘要缺少结果数值，尚无法判断 LLM 预筛是否优于关键词规则或在哪些数据集更划算。

Abstract: Research on video classification models has witnessed significant growth recently, accompanied by a significant increase in the demand for datasets to train such models. Existing video dataset curation approaches share a common bottleneck: candidate videos must be downloaded before any relevance judgment is made, which results in storage, time and annotation costs. In this work, we present a filtering step, using Large Language Models (LLMs) to pre-screen videos based on their metadata, before actually downloading these videos, thereby saving on these costs. We evaluate the LLMs against a strict human majority vote derived from four independent annotators and validate the reliability of the annotations through inter-annotator agreement analysis (Gwet’s AC1). We compare the LLM pre-filter against two no-LLM baselines, namely two keyword heuristics. Via a Leave-One-Rater-Out (LOO) mechanism and an Add-One-Rater-In (AORI), we evaluate the performance of the LLMs as virtual annotators. Finally, we report cost comparisons relative to existing approaches. To ensure reproducability, we made our implementation available on github.

Hongsen Cao, Mona Jaber, Shanxin Yuan, Ahmed Sayed

TL;DR: 该研究通过挖掘 40 万个 nuScenes、KITTI、Waymo 目标的相机/LiDAR 检测差异，发现模态可靠性依目标类别、距离、方向及其交互变化，并据此增强融合检测器。

Details

Motivation: 多模态 3D 检测通常假设相机与 LiDAR 互补，但融合模型的权重如何随空间位置变化并不透明，也难以在部署前审计。

Result: 作者按每个真值目标标记由哪些模态检出，并以空间分层特征归因分析距离条件下的 Shapley 贡献。发现模态偏好边界与目标类别有关：交通锥仅相机检出的概率约为卡车的 5.8 倍（16.8% 对 2.9%），且几乎不依赖距离；相机单独检出在 10–30 米最常见。距离与点云密度交互的影响超过视角主效应；车后目标比车前目标更依赖 LiDAR。将这些规律汇总为模态可靠性图并通过轻量注意力层注入四种预训练融合检测器，在只微调不到 4% 参数、耗时 2–3 GPU 小时的条件下，mAP 最高提升 1.3。

Insight: 模态可靠性并非单一全局权重，而会随类别、距离、方位和密度交互变化；分层归因可暴露全局平均隐藏的结构。增强收益最高为 1.3 mAP，摘要未说明各检测器的逐项提升，部署时也需确认训练场景是否覆盖目标分布。

Abstract: Multi-modal fusion for 3D object detection assumes that cameras and LiDAR are broadly complementary. In practice, how much each modality is trusted at a given location is decided by learned weights that are neither characterized nor auditable. We label every ground-truth object by which modalities detected it, then mine the resulting 0.4M-object database across nuScenes, KITTI, and Waymo. Because the patterns shift with distance from the ego vehicle, we read the mined model through Spatially Stratified Feature Attribution (SSFA), which conditions Shapley attribution on distance zones and exposes structure that global attribution averages away. Three assumptions built into the current fusion design do not hold. The modality-preference boundary tracks the object itself: a traffic cone is nearly six times as likely as a truck to be seen by the camera alone (16.8% vs. 2.9%), regardless of range. Camera-only detections peak at 10-30 m, inside the band usually ceded to LiDAR, and fall away on either side. The drivers of failure are not additive either: the distance - density interaction alone outweighs the entire main effect of viewing direction. We further find that modality dominance is directionally non-stationary, with objects behind the vehicle being markedly more LiDAR-dominant than those in front. We compile the mined structure into a Modality Reliability Map and inject it into four pretrained fusion detectors through a lightweight attention layer. This improves performance by up to 1.3 mean Average Precision (mAP) while fine-tuning under 4% of their parameters in 2-3 GPU-hours. Our work lifts reliability out of the weights and treats it as a quantity that can be examined before deployment. That is what a safety measure for multi-modal perception must rest on.

[148] Vision Models Capture Complementary Representations in Children That Converge in Adults cs.CVPDF

Domenic Bersch, Timothy Schaumlöffel, Michela Proietti, Antonia Franaszek-Traczewska, Hannah Elisabeth Zwad

TL;DR: 该 EEG 表征相似性研究比较儿童与成人阶段视觉—语言模型和纯视觉模型的脑表征对齐，发现两类模型在发展过程中反映互补特征。

Details

Motivation: 视觉语言对齐模型与纯视觉模型的表征如何对应于人类视觉发展过程，目前尚不清楚；研究考察不同年龄段 EEG 与模型表征的时间动态相似性。

Result: 作者对 8 岁、12 岁儿童和成人的 EEG 进行时间分辨表征相似性分析。CLIP 在儿童阶段已显示显著模型—脑对齐，纯视觉模型则在更晚发展阶段对齐更强。残差分析显示二者表征互补：8 岁时 CLIP 对脑活动的解释仍包含与结构匹配纯视觉模型不同的对齐信息。模型—脑对齐很大程度上反映语义结构，类别和有生命性（animacy）的 EEG 表征随年龄增强，而低层图像属性对齐始终较弱。摘要未给出样本人数、效应量或统计检验结果。

Insight: 不同模型类型可能捕捉发展中的视觉表征互补方面，可为随时间发展的 AI 表征设计提供线索。但 EEG—模型表征相关不等于模型模拟了人类发展机制；缺少样本规模与效应量，限制了对结论稳定性的判断。

Abstract: How the relationship between image-text aligned and vision-only models changes across development remains unknown. We applied time-resolved representational similarity analysis to EEG data from 8-year-olds, 12-year-olds, and adults. CLIP showed significant model-brain alignment already in children, whereas vision-only models aligned more strongly later in development. Residual analyses revealed complementary structure, with CLIP retaining alignment beyond an architecture-matched vision-only model from age 8. Much of the model-brain alignment reflected semantic structure, with EEG representations of category and animacy strengthening toward adulthood, while alignment with low-level image properties remained weak across age. Together, these findings suggest that image-text aligned and vision-only models capture complementary aspects of developing visual representations and may provide useful constraints for AI systems that aim to learn and adapt over time.

[149] EviStress: Evidence Stress Learning for Generalizable Face Forgery Detection with Textual Explanations cs.CVPDF

Jiuyao Jing, Xuanyu Yi, Yu Zheng, Chunlei Peng

TL;DR: EviStress 通过训练时扰动高度依赖的图像证据并保留真实性判别信息，提升人脸伪造检测对未知生成方式和图像损坏的泛化，并生成文字解释。

Details

Motivation: 网络传播会使伪造图像经历多样压缩和处理，生成管线也持续变化；标准 LoRA-CLIP 检测器对少量图像 patch 依赖过强，容易在这些证据受损或变化时失效。

Result: 两阶段框架先用 ETPP 从全局、关系和语义线索估计 patch 依赖，并提高高依赖区域被扰动的概率；DVP 约束干净与扰动表示的残差仍保留真实性判别信息。第二阶段将适配 CLIP 编码器移入 LLaVA 作为视觉塔，以参数高效指令微调直接输出真伪预测和文本解释，无需外接检测器。作者报告跨数据集、跨伪造方法泛化和图像损坏鲁棒性改善。摘要未列具体基准、数值或解释忠实性指标。

Insight: 压力测试模型依赖的视觉区域，同时约束扰动前后仍保留判别信息，目标是在证据变化时保持检测稳健；将检测器与解释能力集成也避免外部模块。摘要未量化泛化增益、计算成本或文字解释是否与图像证据逐项一致。

Abstract: The rapid spread of face forgeries on the Web challenges the credibility of visual content. In practical Web environments, manipulated media originate from evolving generation pipelines and undergo diverse processing during dissemination, making reliable authenticity assessment under unseen manipulations and visual changes a critical challenge. We observe that a standard LoRA-adapted CLIP detector exhibits uneven patch-level decision dependence, concen trating a disproportionate share of class-specific relevance on a small fraction of visual patches. To improve robustness to variations in such highly dependent evidence, we propose EviStress, a two-stage framework for generalizable face forgery detection with textual explanations. In the first stage, Evidence-Targeted Patch Perturbation (ETPP) estimates patch dependence from global, relational, and semantic cues and assigns positively correlated perturbation probabilities, explicitly stress-testing highly dependent visual evidence during training. We further introduce Discriminative Variation Preservation (DVP), which constrains the residual between clean and perturbed representations to retain authenticity discriminative information. In the second stage, Intrinsic Evidence Instruction Tuning (IEIT) transfers the adapted CLIP encoder into LLaVA as its vision tower and performs parameter-efficient multimodal instruction tuning, enabling authenticity prediction and textual explanation directly from the learned forgery-sensitive representation without an external detector. Experiments demonstrate improved cross-dataset and cross-method generalization and corruption robustness, with the multimodal model providing authenticity predictions and human-readable explanations.

[150] ConDenseDrive: Knowledge-Dense, Token-Sparse Visual Encoder for End-to-End Autonomous Driving cs.CV | cs.ROPDF

Shengzhen Li, Yongkang Li, Hao Gao, Chenxi Wu, Yajing Sun

TL;DR: ConDenseDrive 先将视觉、语义和几何知识蒸馏到驾驶视觉编码器，再用寄存器引导 token 压缩降低计算量，同时提升端到端规划表现。

Details

Motivation: 端到端驾驶需要丰富时空表示，但常规图像预训练视觉编码器知识有限，且逐相机逐帧密集处理所有 patch token 会增加推理延迟。

Result: 三阶段框架先从三个视觉基础模型蒸馏互补知识，再适配多帧端到端驾驶并把时空上下文写入用于规划的 register token，最后用 register 引导压缩冗余 patch token。作者报告压缩实现 2.0 倍加速且规划性能提高。NAVSIM v1 达到 94.5 PDMS，NAVSIM v2 达到 90.8 EPDMS；在 HUGSIM 零样本评测中 HD-Score 为 37.4。

Insight: 把寄存器作为聚合时空信息和选择视觉 token 的枢纽，兼顾知识丰富度与稀疏计算。报告的 2 倍加速和规划分数适用于所列评测环境；摘要未提供不同 token 压缩比例下的质量曲线、延迟测量硬件或完整消融。

Abstract: End-to-end driving requires rich spatiotemporal representations to support scene understanding and trajectory planning under strict latency constraints. However, most systems initialize their visual encoders from generic image pretraining and densely process all patch tokens for each camera view and input frame, limiting both representation quality and inference efficiency. To address these challenges, we present ConDenseDrive, a knowledge-dense and token-sparse visual encoder for end-to-end autonomous driving. Our three-stage framework first makes the encoder knowledge-dense by distilling complementary visual, semantic, and geometric knowledge from three vision foundation models. It then adapts the distilled encoder to multi-frame end-to-end driving and integrates spatiotemporal context into register tokens for planning. Finally, it makes the encoder token-sparse by applying register-guided token compression, reducing redundant patch computation and yielding a 2.0$\times$ speedup while improving planning performance. ConDenseDrive achieves 94.5 PDMS on NAVSIM v1 and 90.8 EPDMS on NAVSIM v2, and reaches 37.4 HD-Score in zero-shot evaluation on HUGSIM. Extensive experiments and qualitative analyses demonstrate that visual encoder design provides an effective route to reliable and efficient end-to-end driving.

[151] UniMetric: Universal Metric Scale Recovery for Visual Geometry Foundation Models cs.CVPDF

Lei Yang, Yuner Zhang, Ziyang Liu, Ziying Song, Kai WU

TL;DR: UniMetric 从熟悉物体的尺寸先验推导尺度区间，并在多帧重建中联合求解全局公制尺度，使视觉几何基础模型无需额外传感器或公制监督即可恢复真实尺度。

Details

Motivation: 视觉几何基础模型能恢复场景结构和相机运动，但尺度通常不具备可靠的公制意义；具身预训练和规划需要可用的物理尺度标注。

Result: 框架由视觉语言模型获取物体高度区间，沿估计支撑面法线测量重建物体高度并转为尺度区间，再用凸优化融合各帧约束；约束有交集时求共同可行尺度，否则求唯一折中，并将序列级尺度统一作用于点云和相机平移。作者提出含 12 个具身、驾驶、道路子集并具传感器真值的 MetricBench。在 9 个几何基础模型上，平均深度 Abs Rel、重建 Chamfer 距离和轨迹误差分别降低 54.2%、83.8%、65.6%，且三个任务套件层面的所有模型均有改善。

Insight: 多帧、多个物体的尺度先验可互相校验，降低单个物体尺寸假设不准的影响；严格相对高度的受控实验也支持组合观测的价值。尺寸先验来自 VLM，可能受物体识别和先验区间偏差影响；摘要称模型级任务套件均改善，不等于每个子集或样本都改善。

Abstract: Recent advances in multimodal foundation models, exemplified by GPT-6 Astra, raise the prospect of general-purpose embodied reasoning and planning. Using Internet videos for embodied pretraining calls for reliable physical-scale annotations. Yet visual geometry foundation models often recover scene structure and camera motion without reliable metric scale. We show that uncertain size priors of familiar objects can jointly constrain a reconstruction’s shared scale. Building on this insight, we present UniMetric, a universal and training-free metric scale recovery framework for geometry foundation models, without additional sensors or task-specific metric supervision. Specifically, Metric Prior Acquisition elicits object height intervals from a vision-language model. Semantic-Geometric Alignment measures reconstructed object heights along the estimated support-surface normal in each frame and converts the priors into scale intervals. Global Scale Recovery reconciles these intervals through a convex objective, yielding a jointly feasible scale when they overlap and a unique compromise otherwise. The estimated sequence-level scale is applied uniformly to reconstructed points and camera translations. Controlled experiments with real VLM priors show that combining observations improves scale accuracy with exact relative object heights. To evaluate metric accuracy on GFM reconstructions without ground-truth scale alignment, we introduce MetricBench, covering twelve embodied, driving, and roadside subsets with sensor-derived ground truth. Across nine geometry foundation models, UniMetric reduces mean depth Abs Rel, reconstruction Chamfer distance, and trajectory error by 54.2%, 83.8%, and 65.6%, respectively, improving every model on all three tasks at the suite level. We release the UniMetric code and MetricBench data.

[152] FOCUS: A Subspecialty Foundation Model for Intraoperative Central Nervous System Pathology cs.CVPDF

Xing Liu, Yifan Guo, Hongtai Lv, Lei Lou, Yuyi Tan

TL;DR: FOCUS 是以富含中枢神经系统组织的术中冰冻切片预训练的病理基础模型，在诊断与生物标志物任务上表现强，并提高低年资病理医师的分层诊断和淋巴瘤检出能力。

Details

Motivation: 中枢神经系统病变的术中冰冻切片诊断受制于伪影、病理复杂性和时间压力；多数病理基础模型对 CNS 冰冻组织覆盖有限。

Result: FOCUS 使用 27,930 张富集 CNS 组织的冰冻切片预训练。在 11 项内部诊断任务和 12 项生物标志物预测任务上的平均 AUC 分别为 0.969 和 0.877，并在时间外及外部队列中保持领先表现。作者将 10 项诊断任务组织成符合 WHO 的层级，并单独评估淋巴瘤检出。6 位神经病理医师参与读者评估；模型辅助使初级读者在层级诊断中接近或超过无辅助高级读者，并将初级读者淋巴瘤敏感度从 81.6% 提至 94.7%。

Insight: 专科冰冻切片预训练和分层诊断组织能支持术中决策所需的细粒度分类；读者研究还显示辅助效果可能对初级医师尤其明显。摘要未提供受试病例规模、特异度或决策结局，临床可用性仍需结合队列设计和前瞻性验证审慎解读。

Abstract: Intraoperative frozen-section diagnosis of central nervous system (CNS) lesions guides surgical decisions but is challenged by artefacts, diagnostic complexity and time constraints. Most pathology foundation models have limited exposure to CNS frozen tissue. Here we present FOCUS, a subspecialty foundation model pretrained on 27,930 frozen-sections enriched for CNS tissue. FOCUS led 11 internal diagnostic and 12 biomarker-prediction tasks, with mean AUCs of 0.969 and 0.877, respectively, retaining leading performance in temporal and external cohorts. To provide more comprehensive diagnostic information, we further organized ten diagnostic tasks into a WHO-aligned diagnostic hierarchy and separately evaluated FOCUS for lymphoma detection, a critical task for intraoperative surgical decision-making. In reader assessments by six neuropathologists, FOCUS assistance enabled junior readers to approach or exceed unassisted senior readers in hierarchical diagnosis and raised junior readers’ lymphoma sensitivity from 81.6% to 94.7%. FOCUS could improve diagnostic accuracy and granularity, inform lymphoma-related resection decisions and provide provisional molecular information. Such capabilities highlight the potential of FOCUS to provide robust, clinically actionable support for intraoperative CNS pathology.

[153] The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors cs.CV | cs.AIPDF

Shu Zhao, Tan Yu

TL;DR: SMILE 把视觉语言模型自身过去犯过的错误作为模型专属难负例，在测试时按当前预测方向检索相似错误，作为闭集分类和开放式 VQA 的视觉上下文示例。

Details

Motivation: 视觉 ICL 的表现依赖示例选择。已有反事实示例构建方法需经过属性抽取、描述工程和组合图像匹配等多阶段处理。

Result: SMILE 离线收集目标 VLM 自身预测错误，建立模型专属难负例池；推理时将当前预测作为方向信号，在语义嵌入空间检索最可能重复的错误示例，无需为分类或开放问答分别设计流程。在 4 个基准和 4 个 VLM 上，作者报告相较最强基线平均任务分提高 3.09 分，同时延迟降低至其五分之一。摘要未说明具体基准、分数范围或离线建库成本。

Insight: 模型自身的失败轨迹可直接揭示其特有混淆模式，并用作针对性上下文示例；模型专属错误库也意味着更换模型后可能需要重新积累。平均提升和 5 倍速度对比需结合基线延迟定义及错误库构建成本解读。

Abstract: Visual in-context learning (ICL) enables vision-language models (VLMs) to adapt at test time via demonstration examples, but its effectiveness depends critically on the examples selected. Recent counterfactual selection methods improve ICL by constructing demonstrations that expose how individual attribute changes affect the answer. However, they require multi-stage pipelines involving attribute extraction, caption engineering, and composed image matching. We propose a simpler alternative grounded in a key observation: a model’s own prediction errors are natural counterfactual demonstrations that directly encode its failure modes. Our method, SMILE (Self-Mined Visual In-Context Learning from Errors), builds a model-specific pool of hard negatives offline by recording the target VLM’s own mistakes. At test time, the prediction serves as a direction signal that surfaces, in the semantic embedding space, the past mistakes most likely to recur for the current query. This prediction-conditioned identification provides a unified mechanism for both closed-set classification and open-ended visual question answering without task-specific design. Across four benchmarks and four VLMs, SMILE delivers an average task-score gain of +3.09 points over the SOTA baseline at 5x lower latency. Project page: https://tree-shu-zhao.github.io/SMILE/

[154] Proactive Physical AI through Continuous Observation and Timely Intervention cs.CV | cs.AI | cs.ROPDF

Zhenhua Yang, Jiafeng Liang, Xingyu Zhou, Yi Jiang, Xiayu Cao

TL;DR: ProAgent 持续观察物理环境并判断何时应主动干预，再由响应模型生成建议或调用设备执行，以处理尚未由用户明确提出的新兴需求。

Details

Motivation: 许多物理 AI 假设用户先给定明确目标，但现实中重要任务可能先随环境状态变化而出现，用户尚未组织成指令；系统需主动察觉并在适当时机介入。

Result: ProAgent 由 StreamingJev 主动时机模型判断是否及何时介入，以及负责生成面向人的响应或调用设备能力的响应模型构成。作者建立自动化视频转监督流程以规模化训练主动时机。实验称在明确目标和持续角色两种设定下，相较源骨干基线能更准确把握介入时机并减少误触发；真实机器人评估也展示其识别新兴需求并执行适当物理动作的能力。摘要未给出具体指标、任务数或误触发率。

Insight: 主动型物理 AI 将介入时机视为核心预测对象，而不是等待显式指令后才规划动作。摘要的机器人结果为定性表述，缺少用户接受度、安全边界和误触发的数值评估，尚不能量化实际部署可靠性。

Abstract: Physical AI is commonly formulated around an explicitly specified task: given a goal or instruction, a system perceives, reasons, and acts. In many real-world settings, however, important tasks emerge from continuously changing world states before a person explicitly formulates them. We study this complementary problem as Proactive Physical AI, where an AI system continuously observes the physical world, identifies situations that warrant intervention, and initiates appropriate actions at the right time. We develop ProAgent, an end-to-end system built around two components: a Jev-style proactive timing model (dubbed StreamingJev) that determines whether and when intervention is needed, and a Response Model that generates responses for people or invokes available device capabilities for physical execution. To support learning at scale, we develop an automated data pipeline that converts videos into supervision for proactive timing. Experiments demonstrate strong proactive performance under both explicit objectives and standing roles, combining more accurate intervention timing with fewer false triggers than the source-backbone baseline. Real-robot evaluations further demonstrate ProAgent’s ability to identify emerging needs and translate them into appropriate physical actions.

[155] AUGUstus: Autonomous Driving VLA Model with Unified Generation and Understanding cs.CV | cs.AIPDF

Yuqi Ye, Caigao Jiang, Zhuonan Wang, Congyuan Pei, Chao Qu

TL;DR: AUGUstus 将扩散式动作生成与自回归 CoT 理解统一训练，推理时先生成动作并可跳过 CoT，从而在保留理解监督的同时降低驾驶延迟。

Details

Motivation: 自动驾驶 VLA 面临两种局限：直接回归或 flow 动作模型延迟低，却可能弱化视觉语言理解并产生不安全、难解释动作；自回归 CoT 能利用理解能力，但逐 token 生成延迟较高。

Result: AUGUstus 用联合 CoT 监督在共享表示中保留空间理解，以辅助更准确安全的轨迹生成；AUGUstus-mode 训练时先生成动作、后生成 CoT，推理时只需动作输出；AUGUstus-GRPO 联合优化动作和 CoT。NAVSIM v1 PDMS 为 94.6，NAVSIM v2 EPDMS 为 90.7，作者称无需 scorer module 即达到最佳规划结果。nuScenes 平均 L2 误差 0.27 米、平均碰撞率 0.05%，为所比较方法中最好。模型用约 512 张 H800 训练两天，在内部 1,000 公里数据上评测，并初步部署到车辆，以 12 Hz 运行。

Insight: 把 CoT 留作训练时辅助监督、将推理路径设为动作优先，有机会兼顾语义理解与实时控制。内部道路评测和初步实车部署增强了实际性，但摘要未说明部署场景范围、基线配置及安全接管机制；12 Hz 是报告的运行频率，不等于完整系统时延保证。

Abstract: Vision-language-action (VLA) models show strong potential for end-to-end autonomous driving by leveraging world knowledge and understanding to excel in long-tail scenarios. However, current VLAs have two limitations: 1) Degeneration into vision-action (VA) models. For low latency, regression/flow-based methods directly generate actions but underuse VLM understanding, causing unsafe, unexplainable actions. 2) Autoregressive high latency. Autoregressive methods sequentially generate chain-of-thought (CoT) and action tokens, fully leveraging VLM understanding, but high latency hinders deployment. To enable low-latency deployment while using CoT to assist action generation, we propose AUGUstus, an Autonomous-driving VLA model with Unified Generation and Understanding. We unify diffusion-based action generation and autoregressive CoT understanding in one model, where understanding only serves generation. Joint CoT supervision preserves VLM spatial understanding in shared representation for safer, more accurate trajectories. AUGUstus-mode orders action generation before CoT, so CoT supervises reliable actions during training while inference needs only action output, enabling low latency. AUGUstus-GRPO, a unified GRPO algorithm, jointly optimizes action and CoT, improving action performance. Experiments show AUGUstus achieves 94.6 PDMS on NAVSIMv1 and 90.7 EPDMS on NAVSIMv2, SOTA planning without scorer module. On nuScenes, it achieves average L2 error 0.27 m and average collision rate 0.05%, best among compared methods. To validate AUGUstus practicality, we train on in-house data with 512 NVIDIA H800 GPUs for ~2 days, evaluate on an internal 1,000 km dataset, and preliminarily deploy on a real vehicle, achieving 12 Hz operation. Project page: https://augustus-ad.github.io

[156] Self-Supervised Learning of Semantically Consistent Visual Features via a Simple 3D Prior cs.CVPDF

Sebastian Day, Oisin Mac Aodha

TL;DR: 该自监督方法通过数据集级对比学习和分布匹配学习类别粗形状与外观，再结合几何损失改善视觉特征的语义对应，特别针对对称和重复部件的左右混淆。

Details

Motivation: DINOv2、Stable Diffusion 等特征虽具语义对应能力，但在对称或含重复部件物体上可能无法区分左右。已有改进依赖标注关键点、相机位姿，或假设物体具有球形等受限拓扑。

Result: 方法不使用昂贵的实例级分析—合成流程，而是在数据集层面学习类别形状与外观分布，并用额外几何损失，使学习特征在全局和局部统计上匹配真实图像提取的特征。作者在多个有挑战性的物体类别上报告定量结果并称方法有效。摘要未列出类别名称、基准、数值或与基线的差异。

Insight: 全局类别分布先验可在不逐实例标注的情况下补充几何与外观结构，避免依赖固定简单拓扑。由于摘要未提供评测结果细节，尚不能判断对称消歧改善幅度、类别迁移能力或其与监督方法的差距。

Abstract: Visual features extracted from vision models like DINOv2 or Stable Diffusion enable high quality semantic correspondence between images. However, such features exhibit documented limitations that hinder accurate correspondence estimation such as the inability to disambiguate the left from right for categories exhibiting symmetries or repeated parts. Previous attempts to improve these features rely on using supervised data such as annotated keypoint pairs or camera pose information, or make restrictive topological assumptions, such as modelling the object of interest as a 3D sphere. We introduce a new self-supervised approach that allows us to learn features for topologically complex object categories using a simple prior. Instead of relying on an expensive instance-level analysis by synthesis pipeline, our approach makes use of contrastive learning and distribution matching at the global dataset-level to learn the coarse shape and appearance of a category. Combined with additional geometric losses, this enables us to replicate the global and local statistics of visual features which match those that are extracted from real images. Quantitative results are presented across several challenging object categories, demonstrating the effectiveness of our approach.

[157] MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation cs.CVPDF

Yawen Shao, Jie Xiao, Kai Zhu, Yu Liu, Hongchen Luo

TL;DR: MaskSense 针对掩码图像生成强化学习中的“视觉探索陷阱”，在采样和优化阶段保留主体高熵探索并将梯度集中到关键解码步骤。

Details

Motivation: 直接将 GRPO 用于掩码生成模型会遇到探索冲突：基于置信度的解码倾向先填充低不确定背景、过早耗尽多样主体候选，使依赖 rollout 多样性的强化学习缺乏有效区分信号。

Result: 语义锚定路由采样利用语义先验保留主体 token 的高熵探索，并通过探索/利用路由扩大组内奖励差异；全局熵转移锚定策略识别影响最大的解码步骤，将梯度集中于这些位置。在多个文生图基准上，作者报告 GenEval 准确率从 54% 升至 86%，HPS 从 28.89 升至 36.39，并达到当时最佳表现。

Insight: 改善组内候选多样性有助于产生更有辨别力的相对奖励，解码阶段级的梯度分配则避免对所有步骤均匀更新。摘要未列出基准范围、生成质量其他指标或推理开销；高分提升需结合具体基础模型和评测协议理解。

Abstract: Reinforcement learning (RL) has shown strong potential for aligning generative models with human intent, but adapting RL to masked generative models (MGMs) remains largely underexplored. Existing approaches that directly adapt Group Relative Policy Optimization (GRPO) to MGMs leave a foundational conflict unaddressed: MGMs’ exploitative decoding nature is inherently at odds with the exploratory diversity that GRPO presupposes. In this work, we reveal a structural decoding bias termed the visual exploration trap: confidence-based sampling trades the exploration of diverse subject realizations for the greedy resolution of low-uncertainty background regions, causing premature collapse of the generative exploration space and starving GRPO of the rollout diversity it depends on. To this end, we propose MaskSense, a novel RL framework for MGMs that confronts the visual exploration trap at both the sampling and optimization stages. Specifically, we introduce a semantic anchored routing sampling mechanism that leverages semantic priors to preserve high-entropy subject token exploration while amplifying intra-group reward variance through exploratory-exploitative routing to yield more discriminative advantage estimates. Furthermore, we design a global entropy transition anchoring strategy to identify the most consequential decoding steps and concentrate gradient updates on them. Extensive experiments on multiple text-to-image benchmarks demonstrate that MaskSense substantially improves upon the base model, with GenEval accuracy improving from 54% to 86% and HPS from 28.89 to 36.39, achieving state-of-the-art performance.

[158] DIME: Scaling Facial Representation Learning via Differential Masked Autoencoding cs.CVPDF

Hao Yu, Haoyu Chen, Hui Wei, Yan Jiang, Nicu Sebe

TL;DR: DIME 用同一身份两张脸图像的差分方向监督掩码自编码器，减弱身份、姿态和光照主导的训练信号，学习细粒度面部表示。

Details

Motivation: 常规视觉预训练目标以重建单张图像外观为主，身份、姿态和光照可能掩盖同一身份内部细微表情或面部变化，使扩充数据和模型容量的收益在精细面部分析中趋于饱和。

Result: DIME 在身份匹配图像对上同时重建两张被遮挡图像，并以显式差分方向监督（EDDS）使重建差分方向匹配真实差分，同时确保每张重建只依赖自身图像。方向损失不受差分幅度支配，使细微变化不被大姿态/光照变化淹没。身份标签只用于构造预训练配对，无需稠密标注；作者在 2,000 万张高分辨率人脸图像上预训练。随数据从 600 万扩至 2,000 万，BP4D AU 检测 F1 提升 5.9 分，而 MAE 在相同子集仅提高 0.5 分。DIME 还在 AU 检测、强度估计、跨域反欺骗、关键点、头姿、解析和人脸生成上超过所评专用方法。

Insight: 配对差分让训练聚焦同一身份内部变化，而不是重复学习身份外观；损失关注方向而非幅度也有助于保留微弱表情信号。跨任务领先和数据规模收益限于作者报告的数据与协议；摘要未提供所有任务的具体比较数值。

Abstract: Scaling up data and model capacity has driven substantial progress in visual representation learning. However, conventional pretext tasks exhibit diminishing returns on fine-grained facial analysis. We attribute this to a scaling bottleneck: these tasks model each image’s appearance and are dominated by identity, pose, and illumination, so subtle within-identity facial differences contribute little to the training signal. We introduce DIME, a differential masked autoencoder for large-scale facial representation learning. DIME builds on a simple observation: two images of the same person share identity-specific appearance, so their difference exposes facial changes. DIME reconstructs both masked images in an identity-matched pair and, through an Explicit Differential Direction Supervision (EDDS) objective, aligns the direction of their reconstructed difference with the true difference, while source isolation makes each reconstruction depend only on its own image. Shared appearance cancels in this difference, and because EDDS ignores magnitude, subtle facial changes count as much as large pose or illumination changes. DIME uses identity labels only to form pretraining pairs, requires no dense annotations, and learns dense facial representations with emergent part-level semantics. We pretrain DIME on 20 million high-resolution facial images consolidated from existing face datasets. DIME’s AU detection performance continues to improve as pretraining data scales from 6M to 20M images, with its BP4D F1 increasing by 5.9 points, while MAE gains only 0.5 points on the same subsets. DIME also outperforms specialized facial methods on AU detection, AU intensity estimation, cross-domain anti-spoofing, landmark localization, head pose estimation, face parsing, and face generation.

[159] PerturbBridge: Direct Phenotypic Supervision for Cellular Perturbation Generation cs.CVPDF

Yucheng Yang, Xiaoxiao Sun, Huangqingbo Sun, Elaine Sui, Emma Lundberg

TL;DR: PerturbBridge 在图像生成训练过程中直接对扰动特异的细胞表型施加生物学损失监督，以提升虚拟细胞图像的表型忠实度与采样效率。

Details

Motivation: 仅拟合真实图像分布并不会明确优先生成特定扰动引起的生物响应；已有强化学习后训练或推理时 steering 方案也引出一个问题：生物学响应监督应放在生成流程的哪个阶段。

Result: PerturbBridge 是随机 data-to-data bridge 模型，联合优化桥接回归损失与可微生物学响应损失，生物学梯度经冻结图像编码器传回。研究用下游分类和检索评估扰动响应。在三个显微镜基准上，作者报告其生物忠实度、图像质量和采样效率均超过 CellFlux。BBBC021 的作用机制分类使用形态特征、学习视觉特征时分别提高 5.66、7.5 个百分点；JUMP 与 AllenCell 的检索 Hit@10、Hit@1 分别提高 1.77、7.15 个百分点。BBBC021 和 JUMP 上，FID 更低且模型评估次数减少 80%。

Insight: 将生物学响应目标嵌入训练而非只在生成后筛选，有望同时改善表型匹配与生成效率。指标涵盖分类、检索和 FID，但摘要未给出各自基线绝对值；降低模型评估次数也不直接等于相同比例墙钟时间降低。

Abstract: Virtual cell models should capture phenotypes induced by specific perturbations. However, image-based models are typically trained and evaluated to match real-image distributions, which alone does not explicitly prioritize perturbation-specific biological responses. Recent approaches add biological signals through reinforcement-learning post-training or inference-time steering, raising a central question: where should biological response information enter the generative pipeline? We introduce PerturbBridge, a stochastic data-to-data bridge model that directly supervises perturbation-specific phenotypes in predicted perturbed images during training. PerturbBridge jointly optimizes bridge regression and differentiable biological response losses, with biological gradients passing through a frozen image encoder. We further introduce perturbation-response evaluations based on downstream classification and retrieval. Across three microscopy benchmarks for perturbation profiling and subcellular imaging, PerturbBridge surpasses CellFlux in biological fidelity, image quality, and sampling efficiency. On BBBC021, mechanism-of-action classification improves by 5.66 and 7.5 percentage points with morphological and learned visual features, respectively. On JUMP and AllenCell, retrieval Hit@10 and Hit@1 improve by 1.77 and 7.15 percentage points, respectively. On BBBC021 and JUMP, PerturbBridge achieves lower FID with 80% fewer model evaluations than CellFlux. Overall, these results show that directly supervising perturbation responses during training improves phenotypic fidelity while enabling efficient cellular image generation. Code is available at https://github.com/yyc1215lll-creator/PerturbBridge.

[160] Rethinking LLM-Guided Visual Question Answering: Does Deeper Knowledge Integration Really Help? cs.CVPDF

Aneerban Saha, Vijay Kumar, Sameer Madhav D., Priyam Srivastava, Dhruv Bansal

TL;DR: 该研究在统一冻结 ViT+FLAN-T5 的 OK-VQA 框架中比较 LLM 知识引导的晚期决策融合与注意力矩阵注入，发现数据稀缺时浅层融合效果更好。

Details

Motivation: 知识型 VQA 需要结合图像外部常识，但 LLM 知识应在视觉语言架构的何处、以多深方式整合仍不清楚；更深的表示干预未必带来更好效果。

Result: 作者用 Llama 3.3 70B 生成结构化推理指导、视觉焦点概念、干扰项及外部知识，在冻结 ViT+FLAN-T5 的统一框架和 OK-VQA 上比较晚期融合与注意力矩阵注入。与“整合越深越好”的常见假设相反，数据稀缺微调下浅层晚期融合最好，精确匹配率为 39.4%，比无指导基线高 10.0 个百分点；深层注意力修改为 31.8%。另提出问题释义增强流程，可将训练数据扩充至十倍以上。

Insight: 知识整合深度应与训练数据规模相匹配；强行修改深层注意力可能在数据稀缺时损害效果。研究基于冻结 ViT+FLAN-T5、OK-VQA 和特定 LLM 生成知识，尚不能据此推断大规模数据或端到端训练下的最佳整合策略。

Abstract: Knowledge-based Visual Question Answering (VQA) requires reasoning that extends beyond visible image content, demanding external commonsense and factual knowledge. While recent work leverages Large Language Models (LLMs) as knowledge sources, the question of how deeply such guidance should be integrated into a vision language architecture remains underexplored. In this work, we generate structured reasoning guidance, visual focus concepts, distractor elements and external knowledge using Llama~~3.3~~70B and systematically compare two integration strategies within a unified frozen ViT + FLAN-T5 framework on the OK-VQA benchmark: decision level late fusion and attention-matrix injection. Contrary to the prevailing assumption that deeper integration is superior, we find that shallow late fusion achieves the strongest results (39.4% exact match, +10.0 over the no guidance baseline), outperforming deep attention modification (31.8%) under data scarce fine-tuning. We further contribute an LLM based augmentation pipeline that expands training data more than tenfold via question paraphrasing. Our findings indicate that integration depth must be matched to the data regime.

[161] A unified three-branch deep learning framework for automated detection and counting of wildlife in aerial imagery cs.CVPDF

Zeyu Xu, Tiejun Wang, Andrew Skidmore, Megan Claase, Harry Wells

TL;DR: 该三分支系统将航拍野生动物监测中的样带提取、重复覆盖匹配和动物检测计数整合起来，并在南苏丹四种羚羊的斜视航拍图像上评测。

Details

Motivation: 大规模空中野生动物调查需要可靠地界定观测带、消除连续影像的重叠覆盖，并检测计数动物；单独检测无法保证计数区域和重复样本处理正确。

Result: WildSeg 用杆状标记分割和几何细化提取计数样带；WildMatch 通过教师—学生蒸馏学习稠密图像对应，不依赖地理定位元数据来找重叠区域；WildDet 结合双尺度检测、特征融合、密度监督和空图门控。数据为南苏丹 Great Nile Migration Landscape 中四种羚羊的倾斜航拍图像。WildSeg 能可靠提取标杆，WildMatch 接近教师精度且模型更轻更快；两者在缩放图像上仍保持较高精度。相较 YOLO11，WildDet 对四个物种均提升检测表现，并使基于检测的计数 MAE、RMSE、平均偏差分别降低 21.0%、4.6%、27.8%。

Insight: 将调查空间约束、重复覆盖匹配与动物检测打通，可减少仅靠逐图检测造成的区域偏差和重复计数。结论针对单一地区的四种羚羊航拍数据，摘要未报告各物种绝对误差或跨地区迁移效果。

Abstract: Automating aerial wildlife surveys requires reliable animal detection, consistent delineation of observation strips, and correction for repeated coverage between successive images. We present a unified three-branch deep learning framework that integrates these tasks through WildSeg, WildMatch and WildDet. WildSeg combines rod segmentation with geometric refinement to delineate counting strips from aircraft-mounted markers. WildMatch uses teacher-student distillation to learn dense image correspondences and identify overlapping survey areas without relying on geolocation metadata. WildDet combines dual-scale detection with feature fusion, density supervision and empty-image gating to support animal detection and counting. We evaluated the framework using oblique aerial imagery of four antelope species from the Great Nile Migration Landscape in South Sudan. WildSeg achieved reliable rod extraction, while WildMatch approached teacher-level accuracy with a substantially lighter and faster model. Both modules maintained high accuracy on resized images, improving processing efficiency. WildDet improved detection performance across all four species and reduced detection-based count MAE, RMSE and mean bias by 21.0%, 4.6% and 27.8%, respectively, relative to YOLO11. By connecting spatial survey constraints with animal detection, the framework provides a practical and computationally efficient approach to processing large aerial image collections for wildlife monitoring and population surveys.

[162] Scaling Unified Multimodal Models to Long-Horizon Interleaved Generation cs.CVPDF

Ziqiao Peng, Yutao Cui, Wang Lin, Zhong-Yu Li, Qingwen Gu

TL;DR: 该工作训练原生交错式多模态模型进行长程文本—图像生成，并用视觉记忆管理历史参考、选择相关场景状态和适应不完美的自生成历史。

Details

Motivation: 统一多模态模型可让生成的图像直接影响后续推理和合成，但已有系统多展示短轨迹或简化内容。长自回归过程中图像不断累积，错误或无关画面会干扰后续生成，而角色和场景又需在较长间隔后恢复。

Result: 作者训练了具备图像合成、文本渲染、复杂构图和多模态规划能力的统一模型，并研究如何扩展到更长自回归轨迹。视觉记忆框架保留持久参考，按场景选择性检索历史视觉状态，同时保留完整语义历史；少步自生成 rollout 用于适应不完美的模型生成历史。作者还提出 LongMIG-Bench，评估长程文本到多图及图像到多图生成。摘要称模型实现高保真长程生成并保持较强角色和场景一致性，但未给出基准规模、数值成绩或对比增益。

Insight: 长程生成需要把语义历史保留与视觉参考筛选结合，既避免上下文被无关画面淹没，也防止角色和场景遗忘。摘要没有提供与无记忆模型的量化对照，因此对记忆组件的独立贡献仍无法判断。

Abstract: Unified multimodal models (UMMs) provide an interface for interleaved text-image generation, where generated visual states can directly inform subsequent reasoning and synthesis. However, existing systems are still primarily demonstrated on short trajectories or simplified visual content. We develop a large UMM natively trained on interleaved multimodal data, with strong capabilities in image synthesis, text rendering, complex composition, and multimodal planning, and study how to extend these native capabilities to substantially longer autoregressive trajectories. The main challenge is that generated images continuously accumulate in context: irrelevant or imperfect visual states interfere with future synthesis, while previously established characters and scenes must still be recovered after long intervals. We introduce a visual-memory framework that maintains persistent visual references, selectively retrieves scene-relevant historical states while preserving the complete semantic history, and uses few-step self-rollout to adapt the model to imperfect generated histories. We further introduce LongMIG-Bench for evaluating long-horizon text-to-multi-image and image-to-multi-image generation. Our model achieves high-fidelity long-horizon interleaved generation with strong character and scene consistency, demonstrating how visual memory can scale native interleaved generation to substantially longer trajectories.

[163] Contrastive learning for video reuse detection cs.CVPDF

Tristan J. B. Cann, Arjun Biswas

TL;DR: 该视频复用检测方法以对比学习微调 Siamese 网络，从目标视频关键帧中检索潜在素材来源，并通过关键视觉缩减提高大规模数据集筛查效率。

Details

Motivation: 社交媒体混剪和专业素材库再利用使识别视频内容来源具有版权应用价值；完整视频数据复杂、规模大，不利于逐帧比对。

Result: 方法从目标视频提取关键帧，并训练孪生网络识别其可能来源。作者称关键帧集合通常不到完整视频数据量的 10%，仍可有效定位组成素材。在由素材库视频编成的测试集合上，组成视频识别平均准确率超过 0.95。摘要未给出数据集规模、切分方式、视频变换类型或误检率。

Insight: 将视频压缩为少量关键视觉再做对比检索，有望降低大型素材库的搜索负担。平均准确率超过 0.95 是在作者素材库测试集合上的结果，摘要未说明对裁剪、变速、叠字等复杂变换的稳健性及候选检索规模。

Abstract: The incorporation of footage from other sources has long been a concern from a copyright perspective, as such usage can harm the interests of rights holders. Recent trends in the ``remixing’’ of social media content and the development of bespoke stock libraries for specialist causes mean that methods to identify content reuse will have a range of applications. In this paper, we develop a framework based on a Siamese network fine tuned using contrastive learning to identify potential sources for key frames in a given video. This framework makes two contributions to analysis of large video datasets. Firstly, we demonstrate its utility at reducing complex video data to a set of key visuals typically less than 10% of the size of the full set. Secondly, we show that these reduced sets allow for efficient and effective identification of component footage. We test this method over a set of videos compiled from a stock footage library and demonstrate its effectiveness for constituent video identification with average accuracy greater than 0.95.

[164] EgoProbe: Benchmarking Long-Range and Exact Memory in Streaming Egocentric Videos cs.CVPDF

Haokai Pang, Yuqian Fu, Yifei Wang, Yujiao Shen, Tianwen Qian

TL;DR: EgoProbe 将流式第一人称视频记忆拆成远距离保留和受相似事件干扰时的目标事件细节回忆，并通过两轮由粗到细提问检测“记得事件但记错细节”的失败。

Details

Motivation: 可穿戴助手需要边接收视频边处理远过去的经历；记忆跨度长只是挑战之一，反复出现的物体、动作和地点还要求系统准确绑定具体细节与对应事件。

Result: EgoProbe 基于长时第一人称视频，通过事件级数据引擎延长回忆间隔、找出相关事件并构造事件干扰项，再由人工核验。EgoProbe-Base 测无相关事件插入时的长程回忆；EgoProbe-Echo 插入相似事件和目标事件干扰项，并用多轮粗到细提问：首轮测粗粒度事件回忆，第二轮测试指定事件的目标细节。多模型评测显示长程回忆明显受限，相关经历干扰时更难取回目标事件细节；即使粗粒度回答正确，模型也常忘记指定事件的具体信息。摘要未报告各模型准确率。

Insight: 将回忆距离与事件绑定精确度分开评估，可暴露仅问“发生过什么”时被掩盖的细节混淆。事件级干扰设计贴近重复经历场景，但摘要未说明视频时长、标注规模、流式预算或不同记忆方法的量化差异。

Abstract: Egocentric video provides a continuous first-person record of everyday activities, making it a natural foundation for wearable AI assistants. Unlike offline video understanding, streaming systems must process observations as they arrive, while queries may refer to experiences from the distant past. Recent work has begun to explore whether information can be recalled across long temporal gaps. Yet remembering how far is only part of the challenge: recurring objects, actions, and locations also require models to remember exactly which detail belongs to which event. We introduce EgoProbe, a benchmark that evaluates streaming egocentric memory along these two complementary dimensions: long-range retention, i.e., how far, and event-specific recall under interference, i.e., how exact. Built on long-form egocentric videos, EgoProbe uses an event-grounded data engine to extend recall intervals, identify related events, and construct event-grounded distractors, followed by human verification. It contains two subsets: EgoProbe-Base, which evaluates long-range recall without intervening related events, and EgoProbe-Echo, which challenges models with similar intervening events and event-grounded distractors. EgoProbe-Echo adopts coarse-to-fine multi-turn probing: the first turn evaluates coarse-grained event recall, while the second tests whether the model can retrieve the target-specific detail associated with the designated event. Extensive evaluation across diverse models reveals substantial limitations in long-range recall, with further challenges in recovering target-event details among related experiences. Even when models correctly answer coarse-grained questions, they often fail to recall the specific details of the designated event. EgoProbe exposes this underexplored failure mode and provides a testbed for advancing long-range and exact memory in streaming egocentric video understanding.

[165] When Left Is Right? Camera-Motion Understanding beyond Screen cs.CVPDF

Chong Cheng, Ran Zhang, Xinyuan Li, Alan Zhao, Hao Wang

TL;DR: CameraVLM 将屏幕外相机运动状态与画面内前景/背景光流联合表示，并让语言模型参与运动语义对齐，以提升左右相机运动判断。

Details

Motivation: 尽管 VLM 能处理视觉谜题和长视频，但多数模型判断相机向左还是向右仍接近随机水平；单独增加相机运动数据或几何表示均未解决问题。

Result: 方法同时使用表示屏幕外运动的 camera tokens 和表示画面内主体/背景运动的 flow tokens，并以两阶段训练将它们与相机运动语言语义共同学习。仅 9B 参数的 CameraVLM 在三个基准、20 多个基线中优于所有开源及专用相机运动模型；在 48 人研究中达到或超过 GPT-6 Astra 及人类观察者。摘要未提供具体基准分数或人类研究统计。

Insight: 相机运动理解需要把运动几何与语言语义绑定，单独训练姿态估计器或仅扩充数据可能不足。人类对照来自 48 名参与者，结论限于论文设置的三个基准与评估任务。

Abstract: Can a vision-language model tell left from right? Modern VLMs solve visual puzzles and follow hour-long videos, yet when asked whether the camera moved left or right, most do no better than a coin flip. Adding camera-motion data or geometric representations has been tried, separately, and neither closes the gap. We argue that camera-motion understanding needs two conditions at once: the representation must couple off-screen camera motion with on-screen flow dynamics, and it must be interpreted by the language model rather than trained as a standalone pose or depth estimator. We propose CameraVLM, which pairs a motion-state representation with state-language alignment. Camera tokens capture how the camera moves off screen, flow tokens capture how the subject and background move on screen, and a two-stage training strategy keeps language in the loop so that both are learned together with camera-motion semantics. With only 9B parameters, CameraVLM outperforms all open-source and camera-motion-specific models by a wide margin across three benchmarks and more than 20 baselines, and matches or surpasses GPT-6 Astra as well as human viewers in a 48-participant study. Project page: https://3dagentworld.github.io/CameraVLM/

[166] Beyond Sparse Captions: Aligning Slide-Level Text and Patch-Level Vision in Pathology cs.CVPDF

Daniel Shao, Dongmin Bang, Luca L. Weishaupt, Sophia J. Wagner, Nic G. Reitsam

TL;DR: LIMIT 冻结视觉编码器、只微调文本编码器，使病理报告文本以交叉注意力聚合切片 patch 特征，连接宏观诊断描述与微观组织图像。

Details

Motivation: 计算病理视觉语言预训练依赖精选图像—描述数据，这类数据相对于组织形态多样性仍稀疏，且临床病理医师不会常规撰写逐 patch 描述；切片报告的宏观诊断信息与 patch 的微观形态存在尺度鸿沟。

Result: LIMIT 以报告嵌入作为交叉注意力查询，在 patch 特征袋中生成报告条件化切片表示，并通过 InfoNCE 优化。基于该方法，作者构建 CONCH-Z，将 CONCH v1.5 的文本编码器与 33.5 万份临床报告对齐。在 22 项 patch 分类、肿瘤检测和切片分型任务上，CONCH-Z 的零样本表现优于基于精选图注训练的视觉语言编码器，同时缩小与参数量和计算量显著更大的切片级基础模型之间的差距。

Insight: 利用常规临床报告扩展监督来源，并让报告文本跨 patch 聚合，可以缓解稀疏图注与宏观—微观尺度不匹配。零样本优势针对 22 项任务与特定比较组；摘要未给出各任务数值或报告文本噪声对性能的影响。

Abstract: Contrastive vision-language (VL) pretraining in computational pathology has generally relied on curated image-caption datasets sourced from social media, educational videos, and research publications. These corpora are sparse relative to the diversity of tissue morphology and difficult to scale further, as pathologists do not routinely produce descriptive captions in clinical practice. We propose to instead leverage slide-level pathology reports which can be generated during routine diagnostic workflows in a scalable manner. A key challenge in this weakly-supervised setting is that pathology reports describe macroscopic, diagnostic findings while patch-level visual features describe microscopic findings. We introduce Locked Image Multiple Instance PreTraining (LIMIT), which addresses this gap by freezing a pretrained vision encoder and fine-tuning only the text encoder: report embeddings serve as cross-attention queries over bags of patch features, producing report-conditioned slide representations optimized via InfoNCE. Using LIMIT, we establish CONCH-Z, which aligns the text encoder of CONCH v1.5 to 335k clinical reports. Evaluated across 22 tasks spanning patch-level classification, tumor detection, and slide-level subtyping, CONCH-Z establishes state-of-the-art zero-shot performance over VL encoders trained on curated caption datasets, while simultaneously closing the gap with slide-level foundation models trained with substantially more parameters and compute. The code and model is available at https://github.com/mahmoodlab/LIMIT.

[167] Multics: Generative Multi-Player World Model with Hierarchical Communication cs.CVPDF

Kaiyi Huang, Wangbo Yu, Yukun Huang, Jiwen Yu, Kunhao Liu

TL;DR: Multics 用玩家—群组—全局三级通信和基于摄像机轨迹的动态分组，构建可扩展的多玩家视频世界模型，减少完全互联通信的注意力开销及集中式表征干扰。

Details

Motivation: 多玩家世界模型需在共享环境中保持多视角一致生成，但完全互联通信会随玩家数增加产生二次注意力开销，集中式全局表征则可能使无关事件互相干扰。

Result: Multics 在自注意力中加入群组 token，先聚合局部相关玩家信息，再通过全局 token 进行跨群组通信；分层拓扑限制无关局部事件的直接混合。群组 token 参数在不同组槽位间共享，支持可变玩家数而不新增专属参数；动态分组根据近期相机轨迹，在通信预算内选择相关玩家。作者在 4 人和 6 人设定报告最先进性能，并改善跨玩家一致性、玩家动作及相机可控性和抗干扰能力。模型从 4 人训练后无需再训练即可泛化到 6、8、10 人推理，且比完全互联通信减少注意力计算。

Insight: 通过层级中间表示把局部交互与跨组共享分开，既可控信息传播，也可降低玩家数量增长时的通信成本。摘要未给出速度、显存或一致性提升数值；4 人训练泛化到 10 人是报告的设定，实际泛化质量仍需具体指标验证。

Abstract: Multi-player interactive world models aim to simulate shared environments from multiple players’ perspectives, but achieving robust generation while scaling to more players remains challenging. Existing approaches that rely on fully-connected communication (e.g., Solaris) incur quadratic attention costs as the number of players grows, while centralized approaches (e.g., $γ$-World) suffer from interference between unrelated events in a shared global representation. To address these challenges, we present Multics, a multi-player video generation framework with hierarchical communication that structures information exchange across player, group, and global levels. This hierarchy introduces group tokens into self-attention as intermediate group-level representations that aggregate locally relevant player information. These tokens connect to global tokens for cross-group communication, while the hierarchical topology limits direct mixing between unrelated local events. Group-token parameters are shared across group slots, allowing the same communication module to accommodate varying numbers of players without introducing new group-specific parameters. We further propose budget-constrained dynamic grouping based on recent camera trajectories, which adaptively groups relevant players while limiting unnecessary communication. Experiments demonstrate that Multics achieves state-of-the-art performance in 4- and 6-player settings, with improved cross-player consistency, player-specific action and camera controllability, and interference robustness. Moreover, Multics generalizes from 4-player training to 6-, 8-, and 10-player inference without retraining, while reducing self-attention computation compared with fully-connected communication.

Yating Wang, Kui Wu, Ran Yi, Zixuan Lu, Tongtong Wang

TL;DR: HairFlow 以粗略发型网格约束整体形状、以多视图草图指定局部发丝流向，通过条件扩散生成可渲染和模拟的显式 3D 发丝。

Details

Motivation: 发丝几何复杂、高质量训练数据稀缺，输入观测还可能不完整或有噪声；基础模型产生的发型先验不能直接用于发丝级渲染与物理模拟，既有方案需昂贵的逐实例优化。

Result: HairFlow 结合 Orientation VAE 和条件扩散 Transformer 生成体场，再将其追踪为显式发丝。为解决缺少配对草图—发丝训练数据的问题，作者自动合成几何对齐、多视角且风格多样的草图。实验报告在多种发型上生成的 3D 发丝合理并符合输入条件，显著缩短推理时间，且支持灵活多模态控制。摘要未列出数据规模、耗时或定量质量指标。

Insight: 网格提供全局形状、草图提供局部流向，条件分工使用户控制兼顾粗细尺度；自动合成训练配对则缓解真实监督稀缺。摘要只有定性结果，尚无法比较推理效率和重建质量，也未说明输入噪声下的鲁棒性。

Abstract: Modeling 3D hair strands is challenging due to complex hairstyle geometry, scarce high-quality training data, and incomplete or noisy input observations. Foundation models offer broader hairstyle priors, but their outputs are not directly usable for strand-based rendering and simulation. Existing approaches require costly per-instance optimization to fit hair priors to meshes from foundation models and 2D orientations from specialized predictors trained on limited hair data, restricting efficiency and generalization. We present HairFlow, a multimodal conditioned diffusion framework for efficient and controllable hair strands modeling. HairFlow conditions strand generation on two complementary representations: a coarse mesh that constrains global shape and multi-view sketches that specify local strand flow. HairFlow combines an Orientation VAE with a conditioned diffusion transformer to generate volumetric fields that are traced into explicit strands. To overcome the lack of paired sketch-strand training data, we introduce an automatic synthesis pipeline that generates geometrically aligned multi-view sketches with diverse styles representative of inference-time inputs. Experiments show that HairFlow generates plausible 3D strands that faithfully follow the input conditions across diverse hairstyles, substantially reduces inference time, and supports flexible multimodal control.

[169] WIRE: Weight Space Implicit Representation Encoding of Frozen Video LLMs cs.CVPDF

Ameen Ali, Tamim Zoabi, Lior Wolf, Chaim Baskin

TL;DR: WIRE 将部分视频条件计算从长视觉输入激活迁移为预填充后临时低秩权重更新，并压缩视觉 token 序列，以降低冻结视频语言模型处理长视频的成本。

Details

Motivation: 长视频视觉嵌入在预填充、KV 缓存和生成阶段反复产生计算及内存开销；简单压缩又可能丢失原视频信息。WIRE 保留短视觉上下文，让权重更新作为补偿而非完全替代。

Result: 模型用问题条件双向选择扫描将每帧全分辨率特征压缩为少量载体 token，物理缩短解码器输入；超网络结合压缩前向信息和被移除的全分辨率特征读出，在选定层生成注意力及前馈投影的 LoRA 更新。更新在预填充后合并，之后自回归解码使用未修改的冻结骨干。在 Qwen3-VL-8B 每帧 16 个载体 token 时，视觉流缩短 3.5 倍；256 帧和 1,024 帧预填充时间分别降低 3.35 倍、4.14 倍，KV 缓存同比缩小，单 token 解码延迟基本不变。256 帧九个基准设置中，其平均准确率在所比较高效视频模型中最高。

Insight: 把输入压缩与临时权重补偿结合，有助于削减前向长度而保留被压缩掉的视觉信息；权重更新合并后不增加每 token 解码延迟。报告的效率数据针对 Qwen3-VL-8B 和特定载体数；摘要未说明质量下降绝对值及超网络生成/合并开销。

Abstract: Video language models represent a video as a long sequence of visual embeddings in a decoder’s input. Every visual embedding is processed by every decoder layer during prefill, stored in the key value cache, and attended by every generated token, so a long video raises prefill, memory, and generation cost at once. We present WIRE (Weight Space Implicit Representation Encoding), which reduces this recurring cost by moving part of the video conditioned computation of a frozen video language model from its input activations into a temporary low rank update of its weights. WIRE keeps a short explicit visual context, so the update acts as a correction rather than a replacement of the video. WIRE first compresses the full resolution embeddings of each frame into a small set of carrier tokens using a question conditioned bidirectional selective scan, and physically contracts the decoder input so every decoder layer, cache entry, and decoding step operates on the shorter sequence. Hypernetworks at selected decoder layers read the compressed forward pass together with a readout of the full resolution features removed by contraction and generate LoRA updates to the attention and feedforward projections. The updates are merged after prefill, so autoregressive decoding runs the unmodified frozen backbone. On Qwen3-VL-8B, WIRE with 16 carrier tokens per frame contracts the visual stream by $3.5\times$, reduces decoder prefill time by $3.35\times$ at 256 frames and $4.14\times$ at 1024 frames, shrinks the key value cache proportionally, and keeps per token decoding latency approximately constant. Across nine benchmark settings at 256 frames, it obtains the highest average accuracy among efficient video language models we evaluate under a common 256 frame budget.

[170] Bottlenecked semantic compression by representation reconstruction cs.CV | cs.AIPDF

Xiaoya Tang, Shilong Liu, Daniel Smolyak, Xiwen Li, Quynh Nguyen

TL;DR: ReConV 用紧凑 Summary Bank 重建视觉 token 序列，并保留重建残差大的细节 token，从而实现带语义瓶颈的视觉 token 压缩。

Details

Motivation: 视觉 Transformer 处理长 token 序列计算成本高；常见压缩方法逐个排序 token 并丢弃低重要性项，可能遗漏只有组合后才显现的语义信息。

Result: 在每个压缩层，位置条件解码器从 Summary Bank 重建当前 token；共享信息由摘要库保存，而未被充分重建的细节通过残差显式保留。作者在图像分类和多模态理解上评测，报告 ImageNet 计算量减少 48% GMAC 时仍优于未压缩基线，并可复用图像级表示以支持高效多轮推理。摘要未给出分类准确率、其他基准或多轮推理的效率数字。

Insight: 按摘要表征的重建误差决定保留什么，能将压缩重点从单 token 的独立排序转向“已有记忆是否覆盖此信息”。ImageNet 上以更少计算超过未压缩模型是值得注意的摘要结论，但未给出具体准确率和模型/训练对比，需正文核实。

Abstract: Vision Transformers incur high computational costs when processing long visual token sequences. Existing reduction methods typically rank tokens independently and discard those deemed unimportant or redundant. We propose a reconstruction-guided framework for visual token condensation (ReConV), which summarizes the complete visual sequence with a compact Summary Bank (SMB) and retains tokens whose information is not adequately captured by the SMB, as indicated by their reconstruction residuals. At each compression layer, a lightweight position-conditioned decoder reconstructs the current tokens, leaving shared information compressed in the SMB while explicitly preserving poorly reconstructed details. Experiments on image classification and multimodal understanding demonstrate favorable performance-efficiency trade-offs: ReConV surpasses the uncompressed baseline with 48% fewer GMACs on ImageNet and enables reusable image-level representations for efficient multi-turn inference. Code is available at TokenDrop.

[171] World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models cs.CV | cs.AI | cs.MMPDF

Junyi Wu, Zekun Li, Chenhongyi Yang, Chen Guo, Lingchen Yang

TL;DR: World-T3 将测试时训练作为视频世界模型的持久视觉记忆，以相机位姿为键递归吸收旧场景特征，并蒸馏成少步生成器以支持实时长程交互。

Details

Motivation: 长程自回归世界生成需要记住超出当前上下文的先前场景；现有模型通常隐式依赖生成模型本身保留历史，因而容易出现空间记忆衰减。

Result: World-T3 用递归最小二乘规则在线更新位姿键控快速权重，将历史观测巩固到专用参数中，并根据相机位姿读回先前场景特征；作者称存储和单步计算量不随 rollout 长度增长。再通过双教师分布匹配把自回归模型蒸馏为少步生成器。在长程重访基准上，相较两项指标均最强的基线，重访 PSNR 从 13.52 提至 18.31 dB，绝对旋转误差从 3.344 降至 1.075；单张 NVIDIA B200、1280×704 分辨率下达到 20.6 FPS。

Insight: 把记忆写入在线更新的快速权重而非仅保留上下文，针对的是长时重访时的场景遗忘；位姿寻址和固定存储/单步计算目标也有利于长 rollout。所报速度针对单张 B200 与指定分辨率，摘要未说明记忆更新稳定性、不同场景容量或额外训练成本。

Abstract: Recent advances in video world models have enabled plausible interactive environments, yet long-horizon generation places a fundamental demand on visual memory: the previously observed world must remain consistent as the model continues to generate. Existing autoregressive approaches often implicitly rely on the generative model itself to retain such information, leading to spatial memory decay beyond the active context. We introduce World-T3, a camera-controlled video world model with a test-time training memory that consolidates past observations into dedicated model parameters. To robustly update this memory, we propose a recursive least-squares rule that incrementally absorbs novel visual features via online pose-keyed memory updates. Our TTT memory reads out previously observed scene features from the fast-weight state by camera pose, enabling consistent generation with constant storage and per-step computation regardless of rollout length. To facilitate real-time interactive generation with this robust memory, we further distill the autoregressive model into a few-step generator via dual-teacher distribution matching. On a long-horizon revisit benchmark, World-T3 raises revisit PSNR from 13.52 to 18.31 dB and reduces absolute rotation error from 3.344 to 1.075 relative to the strongest baseline on both metrics, while sustaining 20.6 FPS at 1280 x 704 on a single NVIDIA B200 GPU.

[172] TableVision: Recovering Verified Structure from Historical Tables cs.CVPDF

Todd Y. Zhou, James M. Snyder

TL;DR: TableVision 将历史表格恢复重点放在把数字放回正确单元格并验证列求和，以应对缺行、跨页、错列和原始账目算术不一致。

Details

Motivation: 历史登记表常有缺失行、跨页、列错位及行列合计不符；OCR 即便读对数字，仍可能破坏单元格结构，造成数据无法可靠复用。

Result: 在 7 个干净选举表面板上，最佳开源读取器识别数字准确率 99.0%，但仅 83.7% 数字归入正确单元格，且 30 列中只正确 11 列。三张退化的 1888 年页面上，Amazon Textract 仅正确安置 1/1,494 个数字，TableVision 为 99.2%；其精确总和校验放行的 51 列全部正确。动态规划按列合计边界切分区块，读取器提供单元格证据，门控器将列判为精确对账、忠实标记原表算术错误或暂不采纳。威斯康星选举数据 2,201 列中，接受 2,042、标记 118、搁置 41；一般选举每千单元格约 1.1 处错误，失败率随列长 R 按 1−(1−θ)^R 上升。官方结果验证 934 个可检验接受列中的 933 个及全部 35 个可检验标记列；部署数据集增加 491,359 条经门控选区记录，原数据集 930 万行。

Insight: 字符识别和结构恢复是不同问题，算术总和提供了可审计的结构校验，但无法发现所有会相互抵消的错误；区块大小和列长会影响校验能力。实际部署通过接受、标记和搁置区分证据强弱，避免把总和校验视作对每列都充分的保证。

Abstract: Gaps, skips, and mistakes are the normal state of a historical register: rows go missing, tables cross page breaks, columns slip under wrong names, and column sums disagree with the corresponding row entries. Readers read digits correctly but lose the table structure. On seven clean panels of election returns, the best open-source reader reproduces 99.0% of the digits but files only 83.7% of the figures in the correct cells and gets 11 of 30 columns correct. TableVision treats table structure recovery as the problem. On three degraded 1888 pages, Amazon Textract files only 1 of 1494 figures correctly; TableVision correctly files 99.2%, and all 51 columns its exact-total check releases are correct. A dynamic program cuts a volume into blocks by how many columns sum to their last line; witnesses read the cells; the gate classifies each column as exactly reconciling, flagged as faithful to a page whose arithmetic is wrong, or withheld. A theory of the sum check says which errors it cannot see and why its optimal block is about $θ^{-1/2}$ rows at defect rate $θ$. On 2201 columns of Wisconsin electoral data, the gate accepts 2042, flags 118, and withholds 41; failure rises with column length $R$ as $1-(1-θ)^R$ at 1.1 defects per 1000 cells in general elections; official returns confirm 933 of 934 testable accepted sums and all 35 testable flagged totals; and the deployment added 491,359 gated precinct rows to a 9.3-million-row dataset. One implication is that the design of the verification block is critical.

[173] Know-Thy-Not: Training-Free Targeted Negation in Dual-Encoder Vision-Language Models cs.CV | cs.AIPDF

DoYoung Kim, Changkyu Choi, Adín Ramírez Rivera, YunSeok Choi, Jee-Hyong Lee

TL;DR: Know-Thy-Not 在双编码器视觉语言模型中为图像表示增加显式的否定概念排斥维度，无需训练即可改善对“有某物但没有另一物”这类文本的理解。

Details

Motivation: 双编码器常把“有猫但没有狗”和“有猫且有狗”映射为相似分数。作者指出单嵌入对比框架至多忽略被否定概念，无法直接惩罚图像中出现该概念。

Result: Know-Thy-Not（KTN）以训练免除方式在图像侧引入拒绝维度，对否定概念的存在直接扣分；作者还提出需要真正排斥能力才能通过的 COCO-Confusion 基准。在 9 种骨干上，KTN 使该基准 mAP 相较原始 VLM 最高提高 25.6 分，相较最强既有方法最高提高 9.6 分。

Insight: 明确建模“存在时应排除”的视觉维度，区别于只稀释否定词影响的语义相似度方法。摘要报告的是最高增益而非所有骨干的平均增益；新基准覆盖范围及对复杂自然语言否定的泛化仍需结合正文分析。

Abstract: Dual-encoder vision-language models systematically fail at negation, scoring ‘’a cat without a dog’’ nearly identically to ‘’a cat with a dog.’’ We show that this failure is inherent to the single-embedding framework. Under contrastive training, single-embedding methods can at best ignore the negated concept (dilution) but never penalize its presence (rejection). We propose Know-Thy-Not (KTN), a training-free method that introduces an image-side rejection dimension to penalize the negated concept directly in the visual representation. We also introduce COCO-Confusion, the first benchmark requiring genuine rejection to succeed. Across 9 backbones, KTN improves COCO-Confusion mAP by up to +25.6 over vanilla VLMs and +9.6 over the strongest existing method.

[174] CrossPortrait: Cross-Species Portrait Animation via Adversarial Motion Alignment cs.CVPDF

Gezhong Pan, Zhe Li, Yunfan Tao, Yongheng Zhao, Shiqi Zhou

TL;DR: CrossPortrait 将面部运动拆成刚性头姿和表情关键点，用物种路由编码器与表达级对抗域适配实现跨物种肖像动画并减轻驱动物种信息泄漏。

Details

Motivation: 用物种 B 的驱动视频动画化物种 A 的肖像时，面部解剖结构差异很大；把人像动画模型微调到动物数据上容易把驱动物种特征泄漏到目标动画。

Result: CrossPortrait 支持人到动物、动物到人及不同动物间动画。统一表情编码器内置物种特定 LoRA，以适配不同面部结构；表达级对抗域适配抑制表情潜变量中的驱动物种信息，促使形成跨物种共享表达空间。作者报告在不同物种设定下能生成富有表现力且保持身份的动画，并超过所比较的先进方法。摘要未给出数据集、主客观指标或提升数值。

Insight: 把头部姿态和表情分开建模，并显式去除物种来源线索，有助于在保留目标身份的同时跨越结构差异。结果目前仅定性描述；摘要未提供真实度、身份保持、物种组合覆盖或视频泛化的量化证据。

Abstract: Cross-species portrait animation, which aims to animate a source portrait of species A using a driving video of species B, is particularly challenging due to substantial anatomical differences in facial structures across species. Previous works typically fine-tune human portrait animation models on animal video datasets but often suffer from driving-species leakage. To address this, we propose CrossPortrait, a novel framework for expressive cross-species portrait animation, supporting diverse scenarios including human-to-animal, animal-to-human and cross-species animal-to-animal (e.g. cat-to-dog) animation. Specifically, we decompose the portrait motion into rigid head poses and expression keypoints. To accommodate species-specific facial structures while facilitating cross-species expression alignment, we introduce a species-routed expression encoder that incorporates species-specific LoRAs into a unified expression encoder. Furthermore, we introduce expression-level adversarial domain adaptation (Ex-ADA) to suppress driving-species-specific information in the expression latents, thereby mitigating species leakage and promoting a shared expression space across species. Extensive experiments show that our method enables expressive and identity-preserving portrait animation across diverse species and outperforms other state-of-the-art approaches.

[175] FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching cs.CV | cs.AIPDF

Ruchang Yao, Shi Guo, Junhao Zhuang, Mingde Yao, Xin Cai

TL;DR: FlashVSR+ 通过局部为主、保留稀疏全局交互的混合注意力降低高分辨率视频超分计算与内存，并用奖励加权教师 rollout 的分布匹配蒸馏对齐人类偏好。

Details

Motivation: 扩散蒸馏可实现一步视频超分，但高分辨率推理仍昂贵，且蒸馏生成器难以对齐人类偏好；因此架构和训练策略都需要改进。

Result: 混合注意力利用预训练视频超分模型中的局部性，同时保留长程一致性所需的稀疏全局交互，DiT 最高加速 2 倍、峰值内存降低 45%，支持单张 24GB RTX 4090 对 4K 视频进行不切块推理。Ry-DMD 使用奖励加权教师 rollout 注入黑盒奖励进行分布匹配，无需对奖励模型反向传播；作者还报告该方法可迁移到文生图。整体上，作者称复原质量和用户偏好稳定改善，但摘要未给具体质量分数或偏好实验规模。

Insight: 局部注意力与稀疏全局交互针对高分辨率效率，而奖励加权分布匹配将偏好信号加入一步生成训练。2 倍加速和 45% 内存下降是最高/报告值；摘要未列出各分辨率的质量折损、用户研究统计或 Ry-DMD 在文生图上的量化结果。

Abstract: Recent video super-resolution (VSR) models achieve efficient one-step generation through diffusion distillation. However, high-resolution inference remains costly, and distilled models are difficult to align with human preferences. These challenges call for improvements in both model architecture and training strategy. We present \textbf{FlashVSR+} to address both limitations. On the architecture side, we introduce hybrid attention based on the strong locality observed in pretrained VSR models, while retaining sparse global interaction for long-range consistency. It achieves up to $2\times$ DiT speedup and 45% lower peak memory, enabling spatially untiled 4K VSR on a single 24GB RTX 4090. On the training side, we propose Reward-Yield Distribution Matching Distillation (Ry-DMD) for preference-aligned one-step generation. Ry-DMD uses reward-weighted teacher rollouts to incorporate black-box rewards into distribution matching, without backpropagating through the reward model. We further show that Ry-DMD generalizes beyond VSR to T2I generation. Together, FlashVSR+ delivers stronger restoration quality and substantially improves high-resolution efficiency, with consistent gains in user preference. We release our code, models, and training data at \href{https://www.flashvsr.tech}{\textcolor{FlashBlue}{\textbf{www.flashvsr.tech}}}.

[176] TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models cs.CVPDF

Kai Liu, Anqi Li, Junxian Li, Zhixin Wang, Zhikai Chen

TL;DR: TwinPrune 用统一空间密度信号分别压缩预填充阶段的视觉 token 和解码阶段的 KV 缓存，并强调解码剪枝必须用完整生成任务评测。

Details

Motivation: 现有视觉 token 剪枝方法若只在预填充阶段使用相似度或忽略解码缓存，可能遗漏两个阶段的不同需求；只用首 token 评分的基准也无法验证解码期剪枝效果。

Result: TwinPrune 将预填充 bipartite merge 和解码 KV eviction 分开评测：MMBench、POPE、ScienceQA 首 token 指标主要反映预填充；TextVQA、DocVQA、MM-Vet 完整生成指标覆盖两阶段。实验显示，解码时丢弃全部视觉 KV 几乎不影响首 token 指标，却严重降低 TextVQA 和 MM-Vet，说明这是评分规则盲点而非压缩强度。方法在 7 个模型变体、3 个完整生成基准形成计算、内存、准确率三方 Pareto 前沿；预填充密度压缩对 token 数线性扩展，高分辨率 token 选择快达 100 倍。解码密度 top-k KV 淘汰在三种保留率下优于 H2O、SnapKV、StreamingLLM；TextVQA 保留率 0.125 时比 H2O 高 9.6 分，InternVL3.5 2B–38B 模型优势为 7–10 分。模型从 8B 增至 38B 时，准确率损失反而降低。

Insight: 首 token 准确率不能用于证明解码期剪枝有效；解码主张需要 TextVQA、DocVQA 等完整生成评测。空间密度同时服务两个阶段，但其效果与保留率、骨干规模和任务相关，摘要建议将 MM-Vet 或 DocVQA 作为解码压缩验证项。

Abstract: We propose a unified spatial-density signal to score token importance for both inference phases: bipartite merge for prefill and KV eviction for decode, yielding measurable gains in each stage. To isolate the two phases, we split benchmarks by scoring logic. Prefill probes (MMBench, POPE, ScienceQA) evaluate only the first decoded token; their accuracy reflects prefill logits argmax before any decode KV eviction. Full-pipeline probes (TextVQA, DocVQA, MM-Vet) assess the complete generated sequence and capture both phases. We identify a critical evaluation pitfall: decode pruning claims validated solely on prefill probes measure prefill quality, not decode performance. Experiments show discarding all visual KV during decode barely affects prefill-probe accuracy but degrades TextVQA and MM-Vet drastically; this insensitivity stems from the scoring rule, not compression intensity.Evaluated on seven model variants (InternVL-3.5, Qwen3-VL, LLaVA-1.5) and three full-pipeline benchmarks, our method achieves a three-way Pareto frontier across compute, memory and accuracy. Density-based prefill compression maintains strong accuracy, scales linearly with visual token count, and delivers 100x faster token selection at high resolution. For decode, density top-k KV eviction consistently outperforms strong baselines H2O, SnapKV and StreamingLLM on TextVQA and MM-Vet across three retention ratios. At 0.125 keep ratio on TextVQA, it surpasses H2O by 9.6 points; InternVL-3.5 models (2B to 38B) maintain a 7 to 10 point advantage. Compression scales favorably with model size: every measured 8B to 38B model transition reduces accuracy loss instead of worsening it. We suggest MM-Vet or DocVQA as mandatory full-pipeline validation for any decode-phase claims. Our code is available at https://github.com/Kai-Liu001/TwinPrune

[177] Physical-Field-Bridged Material Parameter Inference from Multi-View Deformation Videos cs.CV | cs.ROPDF

Linrui Shou, Jie Li, Yu Liu, Tao Huang

TL;DR: 该方法用多视角形变视频及稠密物理场监督，从视觉输入单次推断材料类别和连续参数，并以物理响应场作为外观与类别捷径之外的桥接表示。

Details

Motivation: 可供模拟使用的材料参数对可变形物体操作重要，但材料性质只能通过物理响应间接观察；逆物理方法常需逐场景重建、在线调参或仿真闭环优化。

Result: 作者提出物理场桥接推断，同时预测力/接触掩码、投影运动流场和投影应力图，再以这些场为条件推断材料参数。DeformFieldBench 含 5,000 次仿真，覆盖 100 种形状和五种具力学信息的加载模式，并提供像素对齐物理监督。方法在四个连续参数上取得最低中位误差，应力热点召回率从 0.0762 提至 0.5121；未见物体的同模拟器回放达到 SSIM 0.8676、PSNR 36.82 dB。匹配协议的 GSO 评测中，推断和未来状态预测约需 30 秒，逐场景优化需 15–120 分钟；另在五段真实序列上观察到可测回放一致性。跨模拟器和真实场景仍存在域差异与可辨识性缺口。

Insight: 密集预测中间物理响应能把视频变化和材料参数联系起来，减少直接从外观或类别猜测。合成数据及同模拟器评测带来较高数值，但作者也指出跨模拟器与真实世界差距；五段真实序列的结果不等同于全面的实物参数准确性验证。

Abstract: Estimating simulation-ready material parameters from visual observations is crucial for deformable-object manipulation but challenging because material properties are only indirectly revealed through physical responses. We introduce physical-field-bridged material inference, a single-pass framework that uses dense physical responses to connect multi-view RGB deformation videos with constitutive-type classification and continuous parameter regression. Unlike inverse-physics methods requiring per-scene reconstruction, online tuning, or simulation-in-the-loop optimization, our model predicts force/contact masks, projected motion-flow fields, and projected stress-response maps, then conditions material inference on these fields to reduce reliance on appearance or category shortcuts. We construct DeformFieldBench, a 3DGS MPM learning environment comprising 5,000 simulations across 100 object shapes and five mechanically informative loading modes, with dense pixel-aligned physical supervision. Our method achieves the best median errors on all four continuous parameters, improves stress-hotspot recall from 0.0762 to 0.5121, and attains 0.8676 SSIM/36.82 dB PSNR in same-simulator replay on unseen objects. Under a protocol-matched GSO evaluation, inference and future-state prediction take approximately 30 seconds, compared with 15-120 minutes for per-scene optimization. The method also yields measurable replay agreement on five real-world sequences. These results support dense physical fields as an effective bridge for fast material inference, while cross-simulator and real-world tests reveal remaining domain and identifiability gaps.

[178] Benchmarking Self-Supervised 3D Medical Vision Foundation Models and Transferability Estimation cs.CV | cs.AIPDF

Jiaqi Tang, Yu Wang, Yang Liu, Jing Tao, Fandong Zhang

TL;DR: 该研究统一比较 23 个自监督 3D 医学视觉基础模型在 14 项分割、分类和回归任务上的迁移能力，并提出 InsightRank 在微调前低成本筛选候选模型。

Details

Motivation: 已发表的 3D 医学模型评测使用不同数据、终点和微调协议，模型架构、预训练目标及数据也不相同，使结果难以横向比较；对所有候选模型逐一微调目标任务计算成本较高。

Result: 统一基准覆盖多解剖部位和模态。没有任何模型在所有分割任务或所有分类/回归任务中均进入前三；同一预训练目标下卷积编码器可在某任务超过 Transformer、在另一任务落后。InsightRank 以任务适配的冻结特征统计与架构感知成对比较估计微调排名，23 模型池上的平均 Kendall τ-b 在分割为 0.431，在分类/回归为 0.352。分割前三短名单包含 9 个任务中 4 个任务的赢家；图像级任务前三未包含 5 个任务中的任何赢家，尽管总体排名相关性优于匹配的传统估计器。七项分割任务中，三模型短名单令记录的特征提取加训练作业时间减少 85.8%。

Insight: 模型选择对目标任务高度依赖，不能从单一架构或预训练指标推断普遍优劣。InsightRank 可减少候选微调，但分类/回归的排序相关性较低，Top-3 命中也随任务类型变化；它是优先级工具而非替代目标数据微调验证。

Abstract: Self-supervised pretraining has expanded the pool of 3D medical foundation models, but published evaluations use different datasets, endpoints and fine-tuning protocols. Models also differ in architecture, pretraining objective and data, while downstream targets range from voxel-level masks to patient-level diagnoses and outcomes. As a result, published findings provide little directly comparable evidence for choosing a model for a new task. Direct comparison of adaptation performance requires fine-tuning candidates on labelled target data, at substantial computational cost. We benchmark 23 candidates across the same 14 segmentation, classification and regression tasks spanning multiple anatomies and modalities. No candidate ranked among the top three across all segmentation tasks or across all classification and regression tasks; a convolutional encoder could outperform a transformer trained with the same objective on one task but underperform it on another. We benchmark transferability estimators against the fine-tuned rankings and find that their agreement varies across dense segmentation, classification and regression. InsightRank uses task-appropriate frozen-feature statistics and architecture-aware pairwise comparisons to rank candidates before fine-tuning. Across the 23-model pool, mean Kendall tau-b was 0.431 for segmentation and 0.352 for classification and regression. The segmentation Top 3 contained a task winner on four of nine tasks; the image-level Top 3 contained none on the five tasks, despite better full-ranking agreement than the matched conventional estimators. A three-model shortlist reduced recorded extraction-plus-training job time by 85.8% across seven segmentation tasks. Together, the benchmark provides a common reference for model selection and InsightRank offers a lower-cost way to prioritize fine-tuning, with shortlist performance that differs by target type.

[179] Tanawpamana: Benchmarking Vision-Language Models on Philippine Cultural Heritage Sites cs.CV | cs.AI | cs.CLPDF

Ivan Yuri De Leon, Hiroki Ouchi, Sakriani Sakti

TL;DR: Tanawpamana 建立菲律宾文化遗产地点图像数据集及 4,046 项视觉语言模型基准，评估模型对遗产地的识别与知识。

Details

Motivation: VLM 越来越多用于地点识别与描述，但对菲律宾等代表不足地区文化的识别能力和背景知识尚不清楚。

Result: 数据集以菲律宾国家文化艺术委员会的 Talapamana 登记库为元数据基础；基准含 4,046 项任务，评测 7 个 VLM。模型在相似遗产地之间进行细粒度视觉区分时，比粗粒度识别更困难；仅凭图像推断地理知识也弱于文本输入。摘要未报告数据集规模、模型分项成绩或具体准确率。

Insight: 粗分类能力不能代表对相似文化遗产的可靠识别；图像识别与地理知识提取还呈现不同难度。基准聚焦菲律宾遗产地，摘要未说明覆盖地区、图像来源及元数据完整度。

Abstract: Vision-language models are increasingly used to identify and describe places, yet their recognition and knowledge of specific cultures remain poorly understood, particularly in underrepresented regions such as the Philippines. This study introduces the Tanawpamana Dataset, a repository of Philippine heritage site images with metadata based on the Talapamana registry of the National Commission for Culture and the Arts of the Philippines. We also introduce the Tanawpamana Benchmark, comprising 4,046 tasks designed to evaluate models’ recognition and knowledge of heritage sites. Evaluating seven vision-language models, we find that they struggle more with fine-grained visual discrimination among similar heritage sites than with coarse recognition, and that their geographic knowledge is weaker when inferred from image inputs than from text.

[180] Morphology-aware deep learning predicts colorectal cancer survival from whole-slide images cs.CV | cs.AI | cs.LGPDF

Hikmat Khan, Metin N. Gurcan, Anil Parwani, Wendy L. Frankel, Wei Chen

TL;DR: MorphContrast 以组织形态标签作为结构监督，通过形态匹配的跨患者结局对比学习，从全切片图像预测结直肠癌五年生存。

Details

Motivation: 全切片图像可用于风险分层，但弱监督生存任务中哪些组织形态提供预后信息尚未充分刻画；直接图像级学习可能混淆不同组织上下文。

Result: 研究回顾性分析 TCGA-COAD/READ 562 例以做形态分析，Alliance/CALGB 89803 的 424 例用于模型开发，SurGen 218 例用于外部评估。MorphContrast 用 gated-attention 多实例学习，并结合患者内形态对比、在匹配形态间按结局一致/不一致进行跨患者对比，以及防止表示塌缩的多样性正则。单形态模型的预后区分较弱但分布广：脂肪 AUC 0.622、基质 0.604，与肿瘤上皮 0.601–0.605 相近。Alliance/CALGB 单模型 AUC 0.642，事后集成 pooled OOF AUC 0.70，风险组 HR 3.79（95% CI 2.55–5.63）。外部 SurGen 结肠癌队列中，单模型和集成 AUC 分别为 0.669、0.678。

Insight: 形态可作为结局对比学习的结构标签，且预后信号并不只出现在肿瘤上皮；脂肪和基质也表现出一定区分能力。研究为回顾性队列，单模型和事后集成结果应区分；外部队列 AUC 低于开发队列，摘要未提供临床变量增益或前瞻验证。

Abstract: Background: Five-year survival prediction from hematoxylin-and-eosin whole-slide images could support risk stratification, but the task is weakly supervised, and the tissue morphologies carrying prognostic information remain incompletely characterized. Methods: We conducted a retrospective secondary analysis using TCGA-COAD/READ (n=562) for morphology-specific analysis, Alliance/ CALGB 89803 (n=424) for model development, and SurGen (n=218) for external evaluation. We propose MorphContrast, a morphology-aware contrastive multiple-instance learning framework that uses tissue labels as structural supervision for five-year survival prediction. MorphContrast combines gated-attention pooling with three auxiliary objectives: patch-level morphology labels define within-patient contrasts, outcome concordance or discordance defines cross-patient contrasts within matched morphologies, and diversity regularization discourages representational collapse. Morphology-matched patches from patients with discordant outcomes serve as hard negatives, encouraging prognostic discrimination within comparable histologic contexts. Results: Morphology-specific models show weak but distributed prognostic discrimination: adipose tissue achieves an area under the receiver operating characteristic curve (AUC) of 0.622, stroma achieves 0.604, and both perform comparably to tumor epithelium (0.601-0.605). On Alliance/CALGB 89803, the single model achieves an AUC of 0.642, while the post hoc ensemble reaches a pooled out-of-fold AUC of 0.70 and a risk-group hazard ratio of 3.79 (95% confidence interval: 2.55-5.63). On the external SurGen colon cancer cohort (n=218), the single model and ensemble achieve AUCs of 0.669 and 0.678, respectively. Conclusions: These findings support tissue morphology as structural supervision for outcome-aware representation learning in colorectal cancer prognosis.

[181] FrameLite: Decision-Oriented Aesthetic Cropping with a Compact Vision-Language Model cs.CVPDF

Jianze Li, Kai Liu, Ziqing Zhang, Weihang Zhang, Yong Guo

TL;DR: FrameLite 以 Qwen3.5-0.8B 为骨干，分阶段训练紧凑裁剪模型直接给出构图坐标，并通过在线候选质量监督与贪心参考优化提升裁剪质量。

Details

Motivation: 审美裁剪需要平衡主体、上下文和画面边界；大型 VLM 的语义能力有帮助但不适合轻量摄影辅助，而小模型的首次回答未必是其能生成的最佳裁剪。

Result: 三阶段训练先学习空间定位和构图偏好，再通过 QOS 从当前策略生成的候选中学习完整答案质量最高者，最后由 GRO 根据候选相对当前贪心答案的质量差优化策略。在线阶段均使用标注派生反馈；推理仅输出一组坐标，不需搜索或外部评分器。作者报告 FrameLite 在公开基准上超过所比较的大型 VLM 裁剪器；控制实验显示挑选高质量训练答案优于单纯在线采样。摘要未列基准名称、裁剪分数和延迟。

Insight: 将候选质量选择变成训练信号，能把小模型潜在能力转为单次直接预测，从而保持轻量推理。领先结论限于所评基准和大型 VLM；摘要未量化构图质量差异或人工审美一致性。

Abstract: Aesthetic image cropping improves composition by balancing subjects, context, and image boundaries. Large vision-language models bring semantic knowledge to this task, but their scale limits lightweight photography assistance. We propose FrameLite, a compact cropping model built on Qwen3.5-0.8B. Our approach is motivated by the observation that a small model can generate better crops than its first answer suggests. We develop three-stage training to turn this potential into stronger direct predictions. Spatial and composition training establishes localization and framing preferences. Quality-Guided Online Supervision (QOS) then learns the highest-quality complete answers among the current policy’s generated candidates. Greedy-Referenced Optimization (GRO) further refines the policy using positive and negative quality differences relative to its current greedy answer. Both online stages use annotation-derived feedback, while inference returns one coordinate answer without search or an external scorer. Experiments show that FrameLite outperforms the compared large-VLM croppers across public benchmarks with a substantially smaller backbone. Controlled studies confirm that selecting high-quality training answers improves the final model beyond online sampling alone. We will release our code and model weights at https://github.com/JianzeLi-114/FrameLite .

[182] 4LC: 4D Latent Cache for Efficient Conditioning in Dynamic World Modeling cs.CV | cs.AIPDF

Jianhui Zhang, Zuyao Lin, Pengxin Xu, Jingdong Wang, Xingyu Chen

TL;DR: 4LC 用带共享几何和运动的复合 4D 潜缓存存储多种视觉编码器表示，并允许按任意时间和视角读取，以避免每次生成都重新渲染和编码历史场景。

Details

Motivation: 动态世界视频生成会在不同时间和视角反复使用已观察到的静态及动态内容；逐请求进行 RGB 渲染和重新编码效率低，也难以在多种下游任务间复用统一时空条件。

Result: 4LC 将 VAE、DINO、V-JEPA、VGGT 等编码器的视觉表示绑定到共享几何和运动，并提供独立指定时间、视角的时空访问接口，可组合多次观测读出。基于此构建的 4LC-Video 用单一模型处理相机控制、时间控制、新视角合成、场景重访和长时生成等任务，无需按任务微调。实验报告条件获取更高效、时空可控性有效且跨任务场景一致性良好。摘要未给出数据集、速度或一致性数值。

Insight: 缓存潜表示而非反复重建图像，可以将多编码器的丰富视觉信息在时空查询中重复利用；共享几何/运动提供跨表征对齐。摘要只有定性效率和质量结论，缓存构建成本、存储量及长期更新策略仍未明确。

Abstract: For video generation in dynamic world modeling, observed static and dynamic content should be efficiently reused to provide conditions for repeated generation at different times and viewpoints. We introduce 4LC, a 4D latent cache with composite visual representations and a flexible spatiotemporal access interface. As a latent cache, 4LC directly reuses visual representations of static and dynamic content to provide generation conditions, avoiding per-request RGB rendering and re-encoding. As a composite cache, 4LC binds visual representations from multiple visual encoders (e.g., VAE, DINO, V-JEPA, VGGT) to shared geometry and motion, allowing the composite representations to describe the dynamic world with rich information. Besides, we design a flexible spatiotemporal interface, which directly reads the cache at independently specified times and viewpoints and combines readouts across observations, providing a common conditioning format for diverse downstream tasks. Building on 4LC, we develop 4LC-Video to address multiple video generation tasks through different configurations of source observations, spatiotemporal requests, and historical reuse, including camera control, temporal control, novel-view synthesis, scene revisits, and long-term generation. Our experiments show more efficient condition acquisition, effective spatiotemporal controllability, and scene consistency across tasks using a single 4LC-Video model without task-specific fine-tuning.

[183] MemAct: Streaming Spatial Understanding with Memory and Active Exploration cs.CVPDF

Yuxi Wei, Wei Huang, Qirui Chen, Lu Hou, Xiaojuan Qi

TL;DR: MemAct 将流式空间理解闭环扩展到主动探索：共享 VLM 写入和检索结构化历史记忆，并通过移动、转向获取新视角证据后再决定继续探索或作答。

Details

Motivation: 流式智能体需边接收观察边推理，但现有方法多侧重处理新帧，较少研究如何根据当前任务主动寻找新视角；完整空间理解还需回忆过去观察。

Result: MemAct 套件提供约 100 万训练样本，覆盖空间推理、记忆写入/检索和探索；MemAct-bench 有 20,000 个主问答评测当前观察理解、历史证据使用与主动探索，另含 2,000 道 Timing 题。方法以监督微调训练记忆和交互，再通过反事实动作优化，联合累积证据收益、动作成本和终局答案质量来学习探索及停止决策。作者报告其超过既有空间理解和流式基线，并在 MemAct-bench 接近前沿多模态模型，同时保留流式推理效率。摘要未给出具体分数或延迟。

Insight: 空间问答被建模为边回忆、边观测、边行动的统一循环，停止时机也纳入优化，避免把主动探索当成独立前处理。数据集规模较大，但摘要未提供现实/模拟样本比例、动作成本口径或不同能力的单项对比。

Abstract: Spatial understanding is essential for embodied agents navigating and interacting with the physical world. As observations arrive incrementally, agents must reason in a streaming manner. Yet streaming understanding must also be closed-loop: task-relevant evidence may require both recalling past observations and actively acquiring new views. Existing streaming approaches largely emphasize processing incoming observations, leaving task-driven active exploration underexplored. We introduce MemAct, a framework for streaming spatial understanding with memory and action. A shared vision–language model incrementally writes structured historical memory, retrieves records and past frames on demand, and acquires additional evidence through movement and turning. Retrieved evidence and action observations inform subsequent decisions, enabling the agent to continue exploring or answer within a unified reasoning loop. We develop the MemAct suite, a data generation pipeline and collection across real and simulated environments, with approximately 1M training examples for spatial reasoning, memory writing, retrieval, and exploration. Its MemAct-bench comprises 20,000 main QA pairs evaluating current-observation understanding, historical evidence use, and active exploration, plus 2,000 separate Timing questions. MemAct combines supervised fine-tuning for memory and interaction with counterfactual action optimization using cumulative evidence gains, action costs, and terminal answer quality to improve exploration and stopping decisions. Experiments show that MemAct outperforms existing spatial understanding and streaming baselines and approaches frontier multimodal models on MemAct-bench while retaining streaming inference efficiency.

[184] Open-o3-Video-v2: Seek, Ground, and Reason with Verifiable Spatio-Temporal Evidence cs.CVPDF

Jiahao Meng, Zhongyuan Bao, Yue Tan, Haiquan Wen, Renrui Zhang

TL;DR: Open-o3-Video-v2 将视频推理改为主动获取并逐步记录时空证据的智能体流程，用线索覆盖、证据增益和冗余惩罚优化工具调用。

Details

Motivation: 可靠视频推理不仅要答对，也要给出可核验的视觉依据；初代 Open-o3-Video 只能使用初始稀疏采样帧，无法主动补充遗漏信息。

Result: 模型可识别当前推理缺少的线索、获取新观察、定位支持推理的时间点和物体，并把证据写入轨迹后继续或作答。作者构建 STGR-v2-CoT135K（5.2 万条多轮带 grounding 智能体推理轨迹）和 STGR-v2-RL-13K；提出 CS-GRPO，追踪工具调用对累计证据的改变，奖励线索覆盖与逐步增益，并惩罚过多、重复或无信息检索。在 V-STaR 上，4B 版 mAM 为 40.4%、mLGM 为 56.3%，比 Qwen3-VL-4B 骨干分别高 8.5%、12.6%；其他一般视频理解、长视频推理和时间定位基准也报告持续增益。

Insight: 主动补充观察与对象级证据落地是互补能力；工具调用奖励不仅看最终答案，也考察证据覆盖和边际收益。具体结论依赖 V-STaR 指标定义及工具预算，摘要未给出其他基准具体幅度和推理成本。

Abstract: Reliable video reasoning requires models to produce correct answers and expose supporting visual evidence. Open-o3-Video enables verifiable reasoning by embedding timestamps and object-level bounding boxes into reasoning traces, but is restricted to frames from an initial sparse sampler. We introduce Open-o3-Video-v2, which formulates video reasoning as a unified agentic evidencegrounded process. The model identifies missing information, acquires additional observations, grounds the moments and objects supporting its reasoning, and writes this evidence into the trajectory before continuing or answering. To support this behavior, we construct STGR-v2-CoT135K, including 52K multi-turn agentic grounded reasoning trajectories, and STGR-v2-RL-13K for reinforcement learning. We further propose Clue-Seek GRPO (CS-GRPO), which tracks how tool calls change accumulated evidence, rewards clue coverage and step-wise gains, and penalizes excessive, repetitive, or uninformative retrieval. On V-STaR, Open-o3-Video-v2-4B achieves 40.4% mAM and 56.3% mLGM, outperforming its Qwen3-VL-4B backbone by 8.5% and 12.6%, respectively. It also delivers consistent gains across general video understanding, long-video reasoning, and temporal grounding benchmarks. These results show that active evidence acquisition and verifiable object-level grounding are complementary for reliable video reasoning. The code and models are released at https://msalab-pku.github.io/projects/Open-o3-Video-v2.

[185] ScenePercept: Scenario-Conditioned Perceptual Assessment of Real-World User-Generated Content Images cs.CVPDF

Shuhan Xia, Yu Chen, Xiang Wang, Shuai Li, Peipie Li

TL;DR: ScenePercept-Bench 提供符合真实用户生成图像分布的场景感知数据与分层标注，SCPA 通过场景条件化的评分、偏好排序和跨场景校准建模不同情境下的审美标准。

Details

Motivation: 既有图像质量或审美基准往往针对特定任务场景，并对不同类别套用统一评价标准，难以覆盖真实 UGC 图像的场景多样性与感知复杂性。

Result: 基准含 23,100 张图像，覆盖 21 类和 105 个细粒度场景；分层标注包括视觉质量、审美感知、场景语义共 15 个属性，并为每张图像提供总体分、属性分和结构化解释。Scenario-Conditioned Perceptual Alignment 联合优化绝对评分、偏好排序和跨场景校准，保留通用感知能力并适配场景差异。作者报告 SCPA 在 UGC 图像感知上达到当时最佳，并能泛化到既有 IQA 和 IAA 基准。摘要未列各基准数值或标注者规模。

Insight: 场景可改变人类对质量和审美的判断，因此模型需要学会何时采用不同标准，而不是只拟合一个全局分数。基准与分层解释提供丰富监督；摘要未说明属性评分一致性、跨文化差异和场景分布对泛化的影响。

Abstract: The rapid development of mobile imaging and social platforms has led to explosive growth in user-generated content images, making their systematic and comprehensive perceptual assessment increasingly important. Existing image perception benchmarks are typically task-oriented, curating specific scenarios for quality or aesthetic assessment and applying a uniform perceptual criterion across diverse scenarios. This limits their ability to capture the scene diversity and perceptual complexity of real-world UGC images. To address these limitations, we introduce ScenePercept-Bench, a benchmark designed to reflect the distribution of real-world UGC images, comprising 23,100 images across 21 categories and 105 fine-grained scenes. To account for the influence of scene context on human perception, we design a hierarchical, scene-aware annotation protocol, covering visual quality, aesthetic perception, and scene semantics, with 15 attributes in total. Each image is annotated with an overall score, 15 attribute-level scores, and structured explanatory reasoning. Building on this benchmark, we further introduce Scenario-Conditioned Perceptual Alignment, a training framework that tackles the heterogeneity of perceptual criteria across scenarios. SCPA jointly optimizes absolute score prediction, preference ranking, and cross-scenario calibration, preserving the model general perceptual assessment ability while encouraging it to explore and adapt perceptual criteria across different scenarios. Extensive experiments demonstrate that SCPA achieves state-of-the-art performance in UGC image perception and generalizes effectively to established IQA and IAA benchmarks.

[186] SWORD: Scene-Grounded Wasserstein Object-Relation Distance for 3D Scene Graph Prediction cs.CVPDF

U-Chae Jun, Jaeeun Ko, Yoonhyuk Choi, Jiwoo Kang

TL;DR: SWORD 将对齐的 3D 场景图预测转化为带对象角色和关系语义的 Wasserstein 距离监督，使错误图根据与真值的语义接近程度获得不同反馈。

Details

Motivation: 常见 3D 场景图训练只优化对象节点与关系边的精确分类，两个错误图可能受到相同惩罚，即便其中一个在对象角色和关系上下文上更接近真值；通用图编辑距离又需搜索对应关系并设计编辑代价。

Result: 作者利用固定节点索引和受监督边槽，将场景图比较转为对齐图距离问题。SWORD 从训练标注按确定性 base-to-context 顺序构造分块归一化对象/关系地面距离，不使用外部语义先验；再扩展为对齐节点边际分布和加入空边的 Wasserstein 距离。在对齐边际概率图空间中，真值仍是唯一零距离解。理论上作者证明 SWORD 是严格度量、保留场景图语义替换排序、控制可加语义观测，并给出对解码语义误差的代理上界。当前摘要未报告具体下游实验数值。

Insight: 固定对齐协议让语义场景图差异可用可证明的 Wasserstein 度量表达，补充精确分类损失无法区分错误严重程度的问题。理论保证建立在对齐边际概率空间和作者定义的地面距离上；摘要未提供实际预测性能或训练稳定性结果。

Abstract: 3D scene graph prediction is commonly optimized with object-node and relation-edge classification losses. These objectives encourage exact object and relation label recovery, but they do not directly optimize semantic proximity to the ground-truth scene graph: two wrong graphs can receive comparable penalties even when one is much closer in object roles and relation contexts. A direct graph-distance loss is difficult in general, since graph matching and graph edit distance require correspondence search and edit-cost design. Our key observation is that prior supervised 3D scene graph prediction protocols are conducted on fixed object-node indices and supervised edge slots. This can turn a scene graph comparison into a well-defined aligned graph-distance problem. To this end, we propose SWORD, a Scene-Grounded Wasserstein Object-Relation Distance that builds block-normalized object and relation ground distance metrics from training scene graph annotations through a deterministic base-to-context order, without external semantic priors. SWORD lifts these metrics to Wasserstein distances over aligned node and null-augmented edge marginals, providing semantic graph-distance supervision beyond exact-label penalties while preserving the ground-truth graph as the unique zero-distance solution in aligned marginal probability space. We prove that SWORD is a strict metric on aligned marginal probabilistic scene graphs, preserves scene-graph-grounded substitution ordering, controls additive graph-grounded semantic observables, and yields a surrogate upper-bounding decoded semantic error.

[187] Hierarchical Curriculum Learning for Surgical Video Generation: Focusing on Bilateral Axillo-Breast Approach Robotic Thyroidectomy Datasets cs.CVPDF

Youngbin Kong, Seungho Lee, Jae Kyung Shin, Jong-hyuk Ahn, Hyoun-Joong Kong

TL;DR: 该工作用阶段—任务—动作三级术式标注和由粗到细课程迁移，将通用文生视频扩散模型适配到腋窝—乳房入路机器人甲状腺切除术视频生成。

Details

Motivation: 高保真手术视频生成可支持教学、模拟和 AI 训练，但手术解剖结构复杂、器械—组织交互细致、流程具层级性，且多数术式缺少整理好的训练数据。

Result: 作者在尚无整理基准的双侧腋窝—乳房入路机器人甲状腺切除术上实例化框架，将术式拆成 phase、task、action 三级，并逐步适配粗到细的手术上下文；动态采样策略处理不同分辨率和长度片段。相比前一任务级模型，最终动作级模型 FVD 降低 20.1%、FID 降低 15.7%；匹配预算消融显示改善来自课程设计而非额外训练。6 位临床医生的用户研究认为视觉真实度较高，但仍指出解剖准确性和临床可靠性有差距。

Insight: 按手术层级逐步学习能让生成模型从流程语境深入到细粒度动作；匹配训练预算消融增强了课程设计带来收益的证据。仅有单一术式案例且没有现成基准，临床专家也明确指出真实性与临床可靠性仍待提高。

Abstract: High-fidelity surgical video generation is increasingly important for surgical education, simulation, and downstream AI training, as conventional simulation pipelines are limited by cost, realism, and scenario diversity. Adapting generic text-to-video diffusion models to surgical procedures remains difficult, however, due to complex anatomical structures, intricate tool-tissue interactions, the inherently hierarchical nature of surgical workflows, and the scarcity of curated training data outside a few well-studied procedures. We propose a surgical video generation framework that combines hierarchical procedural annotation with curriculum transfer learning. Surgical procedures are decomposed into three levels (phase, task, and action) and a pretrained text-to-video diffusion model is progressively adapted from coarse to fine-grained surgical context, supported by a dynamic sampling strategy that handles variable-resolution and variable-length surgical clips. We instantiate the framework on Bilateral Axillo-Breast Approach robotic thyroidectomy, a procedure for which no curated benchmark currently exists. Relative to the preceding task-level model, the final action-level model reduces FVD by 20.1% and FID by 15.7%, and a matched-budget ablation indicates that these gains arise from the curriculum design rather than from additional training. A user study with six clinicians rates visual realism highly while highlighting remaining gaps in anatomical accuracy and clinical reliability. The proposed framework offers a step toward AI-driven surgical simulation in procedures where curated data are scarce.

[188] DT-LoRA: Reaching Beyond the Tangent Space of Low-Rank Adaptation cs.CV | cs.LGPDF

Yangming Zhang, Ling Yi, Datong Li, Jinliang Ding, Xiaojie Wang

TL;DR: DT-LoRA 将固定秩预算拆为稳定分支和动态分支，以矩阵无关随机 SVD 捕获当前低秩切空间之外的主要残差梯度方向。

Details

Motivation: LoRA 更新受当前低秩因子切空间限制，可能无法利用切空间外的重要梯度，从而落后于全量微调。对十种 LoRA 变体的分析发现，这些未利用残差占多数梯度能量且谱集中。

Result: DT-LoRA 在固定秩内分配稳定与动态分支，并论证残差方向在正交补和秩约束下是局部二次模型的最优更新。矩阵无关随机 SVD 无需显式构造残差梯度；受控写回可保留累计更新并稳定训练。作者称其相较 LoRA 及变体在 CV、NLP 任务上持续提升；作为通用适配方法，它在开放词汇语义分割领域泛化上也超过专用方法并达到当时最佳结果。摘要未列各任务数值。

Insight: 问题不只是低秩容量大小，也在于动态梯度方向是否处于当前因子所张成空间；在相同秩预算内保留一条分支追踪新方向，可扩大可达更新范围。摘要宣称不增加 LoRA 可训练参数或推理成本，但未提供训练开销及不同秩预算的具体对比。

Abstract: Low-rank adaptation (LoRA) often underperforms full fine-tuning. We investigate this performance gap through a tangent-space lens: factorized updates are confined to the tangent space of the current low-rank factors, leaving gradient components outside this space unexploited. An analysis of ten representative LoRA variants reveals that these residual components account for most of the gradient energy and exhibit strong spectral concentration. Motivated by this finding, we propose Dynamic Tangent LoRA (DT-LoRA), which partitions a fixed rank budget into a stable branch and a dynamic branch to capture the principal residual gradient directions. We show that these directions define the optimal update of a local quadratic model under orthogonal-complement and rank constraints. Matrix-free randomized SVD recovers them without materializing the residual gradient, while controlled write-back preserves accumulated updates and stabilizes training within the allocated rank. DT-LoRA adds no trainable parameters or inference cost over LoRA. Experiments across computer vision and natural language processing tasks demonstrate consistent improvements over LoRA and its variants. Moreover, despite being a general-purpose adaptation method, DT-LoRA outperforms task-specific approaches to open-vocabulary domain generalization for semantic segmentation, achieving state-of-the-art results. Code is available at https://github.com/Snipine/DT-LoRA.

[189] A Differentiable 3D Scene Graph Metric via Contextual Hellinger Triplet Geometry cs.CV | cs.LGPDF

U-Chae Jun, Yoonhyuk Choi, Jiwoo Kang

TL;DR: 该研究提出基于上下文 Hellinger 三元组几何的图级监督，将 3D 场景图对象、谓词和关系上下文联合比较，而非只优化局部槽位分类。

Details

Motivation: 局部对象和谓词分类损失没有表达预测图与真值图之间的整体语义距离，可能把轻微标签错误和破坏结构的关系错误惩罚得相似，也不会把主语—谓词—宾语事实作为场景级单元比较。

Result: Contextual Hellinger Triplet Geometry 将分类器输出解释为对齐的概率场景图，其中对象和谓词联合定义上下文三元组，再以可微图级距离训练。其因子化实现效率较高，并可用于既有场景图架构而无需改网络结构。在 3DSSG 上，作者报告该监督改善下游图条件 3D 场景生成；消融显示不同图级项互补，用户研究认为生成图描述性、关系合理性和结构保留更好。摘要未给具体指标数值。

Insight: 从逐槽位预测转向比较整张图的上下文三元组，可让训练信号反映关系结构及语义接近程度；因子化实现降低图距离计算负担。结果仅在 3DSSG 及相关生成评测上定性报告，摘要未提供图预测本身的数值增益。

Abstract: 3D scene graph prediction is commonly supervised with local object and predicate classification losses. Although effective for slot-wise label prediction, this formulation does not provide a graph-level semantic distance between a predicted scene graph and its ground truth: semantically mild label errors and structurally disruptive relational errors can be treated similarly, and subject-predicate-object facts are not compared as scene-level units. We propose Contextual Hellinger Triplet Geometry, a graph-level supervision objective that measures the semantic and structural discrepancy between 3D scene graphs. Our key idea is to reinterpret the conventional classifier outputs of a 3D scene graph predictor as an aligned probabilistic scene graph field, where object and predicate predictions jointly define contextual subject-predicate-object facts. This formulation yields a differentiable metric-based training objective with an efficient factorized implementation and can be applied to existing 3D scene graph prediction architectures without modifying their network design. Experiments on 3DSSG demonstrate that our supervision improves downstream graph-conditioned 3D scene generation, while ablation studies confirm the complementary roles of the proposed graph-level terms and user studies show better graph descriptiveness, relation plausibility, and structural preservation.

[190] From Token Correction to Claim Admission: Training-Free Hallucination Mitigation for Large Vision-Language Models cs.CV | cs.AIPDF

Hyeonggeun Han, Taehyun Cho, Sangwoo Hong

TL;DR: EGVCA 在图像支持足够时才准入视觉语义主张，通过比较真实图像与视觉空输入下的主张似然，训练免除地减少 VLM 对象幻觉。

Details

Motivation: 视觉语言模型常生成输入图像并不支持的对象描述；仅修改 token 分布可能造成生成层面的副作用，而可靠性判断更直接的问题是某个语义主张是否获得图像证据。

Result: EGVCA 用观察图像与视觉空输入下的主张似然差衡量图像依赖支持。图像描述任务中，它控制语义扩张而不改 token 生成；二元 VQA 则结合原子视觉支持和图像对模型原生 Yes/No 分数差的影响。在 CHAIR、POPE、MME 三个基准和三种 LVLM 骨干上，作者报告对象幻觉显著减少且描述质量保留；9 种骨干—采样设置中有 8 种 POPE F1 最高，三个骨干的原生 MME 总分均提高。

Insight: 以主张为单位检查图像相对于空输入增加了多少支持，提供了区别于逐 token 分布修正的训练免除方案。改进集中于所评基准与骨干；摘要未说明主张抽取错误、延迟或长文本中多主张交互的影响。

Abstract: Large vision-language models often exhibit object hallucination, generating claims that are not supported by the input image. To mitigate this issue, we propose Evidence-Guided Visual Claim Admission (EGVCA), a training-free framework that admits a semantic claim only when it receives sufficient image-dependent support. EGVCA measures this support by comparing the claim likelihood under the observed image and a visual-null input. For captioning, it controls semantic expansion without altering token generation. For binary visual question answering, it combines atomic visual support with the image-induced change in the model’s native Yes–No margin. Experiments on CHAIR, POPE, and MME across three LVLM backbones show that EGVCA substantially reduces object hallucination while preserving caption quality. It achieves the highest POPE F1 in eight of nine backbone–sampling settings and improves the native MME total score on all three backbones. These results demonstrate the effectiveness of claim-level admission as a simple alternative to token-level distribution correction.

[191] MHVF: Training-Free GUI Grounding via Verification-Guided Multi-Hypothesis Search cs.CVPDF

Minghua Chen, Xingming Liao, Weiqiang Zhong, Zhuowei Wang

TL;DR: MHVF 通过全屏多假设搜索、局部放大重定位及指令一致性验证，在无需训练或额外模型的情况下提升 GUI 元素 grounding。

Details

Motivation: 密集界面和外观相似控件会让模型选到看似合理但错误的元素；围绕单一预测做细化可能始终探索不到其他候选。

Result: 框架用一个预训练 VLM 生成多个全屏空间假设，再对候选区域放大重定位；标记局部视图中的候选与指令匹配度由同一模型核验，分数引导第二轮局部搜索，并对两轮候选统一排序。在 5 个通用或 GUI 专用模型、8 个基准设置上，多数设置的定位准确率改善。UI-Venus-Ground-7B 在 Mind2Web 和 VisualWebBench 分别提高 4.3、2.1 个百分点；Qwen2.5-VL-7B 在 ScreenSpot-Pro 从 23.6% 提至 53.6%。

Insight: 先拓展候选再用局部图像证据核验，可缓解单次预测路径依赖，尤其适合相似小控件。方法需多轮 VLM 搜索与放大，准确率提升可能伴随推理开销；摘要未报告延迟或所有模型上的一致收益。

Abstract: GUI grounding enables web agents to translate natural-language instructions into actionable interface locations. However, dense layouts and small, visually similar controls can lead models toward plausible but incorrect elements, while refinement around a single prediction may leave alternative targets unexplored. We propose the Multi-Hypothesis Verification Framework (MHVF) for training-free GUI grounding. It combines multi-hypothesis exploration with local visual verification using a single pretrained vision-language model. To explore alternative targets, MHVF constructs spatial hypotheses from multiple full-screen predictions and examines the corresponding regions through zoomed re-localization. To resolve ambiguity among the resulting candidates, it verifies their agreement with the instruction in marked local views. Verification scores then guide a second round of local search and rank candidates retained from both search levels for final selection. Experiments with five general-purpose and GUI-specific models across eight benchmark settings demonstrate improvements in grounding accuracy in most settings. MHVF improves UI-Venus-Ground-7B by 4.3 and 2.1 percentage points on Mind2Web and VisualWebBench, respectively, and raises Qwen2.5-VL-7B’s accuracy on ScreenSpot-Pro from 23.6% to 53.6%.

[192] Multimodal Fusion of RGB and Complementary Modalities for Semantic Segmentation cs.CVPDF

Hu Cao, Yifeng Cheng, Mu He, Mengyu Li, Yinlong Liu

TL;DR: BPCF 以双向极性跨模态融合模块对齐 RGB 与辅助传感器特征，并利用两种模态的互补信息提升多模态语义分割。

Details

Motivation: RGB-X 分割可借助热成像、LiDAR、深度、事件、偏振或光场等模态，在低照度和运动模糊下提高鲁棒性；但传感器特征分布不匹配可能造成融合不稳定。

Result: 作者提出双向极性跨模态融合模块 BPCF，以加强模态间特征对齐并提取互补线索。评估六类配对：RGB-热成像、RGB-LiDAR、RGB-深度、RGB-事件、RGB-偏振和 RGB-光场；在 MFNet、KITTI-360、DELIVER、DDD17、DSEC、MCubeS、ZJU、UrbanLF 八个公开数据集上报告达到当时最佳结果。摘要未提供各数据集的分割指标或增益幅度。

Insight: 跨模态信息是否有用依赖对齐和互补关系的稳定建模，不能简单拼接不同传感器表征。摘要覆盖模态较广但只有定性领先结论，无法判断各模态组合的收益差异与失效情况。

Abstract: Multi-modal semantic segmentation adds an auxiliary sensing stream to RGB images (RGB-X), such as thermal, LiDAR, depth, event, polarization, or light field. This can improve robustness in poor lighting and under motion blur. But different sensors often produce features that do not match well, and fusion can become unstable. To alleviate these issues, we propose a bidirectional polarity-aware cross-modality fusion (BPCF) module that effectively captures complementary cues while enhancing feature alignment. We evaluate the framework on six modality pairings: RGB-Thermal, RGB-LiDAR, RGB-Depth, RGB-Event, RGB-Polarization, and RGB-Light Field. Our model achieve state-of-the-art (SOTA) results on eight public datasets, including MFNet, KITTI-360, DELIVER, DDD17, DSEC, MCubeS, ZJU, and UrbanLF. The codes and trained models will be released upon acceptance.

[193] When, Where, and Why: Hierarchical Causal Explainability for Video Face Understanding with MLLMs cs.CVPDF

Anubhav Kataria, Shreya Ghosh, Jingwen Ye, Abhinav Dhall

TL;DR: VideoFace-X-LLM 将面部视频解释拆成时间点、面部区域和因果必要性三层，并通过弱监督 ExplainFace 语料和定向遮挡评估提升解释可信度。

Details

Motivation: 面部行为视频理解需要回答表情何时发生、模型在当时看哪里，以及该证据为何必要而非仅相关；现有 MLLM 虽能描述和分类，却较少为随时间变化的细微面部动作提供忠实解释。

Result: ExplainFace 从 CelebV-HQ 自动构造弱监督视频—文本语料，经人工评估和双模型审计，矛盾率不超过 2.3%，并在两种骨干上带来一致增益。H-TCC 依次定位关键时间点、面部区域，并通过定向遮挡检验因果必要性。在该数据上微调后，外观 Macro-F1 从 19.48% 提升至 32.87%（+13.39）；动作提高 4.38，时间情绪提高 6.77，深伪检测 F1 提高 11.84。作者还评估解释指向提升、面部/边缘提升、稳定性、Adebayo sanity check 和必要性，而不只展示案例。

Insight: 把解释的时间、空间和必要性拆开测量，比仅提供一段理由更能检查模型是否依赖真正相关的面部证据。数据为弱监督自动构造，矛盾率审计有上限；摘要没有提供各解释指标的具体值和临床/情绪标签泛化结果。

Abstract: Understanding facial behaviour in video requires answering three fundamental questions: when a key facial expression occurs, where the model attends on the face at that moment, and why that evidence is necessary rather than merely correlated. Today’s multimodal large language models (MLLMs) can label and describe video, but they rarely provide faithful explanations, especially for small facial movements that change over time. We present VideoFace-X-LLM, a framework for explainable facial video understanding comprising two key components. First, we introduce ExplainFace, a weakly supervised video-text corpus for facial behaviour understanding, automatically constructed from CelebV-HQ videos and validated through human evaluation and a two-model audit (at most 2.3% contradictions), demonstrating consistent gains across two backbone architectures. Second, H-TCC (Hierarchical Temporal Spatial Causal Concepts), which finds the key temporal moments (When), the facial regions (Where), and verifies its causal necessity (Why) through targeted occlusion. Fine-tuning on ExplainFace raises appearance Macro F1 from 19.48% to 32.87% (+13.39), with gains of +4.38 on actions, +6.77 on temporal emotion, and +11.84 F1 on deepfake detection. We also measure the explanations pointing lift, face/edge lift, stability, an Adebayo sanity check, and a necessity test rather than only showing examples. Our models, code, and corpus are available on request for research use.

[194] QK-Edit: Revisiting Attention-based Injection in MM-DiT for Image and Video Editing cs.CV | cs.AIPDF

Tiancheng Shen, Zilong Huang, Xiangtai Li, Zhijie Lin, Jiyang Liu

TL;DR: QK-Edit 利用 MM-DiT 的查询—键注意力特性进行免训练编辑，通过推理时特征替换在文本指令忠实度与图像/视频结构保持之间折衷。

Details

Motivation: MM-DiT 将图像和文本 token 拼接做多模态自注意力，传统 U-Net 注意力操控编辑法难以直接迁移到这种结构。

Result: QK-Edit 提出查询—键操控策略，分离并调节关键注意力成分，可执行物体添加、删除、替换、背景修改、材质/颜色变化和风格转换；实现上可在推理时用特征替换，无需训练。作者在 FLUX、HunyuanVideo 等模型上报告编辑表现优于当时先进方案。摘要未提供基准、数值或不同编辑类型的独立结果。

Insight: 针对 MM-DiT 的注意力交互设计轻量干预，利用已有生成模型而不重新训练，旨在弥补生成能力和可编辑性之间的落差。摘要没有量化结构保持与提示遵循的权衡，也未说明干预对不同采样器及分辨率的影响。

Abstract: Multimodal Diffusion Transformers (MM-DiTs) have recently emerged as a powerful framework for unified text-vision synthesis, surpassing traditional U-Net architectures in generative tasks. One key innovation lies in its Multimodal Self-Attention (MM-SA) interaction, where image and text tokens are concatenated and processed via self-attention. However, this mechanism poses significant challenges for editing, rendering conventional U-Net-based attention manipulation methods ineffective. To address this limitation, we propose QK-Edit, a training-free framework that exploits the unique attention dynamics of MM-DiTs for precise text-guided image and video editing. By introducing a novel query-key manipulation strategy, our method isolates and adjusts critical attention components to achieve an optimal balance between prompt fidelity and structural consistency. This enables seamless edits across various tasks, including object addition, object removal, object replacement, changing background, changing material, changing color, and style transformation. Notably, it can be easily implemented with feature replacement in inference. QK-Edit demonstrates superior editing performance on state-of-the-art models, such as FLUX and HunyuanVideo, effectively bridging the gap between generative power and editable flexibility in MM-DiTs, and paving the way for scalable multimodal content creation.

[195] DVTR-KD: Distilling Vision-Language Models via Intermediate Visual Token Reconstruction cs.CVPDF

Haolan Wei, Jiahui Wang, Junli Zha, Xinkai Lu, Yunxiao Zou

TL;DR: DVTR-KD 让视觉 token 压缩后的学生重建教师的稠密视觉状态，并在关键中间层以教师响应及注意力读出监督，避免直接逐 token 对齐。

Details

Motivation: 知识蒸馏可压缩 MLLM，但视觉 token 压缩导致师生 token 数不一致，增加直接逐 token 迁移知识的难度。

Result: 框架在 LLM 最后一层及一个对视觉信息读出重要的中间层重建教师视觉状态。中间层监督基于教师响应预测，而非直接进行 token 级 KL 或 MSE；辅助注意力读出损失对齐学生与原教师整合上下文以回答问题的方式。多基准实验报告 DVTR-KD 提升压缩学生性能，同时保留视觉 token 压缩带来的推理效率。摘要未提供模型、基准、准确率或压缩比例。

Insight: 对齐重建目标可绕过师生视觉 token 数量不一致，并用响应级监督约束学生最终如何使用视觉信息。效率与质量的具体权衡在摘要中未量化，需依正文评估重建额外训练成本及不同压缩倍率下的表现。

Abstract: Multimodal large language models (MLLMs) have achieved strong performance on diverse vision-language tasks, but their computational demands limit deployment under constrained budgets, particularly on edge devices. Many works focus on Knowledge Distillation (KD) and try to transfer knowledge from powerful teacher models to students. In addition to reducing model size, visual token compression provides a complementary route to efficiency, but introduces mismatched visual token counts that complicate direct token-wise knowledge transfer. To mitigate this, we propose DVTR-KD, a knowledge distillation framework based on \textbf{D}ense \textbf{V}isual \textbf{T}oken \textbf{R}econstruction, which reconstructs the teacher’s dense visual tokens from the student’s compressed tokens. We reconstruct visual states after the final LLM layer and at the input to an intermediate layer identified as important for visual-information readout. At the intermediate layer, reconstruction is supervised through the teacher’s response predictions rather than by direct token-wise KL or MSE matching. An auxiliary attention-readout loss complements this response-level supervision by aligning how the student and the original teacher integrate contextual information when predicting responses. Evaluations on diverse benchmarks demonstrate that DVTR-KD improves the performance of token-compressed students while retaining the inference efficiency afforded by visual token compression.

[196] What Makes a Good Visual Representation for Spatial Understanding in MLLMs? cs.CVPDF

Xiaohu Huang, Yumin Guo, Kai Han

TL;DR: 该研究系统比较不同视觉表征及其迁移方式对 MLLM 空间理解的作用，并发现表征自身的 3D 结构与收益相关，潜特征和显式几何需要不同注入策略。

Details

Motivation: MLLM 视觉预训练主要来自 2D 图像，视觉 token 缺少固有 3D 结构。虽然将外部视觉或几何特征迁移进模型通常有帮助，但哪类表示更有效、应如何迁移尚不清楚。

Result: 作者将迁移因素分为“迁移什么”与“如何迁移”：来源包括 3D 几何、自监督图像/视频、视觉语言及生成模型；形式包括潜特征或深度/点图；方式包括蒸馏监督或作为输入融合。实验发现，表征的 3D 结构与空间理解增益强相关：潜特征看跨视角一致性，显式几何看深度和重建精度。潜特征融合与蒸馏均可获益；显式几何单纯融合收益较小，大部分蒸馏增益依赖几何解码头。按此设计将两类表示蒸馏至同一 MLLM，仅用 RGB 视频输入，在静态 3D 和动态 4D 空间推理基准上均稳定优于基座模型。

Insight: 不同监督表示的“信息形态”决定合适的注入机制，不能把潜特征和几何图统一当作可互换输入。作者提出的相关性分析与双表示蒸馏支持这一设计原则，但摘要未列基准、量化增益及额外训练成本。

Abstract: Spatial understanding is a necessary step in moving multimodal large language models (MLLMs) toward reasoning about the 3D physical world. Their visual pretraining data, however, is almost entirely 2D, so the visual tokens processed by MLLMs often lack inherent 3D structural information. A growing number of methods inject such structure by transferring representations from external vision or geometry models, and collectively establish that such transfer helps. Behind these methods lies a largely underexplored question: what makes a visual representation good for spatial understanding? To answer it, we first separate what is transferred from how it is transferred, and then trace the resulting differences back to properties of the representations themselves. What we transfer comes from 3D geometry models, self-supervised image and video models, vision-language models, and generative models, in the form of latent features or explicit depth and point maps. We transfer it either by distillation, using it as a supervision target, or by fusion, feeding it as an additional input. We find that a representation’s 3D structure correlates strongly with how much it helps, measured by cross-view consistency for latent features and by depth and reconstruction accuracy for explicit geometry. We further show that the form of a representation dictates the mechanism it needs. Latent features gain as much from fusion as from distillation; explicit geometry gains little from fusion, and most of its distillation gain depends on a geometry decoding head. Guided by these findings, we distill both latent features and explicit geometry into a single MLLM. Taking only RGB video as input, it consistently improves over its base model on both static 3D and dynamic 4D spatial reasoning benchmarks.

[197] Visual Dependence Does Not Imply Answerability: A Systematic Analysis of Answerability Signals in Frozen InstructBLIP cs.CV | cs.AIPDF

Minjun Kim, Hyunmin Song, Changwon Lim

TL;DR: 该研究在冻结 InstructBLIP 上系统检验视觉依赖、置信度、跨模态对齐和内部表征能否预测 VQA 问题可回答性，发现这些信号区分能力普遍很弱。

Details

Motivation: 即使问题缺少足够视觉证据，VQA 模型也可能生成看似合理答案；已有拒答基准和机制虽多，但哪些冻结 VLM 信号能在未见图文对上可靠判断可回答性仍不清楚。

Result: 作者在 UNK-VQA 分层抽样的 1 万例上考察预测不确定性、跨模态对齐、视觉 grounding/依赖、内部表示和探针。测试 AUROC 大多仅为 0.50–0.54；CLIP 对齐点估计最高，但配对 bootstrap 下未显著超过生成置信度。移除问题相关视觉证据比随机遮挡更大幅改变答案置信度，却几乎不能区分可回答与不可回答。增加探针容量、组合信号、加入理由监督或冻结后交叉验证均未带来稳健提升；低风险选择性回答下可用覆盖率有限。

Insight: 模型确实依赖图像，不代表图像证据足以回答问题；视觉依赖与证据充分性是不同性质。结论限于 UNK-VQA 子集和冻结 InstructBLIP，且单一信号不是可靠独立代理；对其他模型与主动拒答策略还需重新验证。

Abstract: Visual question answering models often produce plausible responses even when the available visual evidence is insufficient to support a valid answer. Although prior work has introduced benchmarks and abstention mechanisms for unanswerable visual questions, it remains unclear which signals derived from a frozen vision-language model reliably indicate answerability on unseen image-question pairs. We systematically evaluate answerability signals on a stratified 10K subset of UNK-VQA using frozen InstructBLIP, covering predictive uncertainty, cross-modal alignment, visual grounding, visual dependence, internal representations, and learned probes. Answerability discrimination remains weak, with test AUROC largely confined to 0.50-0.54. CLIP alignment yields the highest point estimate but does not significantly outperform generation confidence under paired bootstrap analysis. Targeted removal of question-relevant visual evidence causes substantially larger changes in answer confidence than same-size random masking, yet intervention sensitivity provides essentially no answerability discrimination. Increasing probe capacity, combining multiple signals, and adding reason-aware supervision do not provide robust improvements, while post-freeze cross-validation yields similarly weak discrimination. Selective-answering analysis shows limited usable coverage under low-risk operating criteria. Failure-mode analysis shows that the relative utility of individual signals varies across perturbation mechanisms. Taken together, these results show that, on our UNK-VQA subset with frozen InstructBLIP, visual dependence, uncertainty, semantic alignment, and static representations are not robust standalone proxies for answerability. Our findings emphasize the need to distinguish a model’s reliance on visual evidence from whether that evidence is sufficient to support an answer.

[198] LoopREG: Representation Supervision Enables Test-Time Recurrent Depth in Diffusion Transformers cs.CVPDF

Zipeng Wang, Anping Xiong, Jing Zhang, Xiaoya Zhu, Yi Jiang

TL;DR: LoopREG 通过视觉表示监督训练共享 Transformer 块的循环扩散模型，使测试时增加循环次数可扩展计算深度并改善图像生成质量。

Details

Motivation: 复用 Transformer 块可在测试时增加扩散计算，但若训练中未覆盖更深循环，朴素递归可能超出训练深度后损害生成质量。

Result: LoopREG 在循环模块最后一次访问后施加预训练视觉编码器表示对齐，推理时将同一学习算子扩展到更多次循环。ImageNet 256×256 上，训练 40 万步时 SiT-B/2 FID 从局部 REG 基线 15.08 降至 4.91（降低 67.41%）。在仅训练两次访问的检查点上，不改权重额外增加一次推理访问，FID 从 14.71 降至 5.34；SiT-L/2 和 SiT-XL/2 也保持收益。控制干预显示有效递归依赖表示监督位置，REPA 对齐的外推收益更小。

Insight: 训练时在递归状态上施加表示监督，使同一块在额外推理访问时仍处于有意义的表征区域；因此推理深度可作为测试时计算旋钮。结果针对 ImageNet 256×256 扩散模型，摘要未给采样延迟变化或更大循环次数下的质量/稳定性边界。

Abstract: Reusing Transformer blocks offers a way to increase diffusion-model computation at test time, but naive recurrence can degrade generation beyond its training depth. We introduce LoopREG, a representation-supervised recurrent diffusion Transformer that repeatedly applies a shared block at the state supervised by a pretrained visual encoder. Alignment is applied after the final recurrent visit during training; inference extends the same learned operator to additional visits. On ImageNet 256 x 256, LoopREG reduces SiT-B/2 FID from the local REG baseline’s 15.08 to 4.91 at 400K training steps, a 67.41% reduction. Within a checkpoint trained with two visits, one extra inference visit improves FID from 14.71 to 5.34 with unchanged weights. The benefit persists on SiT-L/2 and SiT-XL/2. Controlled interventions show that useful recurrence follows the location of representation supervision, while REPA alignment produces a smaller extrapolation benefit. These results make representation-supervised recurrence a practical route to improving diffusion generation through additional inference depth.

[199] MicroCLIP: Connecting Text and Microscopy Images via Contrastive Language–Image Pretraining cs.CVPDF

Yanran Ruan, Zhixiang Wei, Shiyao Hong, Xiaoxiao Ma, Yuxuan Gu

TL;DR: MicroCLIP 用约 250 万对荧光显微图像—文本构建开放词汇细胞视觉语言模型，并从类别、知识和细节三个层级建立生物语义监督。

Details

Motivation: 现有显微方法常从封闭类别学习预定义概念，跨任务和生物场景迁移受限；多数显微图像缺少描述性注释或生物知识标注，图文配对难以直接获取。

Result: 作者构建 MicroPool-2M（摘要称约 250 万对），涵盖不同细胞结构、细胞系和成像条件。三级语义关联包括：标准化“亚细胞结构—细胞系”标签的类别对齐；利用标签衍生生物先验建模物种和细胞类型关系；用 MLLM 大规模重描述并验证图文对以提供细节级监督。多个实验中，MicroCLIP 在显微图像零样本分类和图文检索上取得领先，可从数千候选中检索特定细胞特征，并能无额外监督泛化到未见生物概念。摘要未报告具体基准分数。

Insight: 多层级标签和生物先验弥补自然语言描述稀缺，自动重描述则补充细粒度视觉信息。数据集名为 MicroPool-2M 但摘要约述 2.5M 配对，规模统计口径需确认；“未见概念”泛化结论也需看类别切分和评测范围。

Abstract: Understanding fluorescence microscopy images is fundamental to automating biological research. However, existing methods typically learn predefined concepts from closed-set data, limiting transfer across tasks and biological contexts. In this work, we introduce MicroCLIP, a vision-language model that learns generalizable cellular and subcellular representations from natural language supervision for open-set image classification and retrieval. To pretrain MicroCLIP, we construct MicroPool-2M, a dataset of approximately 2.5 million fluorescence microscopy image-text pairs spanning diverse cellular structures, cell lines, and imaging conditions. Aligning microscopy images with biological concepts in natural language is challenging because most images lack descriptive captions and biological knowledge annotations. We therefore progressively establish image-text associations at three levels to provide MicroCLIP with richer semantic supervision: (1) category-level alignment links images to standardized labels defined by subcellular structure-cell-line combinations; (2) knowledge-level alignment incorporates biological priors derived from these labels to capture relationships across species and cell types; and (3) detail-level alignment uses a multimodal large language model to recaption and verify image-text pairs at scale, emphasizing fine-grained visual characteristics. Extensive experiments demonstrate that MicroCLIP achieves leading performance in zero-shot classification and text-image retrieval on microscopy images. It retrieves images with specific cellular characteristics from thousands of candidates and generalizes to unseen biological concepts without additional supervision. These results highlight MicroCLIP’s potential as a foundation model for fluorescence microscopy.

[200] DriveWLA: Unifying Future Dynamics and Scene Semantics through World-Language-Action Modeling for Autonomous Driving cs.CV | cs.AI | cs.ROPDF

Shuzheng Wang, Haoxuan Li, Shengzhen Li, Chenxi Wu, Yajing Sun

TL;DR: DriveWLA 以世界、语言、动作三个专家和跨专家注意力共同建模场景语义、未来动态和驾驶轨迹，并用 Flow-GRPO 强化学习优化动作专家。

Details

Motivation: VLM 擅长语义理解，世界模型擅长未来场景演化，但将两者有效结合用于轨迹生成仍有挑战。

Result: 基于 Mixture-of-Transformers 的 DriveWLA 分为负责未来预测的 World Expert、语义理解的 Language Expert 和轨迹生成的 Action Expert，Cross-Expert Attention 控制专家间信息流并对齐语义与动态表示；Flow-GRPO 用于模仿学习后的动作优化。NAVSIM v1 PDMS 为 92.3，NAVSIM v2 EPDMS 为 90.1。作者还报告无需专项微调即可零样本迁移到 HUGSIM、nuScenes 和 Bench2Drive，并在开环与闭环评估中表现优于基线。

Insight: 将场景演化、语言语义和动作生成分工建模，再通过受控交互连接，可共同为规划提供信息。摘要未列开环/闭环其他指标及各迁移数据集分数，NAVSIM 结果也应限于作者所比较的规划器。

Abstract: Vision-Language Models (VLMs) and world models have demonstrated strong capabilities in semantic understanding and future scene evolution for autonomous driving. However, effectively integrating both capabilities for trajectory generation remains challenging. In this work, we propose DriveWLA, a unified World-Language-Action model based on Mixture-of-Transformers (MoT) that jointly models semantic understanding, future scene dynamics, and trajectory generation. DriveWLA consists of three experts: a World Expert for future prediction, a Language Expert for semantic understanding, and an Action Expert for trajectory generation. We further introduce Cross-Expert Attention, which enables controlled information flow across experts while aligning semantic and dynamic representations. To improve trajectory generation beyond imitation learning, we apply Flow-GRPO reinforcement learning post-training to refine the Action Expert. Extensive experiments show that DriveWLA achieves superior performance in both open-loop and closed-loop evaluation. DriveWLA achieves 92.3 PDMS on NAVSIM v1 and 90.1 EPDMS on NAVSIM v2, demonstrating strong performance among VLM-based and world-model-based planners. Without specific fine-tuning, DriveWLA further demonstrates strong zero-shot generalization to HUGSIM, nuScenes, and Bench2Drive. These results validate the effectiveness of our unified World-Language-Action modeling, establishing DriveWLA as a generalizable framework for autonomous driving.

[201] CineMem: Attention-Aligned Fine-grained Memory Retrieval for Multi-Shot Long Video Generation cs.CV | cs.AIPDF

Xiaoxue Wu, Xinyu Gao, Fang Yu, Yu Qiao, Xinyuan Chen

TL;DR: CineMem 根据注意力访问模式在线组织历史 KV 为语义簇，并按查询 token 自适应检索，使多镜头视频生成能在固定记忆预算下精细读取相关历史。

Details

Motivation: 长视频自回归生成依赖历史记忆，尤其多镜头切换时既要适应画面大幅变化又要保留语义连续性；按帧检索并让所有查询共享同一历史上下文，会忽视不同 token 的信息需求并造成冗余。

Result: CineMem 以在线聚类 KV 缓存管理历史 token，使语义相关记忆分组；查询级检索在固定预算下提供更广的历史访问。作者还提出 VLM 辅助的任务专用指标评估镜头间一致性。在跨镜头重访场景中，方法报告能维持长程一致性，并在受限内存下近似完整记忆访问；实验总体支持框架有效。摘要未给出数据集、检索预算或质量指标数值。

Insight: 按语义访问模式组织缓存并细化到查询级检索，针对同帧/同上下文对所有 token 一视同仁造成的记忆浪费。结果主要定性描述，且评测依赖 VLM 辅助指标；摘要未报告人工一致性或检索延迟。

Abstract: Autoregressive video generation relies on memory retrieval to model long-range dependencies. This capability becomes particularly critical for multi-shot generation, where models must reconcile substantial visual changes across shot transitions with persistent semantic continuity. Existing methods tend to perform memory retrieval at the frame level and share the same historical context across all query tokens, overlooking the heterogeneous information needs and leading to redundant retrieval and inefficient use of historical memory. To bridge this gap, we introduce CineMem, an attention-aligned fine-grained memory retrieval framework that combines online memory organization with query-adaptive retrieval for autoregressive multi-shot video generation. Specifically, guided by intrinsic memory-access patterns revealed through attention analysis, we propose an online-clustered KV cache as a memory organization strategy over all historical tokens, enabling semantic-aware memory grouping and more precise retrieval. In addition, we introduce a query-level retrieval mechanism that provides broader access to historical information under a fixed retrieval budget, thereby improving long-term memory utilization. Together, these designs enable the multi-shot video generator to preserve long-range consistency across shot transitions, particularly in revisitation scenarios, while approximating full-memory access under a constrained memory budget. To facilitate evaluation, we further introduce a suite of task-specific, VLM-assisted metrics for faithful inter-shot consistency assessment. Extensive experiments demonstrate the effectiveness of our framework.

[202] LinMorph: Learning to Morph Weights for Efficient Video DiT Linearization cs.CV | cs.AIPDF

Lingyi Long, Congyue Zhao, Xiaoou Li, Yanmin Qian, Tao Huang

TL;DR: LinMorph 将视频 DiT 线性化前的权重初始化视为可学习的跨参数形变，从预训练 Q/K/V/O 投影共同构造 Gated DeltaNet 权重，再进行渐进蒸馏和联合恢复。

Details

Motivation: 后训练线性化常把注意力替换后的质量损失当作恢复问题；随机初始化丢失预训练知识，手工权重复用又依赖僵硬且不完整的源—目标对应关系。

Result: LinMorph 学习从 softmax 注意力参数到线性 GDN 参数的共享 morphism，让每个目标参数可联合利用预训练 Q、K、V、O 投影，即便没有直接对应源参数也可初始化。共享映射通过局部教师对齐学习，再针对各层预训练权重生成层特定初始化；随后物化为独立 GDN 模块进行渐进蒸馏和联合恢复。由于映射工作在权重空间，可用短低分辨率视频训练并复用于更大尺度。作者报告在多个预训练视频 DiT 上，恢复速度和最终生成质量均优于随机初始化及启发式复用；摘要未给具体模型、数值或速度倍数。

Insight: 线性化效果取决于恢复之前如何迁移预训练知识，初始化质量是关键瓶颈，而不仅是后续恢复算法。共享权重映射实现跨层复用且支持低分辨率学习，但摘要未量化训练开销、质量—速度折衷或长视频上的外推边界。

Abstract: Post-hoc linearization is an attractive way to accelerate pretrained Video Diffusion Transformers (Video DiTs), yet existing methods largely treat the quality loss induced by attention replacement as a recovery problem. We show that a more fundamental bottleneck lies upstream: \emph{initialization}. Random initialization discards pretrained knowledge, whereas manual weight reuse transfers it through rigid and incomplete source–target correspondences. We introduce \textbf{LinMorph}, which recasts Video DiT linearization as a learnable softmax-to-linear weight transformation problem. Instead of copying predefined counterparts, LinMorph learns a cross-parameter morphism that lets each target Gated DeltaNet (GDN) parameter draw jointly from pretrained query, key, value, and output projections, including for parameters with no direct softmax counterpart. A single morphism is shared across layers and learned through local teacher alignment, while producing layer-specific initializations from their respective pretrained weights. The resulting weights are then materialized as independent GDN modules for progressive distillation and joint recovery.Since the morphism operates in weight space rather than token space, it can be learned on short, low-resolution clips and directly reused at larger video scales. Across pretrained Video DiTs, LinMorph enables markedly faster recovery and better final generation quality than random initialization and heuristic weight reuse. Results show that efficient linearization depends not only on how a converted model is recovered, but critically on how pretrained knowledge is transferred before recovery. Project page: https://lonelyling.github.io/LinMorph.

[203] MERIT: Multimodal Evidence Reasoning over Image Tables cs.CV | cs.AIPDF

Haoyang Jiang, Zhuotong Fan, Mingyang Yi, Ju Fan, Lanqing Hu

TL;DR: MERIT 学习从表格图像中获取证据、逐步推理和核验答案，针对文本序列化可能破坏单元格内容、表格关系及视觉信息的问题。

Details

Motivation: 表格问答通常假设可读取干净文本表格，但将表格序列化为文本可能损坏单元格或结构并丢失视觉线索；图像输入保留这些信息，却要求模型既准确感知又基于相关单元格推理。

Result: MERIT 包含三阶段：先用重建、定位和颜色条件监督适配表格专用视觉编码器；再通过经验证的轨迹监督微调显式选择证据并推理；最后用强化学习分别奖励证据正确性、计算、最终答案和输出格式，以反馈中间步骤错误。文本序列化表格仅在可用时作为可选辅助输入。作者在表格问答和事实核验任务上评测，并分析视觉输入与训练组件贡献及对不同结构/外观数据集的泛化。摘要未列基准名称或量化结果。

Insight: 中间证据奖励能区分检索错误、计算错误和答案格式错误，弥补只监督最终答案的不足。摘要未披露实验数值，因此无法评估视觉感知提升、文本辅助收益及不同表格样式下的泛化幅度。

Abstract: Table reasoning with language models typically assumes access to clean textual tables. In practice, tables are often serialized into text, which can corrupt cell contents, distort structural relationships, and discard visual cues, leaving the model to reason over incomplete evidence. In contrast, table images retain this information, but using it requires both accurate perception and reasoning grounded in the relevant cells. To address this, we propose MERIT (Multimodal Evidence Reasoning over Image Tables), a three-stage framework for learning to reason over table images, with serialized tables serving as optional auxiliary inputs (if provided). Our analysis identifies visual evidence acquisition as a bottleneck, motivating table-specific encoder adaptation through reconstruction, localization, and color-conditioned supervision. Building on the adapted representations, supervised fine-tuning teaches explicit evidence selection and reasoning using verified trajectories. Finally, reinforcement learning refines the policy with separate rewards for evidence correctness, calculation, final answers, and output format. By providing feedback on intermediate steps, the framework targets retrieval and computational errors that final-answer supervision alone does not distinguish. We evaluate the approach on table question answering and fact verification, analyze the contributions of visual inputs and training components, and assess generalization to additional datasets with different table structures and visual presentations.

[204] Dynamic Interpolation for Training-Time Prediction Reshaping in Low-Light Image Enhancement cs.CVPDF

Peirui Zhu, Shijie Hao, Jingxi Liao, Richang Hong, Meng Wang

TL;DR: DITM 在低照增强训练期间按目标均值亮度动态插值模型预测，使全局亮度误差收敛后，局部噪声和结构伪影在反向传播中获得更大权重。

Details

Motivation: 低照增强需同时处理照度不足、噪声放大和结构退化；复杂网络会增加部署负担，人工设定多项损失权重又依赖经验。作者观察亮度误差通常先收敛，残余噪声和结构问题在后期优化不足。

Result: 训练期模块将预测与按真值均值亮度缩放得到的亮度对齐图像插值，降低全局亮度差异的相对主导性，从而让局部残差更影响后期梯度。DITM 没有可训练参数，只在训练中使用，推理时完全移除，不增加推理开销。多基准实验覆盖 CNN、Transformer、Mamba 骨干，作者报告保真度与感知质量持续改善，并保持原推理架构。摘要未列具体指标或增益。

Insight: 利用训练动态重新分配误差信号，不必增加推理模块或手工调整额外损失权重，重点优化亮度收敛后的局部质量。摘要没有给不同插值策略、基准增益和失败案例数值，需查看正文验证各骨干与数据集的一致性。

Abstract: Low-light image enhancement (LLIE) is an important image/video processing problem for reliable visual systems under poor illumination, yet it remains challenging due to the strong coupling among insufficient illumination, noise amplification, and structural degradation. Existing methods typically rely on sophisticated network architectures or engineered objective functions. While effective, the former increases model complexity and deployment burden, whereas the latter relies on empirically tuned loss weights. We revisit supervised LLIE from a training-dynamics perspective and observe that brightness-related errors tend to converge early, while residual noise and structural artifacts often remain insufficiently optimized in later stages. To address this issue, we propose a Dynamic Interpolation Training-only Module (DITM), which introduces no additional learnable parameters and adaptively reshapes the model prediction for supervised feed-forward LLIE training. Specifically, DITM employs brightness-aware interpolation, in which the model prediction is combined with a mean brightness-aligned image constructed by rescaling the model prediction according to the mean brightness of the ground-truth image. By reducing the relative dominance of global brightness discrepancy, DITM makes residual local errors more influential during backpropagation after global luminance has largely stabilized, thereby facilitating artifact suppression and fine-detail restoration in later training stages. DITM is used only during training and is fully removed at inference, introducing no additional inference cost. Extensive experiments on multiple benchmark datasets and CNN-, Transformer-, and Mamba-based LLIE backbones show that DITM consistently improves fidelity and perceptual quality while preserving the inference-time architecture and introducing no additional inference-time computation.

[205] PolUR: Polarization Aware Multimodal Degradation Perception for Underwater Image Restoration cs.CVPDF

Wenshuo Zhang, Xiangyue Zhang, Junlin Li, Wenjie Tian, Yubo Zheng

TL;DR: PolUR 联合 RGB 和偏振观测判断水下图像混合退化类型及严重程度，再用退化语义特征引导图像复原，并构建含真实退化图像与描述的 MixPolar 基准。

Details

Motivation: 偏振观测在清水条件下可提供目标辐射恢复所需的互补物理线索，但现有方法尚未联合建模原始强度与偏振信息，以理解色偏、雾化和低照等混合退化。

Result: PolUR 采用多模态、语言驱动框架，通过结构化语义指令估计退化类型与严重度，并以退化感知语义特征增强联合复原表示。MixPolar 收录具有真实世界退化的水下图像及对应退化描述，支持退化感知、图像复原和下游任务评估。作者报告在退化感知和复原上取得有竞争力的结果，并展示泛化能力及下游实用性。摘要未列数据规模、基准名称或数值。

Insight: 退化识别被显式用于指导恢复，可以让模型针对混合退化调整复原策略；偏振提供物理线索，RGB 保留常规强度信息。摘要主要定性报告结果，未说明偏振采集条件、不同水体迁移性能或与 RGB-only 方法的量化差距。

Abstract: Recovering high-quality images from degraded underwater observations is crucial for underwater exploration. Polarization observations show promise as an effective source of complementary physical cues for recovering target radiance under clear-water conditions. However, existing methods have yet to jointly model raw intensity information and polarization cues to enable image restoration driven by an understanding of mixed degradations (color casts, haze, and low light). To overcome these limitations, we propose PolUR, a unified multimodal language-driven framework that jointly uses polarization and RGB information to identify degradation types and severity levels through structured semantic instructions and enhances joint restoration representations under the guidance of degradation-aware semantic features to improve restoration robustness under mixed degradations. We further construct a new benchmark named MixPolar, which contains underwater images exhibiting real-world degradations and their corresponding degradation descriptions. MixPolar supports systematic evaluation of degradation perception, image restoration, and performance on downstream tasks. Extensive experiments demonstrate that PolUR effectively integrates polarization and RGB information, achieving competitive performance in degradation perception and image restoration, while also demonstrating generalization capability and utility for downstream tasks.

[206] COMPASS: Contextual Moment Prior Alignment for Generalizable AIGI Detection cs.CVPDF

Zhan Yang, Tiancheng Shen, Weijie Lyu, Shuo Chen, Bernard Ghanem

TL;DR: COMPASS 用无标注语义代理集估计目标相关激活统计，并在权重子空间分解时保留相关表征，使伪造检测适配更关注目标语义而非仅依赖参数幅度。

Details

Motivation: 少量训练分布上的适配可能损害视觉基础模型的可迁移表征；已有子空间方法冻结高幅度主成分、适配残差，但该分解与目标数据无关，可能造成语义错位。

Result: COMPASS 用小型未标注 Semantic Proxy Set 估计目标激活统计，并将协方差用于预训练权重分解：与目标相关的表征作为稳定语义参照保留，残差适配器学习伪造变化。代理校准不需基于梯度优化。作者在跨数据集、跨方法、联合跨数据集/方法和开放世界跨生成器评测中，报告其在 deepfake 和 AI 生成图像检测基准上达到当时最佳表现。摘要未列具体基准名或数值。

Insight: 权重子空间适配不仅要看参数幅度，还应参考目标域激活统计，以减少对目标语义有用的方向被错误修改。结果在多种迁移设置中定性领先，但摘要未量化代理集大小、不同伪造域增益及校准开销。

Abstract: Generalizable AI-generated image (AIGI) and deepfake detection remain challenging due to the asymmetry phenomenon, where adaptation to limited training distributions can degrade the transferable representations of Vision Foundation Models (VFMs). Recent subspace adaptation methods mitigate this issue by freezing high-magnitude principal components while adapting the residual subspace. However, such magnitude-based decomposition is data-agnostic and does not explicitly account for target-relevant semantics, potentially leading to semantic misalignment. We propose COMPASS (COntextual Moment Prior Alignment for Subspace Separation), a data-aware subspace adaptation framework that uses a small unlabeled Semantic Proxy Set to estimate target activation statistics. COMPASS incorporates activation covariance into the decomposition of pre-trained weights, preserving target-relevant representations as a stable semantic reference while the residual adapter learns forgery-related variations. The proxy calibration requires no gradient-based optimization. Extensive experiments demonstrate strong generalization across cross-dataset, cross-method, joint cross-dataset-method, and open-world cross-generator evaluations, achieving state-of-the-art performance on both Deepfake and AIGI detection benchmarks.

[207] Scaling Vision Transformers Without Training: Prune-and-Merge Token Compression cs.CV | cs.AIPDF

Vatsal, Soumil Malhotra, Priyanshu Chahal, Sanjeev Kumar

TL;DR: Prune-and-Merge ViT 将类别注意力剪枝与相似度合并结合，在不训练的情况下减少视觉 token，并用相对准确率下降和预测翻转率补充鲁棒性评估。

Details

Motivation: ViT 的 token 两两交互带来较高计算成本，免训练 token 缩减方法虽能提效，但对图像变换下鲁棒性与预测稳定性的影响研究不足。

Result: 方法用 class attention 剪除低价值 token、用特征相似度合并冗余 token，目标是在降低计算量的同时保留语义信息。作者提出 Relative Accuracy Drop（RAD）衡量变换下相对准确率退化，以及 Prediction Flip Rate（PFR）衡量预测一致性。在预训练 ViT、DeiT 上，作者报告方法以有竞争力的计算缩减保持鲁棒性，并比既有 token 缩减方法准确率更高；分析称混合缩减未引入类似幻觉的行为，常见图像变换下预测稳定。摘要未报告模型、数据集或数值。

Insight: 结合剪除与合并可同时处理低价值和冗余 token；用预测翻转率等指标能发现单一准确率无法反映的稳定性问题。鲁棒性结论仅针对摘要所述常见图像变换，未量化压缩率与性能折衷。

Abstract: Vision Transformers achieve strong performance but incur high computational costs from quadratic token interactions. Training-free token reduction methods improve efficiency, but their robustness remains underexplored. We propose Prune-and Merge ViT, a training-free framework for off-the-shelf models. It combines class attention-based pruning with similarity-based merging to reduce redundancy while preserving semantic information. To evaluate robustness we introduce two metrics. Relative Accuracy Drop (RAD) measures relative accuracy degradation under transformations, and Prediction Flip Rate (PFR) quantifies prediction consistency. On pretrained ViT and DeiT, Prune-and-Merge reduces computation competitively while maintaining robustness and better accuracy than existing token reduction methods. Our analysis suggests that hybrid token reduction does not introduce hallucination-like behavior and preserves prediction stability for common image transformations.

[208] LensErase: A Foundation Model for Erase Anything cs.CVPDF

Yunuo Chen, Xiuyu Wu, Qinhong Yang, Jinjing Zhao, Yanchen Dong

TL;DR: LensErase 用超过 1.01 亿组图像—掩码—目标配对训练通用掩码擦除基础模型，并通过 mask-aware 位置编码和区域奖励改善多类真实擦除任务。

Details

Motivation: 真实擦除任务覆盖物体、文字、水印、相关视觉效果及不同尺度/分辨率；通用生成先验需要有效结合源图与掩码，并在擦除目标和保护未编辑区域之间平衡。

Result: LensErase-100M 含 101.07M 源图—掩码—目标样本；模型从文生图模型 Lens 预训练，提出 3D Mask-Aware RoPE 将掩码信息注入源图 token 而不另建掩码序列。后续以监督微调和契约引导的区域感知奖励强化学习优化，并在推理时用 TraceFusion 保留有效编辑区之外的内容。在六个公开基准及自建 LensErase-Bench 上，作者报告参考型客观指标和无参考 VLM 评估均有显著提升，但摘要未给具体数值。

Insight: 大规模配对目标、位置编码和区域奖励共同覆盖“擦除目标”和“维持其余内容”两类需求；推理时融合进一步强调保真。摘要未列数据去重/来源细节、各擦除类别表现及外部 VLM 评估偏差，因此不能量化不同模块收益。

Abstract: We introduce LensErase, a general-purpose mask-based Erase Anything foundation model designed to cover a broad spectrum of real-world erase scenarios. LensErase supports diverse objects, visual text, watermarks, associated visual effects, multiple-object removal, and a wide range of object scales, aspect ratios, and image resolutions. First, to enable such broad coverage, we construct LensErase-100M, a large-scale dataset containing 101.07 million paired source-mask-target examples spanning diverse erase scenarios. Second, we pretrain LensErase from the text-to-image model Lens, adapting its strong generative prior to mask-based erasing. To efficiently incorporate the source image and erase mask, we propose Mask-Aware RoPE, a 3D rotary positional encoding that directly injects mask information into the source-image tokens without introducing a separate mask token sequence. Finally, we refine the model through supervised fine-tuning and reinforcement learning with contract-guided, region-aware rewards, and apply TraceFusion at inference to preserve high-fidelity content outside the effective edit region. Across six public benchmarks and the proposed LensErase-Bench, LensErase achieves substantial improvements under both reference-based objective metrics and reference-free VLM evaluation. We hope LensErase can serve as a foundational model for erase-related applications and systems.

[209] APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation cs.CV | cs.AIPDF

Sirui Zhang, Fangyun Wei, Dong Chen, Jinjing Zhao, Yitong Wang

TL;DR: APT 用三个智能体自动构造评测集、诊断文生图模型的特定能力缺口并生成偏向弱项的强化学习数据，形成模型行为驱动的后训练循环。

Details

Motivation: 文生图后训练通常依赖人工整理训练数据和固定分布，难以随模型能力变化自动调整训练重点。

Result: APT 协调评测集构建、模型优缺点诊断和强化学习数据生成三个角色，并根据评估反馈动态改变提示分布以强调不足能力；模型反复评估、诊断和更新，使当前生成器表现影响下一轮训练任务。无用户指定场景时另提供通用评测 harness，以生成多样评测集。作者在多个基座模型、基准和生成架构上报告框架具有模型、基准、架构无关性并能全自动运行。摘要未列性能数值或自动生成数据质量指标。

Insight: 反馈驱动的数据分布可减少固定训练集与不断变化模型能力之间的错位；自动评测也试图支持更广泛场景。摘要未量化能力增益、评测集噪声或闭环稳定性，因此“通用”主要是作者的适配范围结论，需结合具体实验判断。

Abstract: Post-training is crucial for improving text-to-image models, yet existing pipelines typically rely on manually curated training data and fixed training distributions. We introduce APT (Agentic Post-Training), a fully automated framework that adaptively post-trains a base model according to its evolving capability profile. APT coordinates three agents to construct evaluation sets from user requirements, diagnose model-specific strengths and failure modes, and generate reinforcement learning data with an adaptive prompt distribution that emphasizes underperforming capabilities. By repeatedly evaluating, diagnosing, and updating the generator, APT lets the current model’s behavior shape its next training tasks, forming a feedback loop inspired by recursive self-improvement. For general-purpose settings without user-specified evaluation scenarios, we further develop an evaluation harness that enables broad and diverse evaluation set generation. Experiments across multiple base models, benchmarks, and generation architectures demonstrate that APT is model-agnostic, benchmark-agnostic, architecture-agnostic, and fully automated.

[210] Benchmarking Multimodal Large Language Models for Zero-Shot Human-Object Interaction cs.CVPDF

Oriol Rabasseda, Zenjie Li, Kamal Nasrollahi, Sergio Escalera

TL;DR: 该研究系统评测 MLLM 对人—物交互的零样本识别与检测，并提出将词元概率转为交互分数的公平比较协议和两种视觉提示策略。

Details

Motivation: 传统 HOI 模型依赖大量标注和固定交互分类体系，MLLM 的开放推理可能适合动态场景，但其零样本 HOI 能力尚未系统评估。

Result: 框架使用外部目标检测器定位实体，再由 MLLM 零样本推理交互；构建统一零样本提示策略分类法，并把 token logits 转成交互分数，以便按标准指标与专用模型比较。作者还为标准 HOI 数据集新增问答标注，并提出 Objects as Context 与 Black Other Objects 两种保留场景上下文、突出目标物体的视觉提示。在 HICO、HICO-DET、V-COCO 上评测，HICO mAP 为 53.50，HICO-DET 为 23.69，作者称达到零样本最先进水平；摘要未给 V-COCO 数值。

Insight: 外部检测器承担实体定位、MLLM 负责交互理解，是把开放词汇推理纳入标准 HOI 评估的一种拆分方式。结果为零样本设定，受物体检测器质量、提示策略和新标注影响；HICO 与 HICO-DET 的指标不能直接代表所有开放世界交互场景。

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated strong zero-shot reasoning across diverse domains, but their potential for Human-Object Interaction (HOI) remains unexplored. Unlike traditional HOI approaches that rely on extensive labeled datasets and fixed interaction taxonomies, MLLMs offer flexible reasoning, making them well-suited for dynamic, real-world scenarios. In this work, we present the first systematic study of MLLMs for HOI recognition and detection within a pipeline that combines an external object detector for entity localization with zero-shot multimodal reasoning for interaction understanding. We introduce a benchmark framework for principled zero-shot evaluation, consisting of (i) a unified taxonomy of zero-shot prompting strategies, and (ii) a methodology that transforms token-level logits into interaction scores, enabling fair comparison with specialized HOI models under standard metrics. To support prompt-based evaluation, we additionally construct new question-answer annotations for standard HOI datasets, transforming them into a format suitable for MLLMs. Building on this, we propose two visual prompting strategies, namely Objects as Context and Black Other Objects, that guide MLLMs to target objects while preserving scene context. Experiments on HICO, HICO-DET, and V-COCO demonstrate state-of-the-art zero-shot performance, achieving 53.50 mAP on HICO and 23.69 mAP on HICO-DET. Overall, our work provides a principled benchmarking framework for MLLMs in HOI, opening the door to new models and datasets and enabling flexible open-world interaction understanding.

[211] Human-Inspired Asymmetric Collaboration for Dual-Student Semi-Supervised Medical Image Segmentation cs.CVPDF

Huiyi Ren, Ziming Fang, Jiaming Cai, Rui Xu, Haojie Li

TL;DR: HAC 通过强化学习动态调节双学生半监督分割的强增强难度，并用 Gram 矩阵几何约束保持学生差异，避免增强不足同质化或过强破坏解剖证据。

Details

Motivation: 双学生半监督医学分割需用强弱增强形成互补预测，但强增强不足会让学生趋同，过强则可能破坏解剖信息并使协作不稳定。

Result: RDC 将强增强控制作为窗口级强化学习过程，根据增强历史、实际扰动和训练反应动态调整增强幅度、操作数量及取值范围。G2PC 通过全局体积和成对面积约束保留早期三分支几何关系，防止共享监督造成学生同质化。在 LA、Pancreas 和 BraTS-2019 上，作者报告 HAC 在分割效果和按模型能力校准增强难度方面均有效。摘要未给出 Dice 等数值或基线差距。

Insight: 强增强不应固定设定，而可依据模型当前能力闭环调整；几何约束则保留双学生的互补性。结果来自三个医学分割数据集，摘要未量化增强策略对不同器官、标签比例或训练稳定性的贡献。

Abstract: Dual-student learning enriches predictive diversity in SSMIS by introducing a strongly augmented view alongside a weakly augmented view. Insufficient strong augmentation may homogenize the students, whereas excessive augmentation can corrupt anatomical evidence and destabilize collaboration. This challenge echoes feedback-regulated human instruction, where learners receive differentiated tasks and the task difficulty is continually adjusted according to their evolving responses. Inspired by these, we propose Human-Inspired Asymmetric Collaboration (HAC), a closed-loop dual-student framework that maintains weak–strong learning through adaptive difficulty calibration and geometry preservation. Specifically, Reinforcement-guided Difficulty Calibration (RDC) formulates strong-augmentation control as a window-level reinforcement learning process, using augmentation history, realized perturbations, and training responses to dynamically adjust the augmentation magnitude, operation number, and magnitude range. Meanwhile, the Gramian Geometry Preservation Constraint (G2PC) preserves the early-stage three-branch geometry through global volume and pairwise area constraints, preventing shared supervision from homogenizing the two students. Comprehensive evaluations across LA, Pancreas, and BraTS-2019 validate both the segmentation effectiveness and capability-aware difficulty calibration of HAC.

[212] H-DDCM: Hadamard Codebooks for Fast Generative Image and Video Compression cs.CV | cs.AI | cs.LGPDF

Or Ronai, Vladimir Kulikov, Tomer Michaeli

TL;DR: H-DDCM 用隐式随机 Hadamard 字典替代稠密高斯码本，并提出无查表字典序编码，使免训练生成式图像/视频压缩能扩展至更高维潜空间。

Details

Motivation: 扩散或 flow 模型先验可在不重训专用编解码器的情况下实现感知压缩，但稠密高斯码本在现代图像/视频生成器的大潜维度下内存、计算和组合编码代价高。

Result: 快速 Walsh–Hadamard 变换隐式生成全维随机字典，投影和重建计算复杂度为 O(d log d)、内存 O(d)，无需显式物化字典；另用无查表字典序编码降低组合码流成本。方法扩展到自回归视频 flow 模型，并以流式方式定期把近期重建帧重新编码为新的 KV-cache 前缀，维持分段间时间上下文且不增加传输比特。作者称单 GPU 上可做长视频近实时生成式压缩，在图像和视频基准上实现有竞争力的率—失真—感知质量，同时显著降低码本类免训练压缩开销。摘要未给出具体比特率、延迟或质量指标。

Insight: Hadamard 结构用快速变换避免存储高维稠密码本；视频分段重新编码 KV 前缀则将时序记忆纳入压缩流。效率与质量目前以定性结论呈现，尚不能据摘要比较不同码率、显存和画质折衷。

Abstract: Training-free diffusion-based compression methods repurpose the generative priors of a diffusion or a flow model to compress data without retraining them as task-specific codecs. Recent codebook-based approaches such as Turbo-DDCM achieve strong perceptual quality at ultra-low bitrates, but their reliance on dense Gaussian codebooks remains costly in both memory and computation, and scales poorly with the large latent dimensions of modern image and video generators. Their combinatorial bitstream encoding further becomes computationally heavy at this scale. We introduce H-DDCM, a scalable generative compression framework that replaces Gaussian codebooks with implicit randomized Hadamard dictionaries. The fast Walsh-Hadamard transform allows for implicitly synthesizing full-dimensional dictionaries with $O(d \log d)$ compute for projection and reconstruction and $O(d)$ memory, without materializing the full-dimensional dictionary. We further introduce an efficient table-free lexicographic coding scheme that makes combinatorial encoding practical at larger dimensions. Beyond images, we adapt the framework to an autoregressive video flow model and develop a streaming compression scheme that periodically re-encodes recently reconstructed frames into a fresh KV-cache prefix, preserving temporal context across segments at no additional transmitted bits. This enables long-form, close to real-time generative video compression on a single GPU. Across image and video benchmarks, H-DDCM achieves competitive rate-distortion-perception performance while substantially reducing the memory and computational overhead of codebook-based training-free generative compression.

[213] FlowCred: Hierarchical Credit Assignment and Functional Coordination for Joint Text-to-Video Post-Training cs.CVPDF

Mingxuan Zhu, Yanwei Huang, Shiyuan Liu, Xinzhi Zhang, Shujie Li

TL;DR: FlowCred 通过共享随机种子的提示—生成 rollout 网格区分提示策略信用与生成器残差，再用真实生成轨迹中的替代提示指导二者协同优化。

Details

Motivation: 联合训练提示策略和视频生成器时，共用同一终局奖励会混淆提示质量与随机生成影响，且两个模块可能重复承担相同局部改进。

Result: 两级框架先用匹配的提示—种子 rollout 网格，在共享随机性下分离提示层信用和提示条件生成器残差；功能协调阶段则在实际生成状态探索当前策略的替代提示，为提示优化提供生成器相关证据，并减弱已可由条件变化解释的重复生成器更新。在多个文生视频基准上，作者报告其稳定优于匹配联合训练基线。LTX-2.5 上，T2V-CompBench 从 61.47 提至 64.46，StoryEval 从 55.58 提至 58.08，解码视频和奖励模型调用预算均未增加。因子实验、实际更新分析、人工评估及跨骨干/更新器迁移也支持方法。

Insight: 共享随机数和分层信用分配旨在把提示贡献与生成随机性分开，避免两个模块重复优化同一效果；报告改进不增加视频解码或奖励调用预算。摘要未说明其他基准的增益数值、额外提示评估计算和人评规模。

Abstract: Prompt optimization and generator post-training are two complementary ways to improve text-to-video (T2V) generation, and recent systems increasingly train both modules jointly. However, optimizing a prompt policy and a video generator from the same terminal reward creates two ambiguities: prompt quality is confounded with stochastic generation, and the two modules may realize overlapping local improvements. We introduce FlowCred, a two-level framework for joint T2V post-training. First, hierarchical credit assignment uses matched prompt-seed rollout grids to separate prompt-level credit from prompt-conditioned generator residuals under shared randomness. Second, functional coordination probes alternative current-policy prompts at states from actual generation trajectories, providing generator-aware evidence for prompt optimization while selectively attenuating generator updates already supported by conditioning changes. Across multiple T2V benchmarks, FlowCred consistently improves matched joint-training baselines. On LTX-2.5, FlowCred improves a strictly matched joint-training control from 61.47 to 64.46 on T2V-CompBench and from 55.58 to 58.08 on StoryEval, without increasing decoded-video or reward-model-call budgets. Factorial and realized-update analyses show complementary contributions from prompt shaping and generator routing, while human evaluation and transfer across backbones and updater families provide further supporting evidence.

[214] Heading Hierarchy Reconstruction for Complex and Long Documents with Level-Wise Multimodal Memory cs.CV | cs.AIPDF

Qian Chen, Xianyin Zhang, Lifan Guo, Chi Zhang

TL;DR: 该 VLM 文档解析框架用逐层级多模态记忆渐进重建长文档标题树，使模型可在有限上下文中参考此前各级标题的视觉、文本和版式模式。

Details

Motivation: 页面级 VLM 能提取文字并给标题加层级标记，但逐页预测缺少全篇一致性；数十至数百页文档又难以一次输入所有页面图像。

Result: 框架增量维护历史页面的紧凑子集，通过锚点选择记忆；本文每个层级保留最近节点，让 VLM 同时观察各级标题的图像、文本和布局特征。作者构建 CLDoc-HHRBench，含 250 份文档、11,212 页及专家核验标注，并在该基准和三个既有基准上报告方法有效。摘要未给出具体层级准确率或基线差异。

Insight: 层级记忆把长文档结构转成逐页可处理的问题，较全量上下文更适合长文档；最近节点策略则提供层级模式参照。摘要未说明锚点选择的敏感度、记忆长度与延迟，也未量化长文档收益。

Abstract: Vision-language models (VLMs) have been widely adopted for document parsing and optical character recognition (OCR), as they can accurately extract page-level elements. However, downstream applications such as retrieval-augmented generation (RAG) also require coherent document-level structure. Although these models indicate heading levels using different numbers of preceding ``#‘’ symbols, page-by-page prediction often lacks document-level consistency. In practice, input-resolution and context-length constraints make it infeasible for a VLM to jointly process all page images in documents containing tens or hundreds of pages. To address this limitation, we propose a VLM-based framework that incrementally reconstructs heading hierarchies with a \emph{level-wise multimodal memory}. The memory retains a compact subset of all previously processed pages, selected by anchors. Our framework supports flexible anchor selection, and in this work we select the most recent node at each level, allowing the VLM to observe the visual, textual, and layout patterns of headings at every level. For comprehensive evaluation, we also introduce the Complex and Long Document Heading Hierarchy Reconstruction Benchmark (CLDoc-HHRBench), comprising 250 documents and 11,212 pages with expert-verified annotations. Experiments on CLDoc-HHRBench and three existing benchmarks demonstrate the effectiveness of our method.

[215] VARI: Robust Text-to-Video Generation via Group-Structured Intent Alignment cs.CVPDF

Mingxuan Zhu, Hohei Chan, Shiyuan Liu, Yuanxing Zhang, Lu Chen

TL;DR: VARI 将同一视觉意图的多种等价提示视为一组监督，通过跨提示视频核验、组一致蒸馏和形式对比偏好优化，提升文生视频对措辞变化的鲁棒性。

Details

Motivation: 意义等价的提示改写可能使视频模型生成不同语义内容或质量；逐条提示优化不能显式约束同一意图在不同表达下保持一致。

Result: VARI 建立匹配的提示—随机种子 rollout 池，对每个视频与组内各提示交叉评分，并用措辞无关的意图审计判定有效实现和语义失败。GCD 将每种已验证表达与可互换的有效生成对齐；FC-DPO 使用审计后的正负样本，优先比较相同种子、不同措辞。HunyuanVideo-1.5 在所评方法中取得最高 T2V-CompBench、VBench、StoryEval 分数。在提示扩展版 T2V-CompBench/VBench 上，最弱措辞对齐分分别提高 2.45、3.55 分，跨措辞性能跌幅降低 30.5%、32.2%，且视觉质量保持；无需额外推理模块。人工评估、自然描述控制、提示增强、生成多样性及迁移到 LTX-2.3 的实验也支持其鲁棒性。

Insight: 组级核验与训练使提示表达变化成为显式不变性目标，且相同随机种子对比有助于隔离措辞影响。提升集中在提示形式扩展评测；摘要未给绝对得分差和组内表达的构建覆盖，仍需注意意图审计本身的可靠性。

Abstract: Modern text-to-video (T2V) models generate visually compelling videos, yet meaning-preserving changes in prompt wording can substantially alter semantic realization and overall quality. Existing prompt rewriting modifies the inference input, whereas conventional post-training constructs supervision from one prompt formulation at a time; neither explicitly optimizes reliability across equivalent forms. We introduce VARI (Verified Alignment over Representations of Intent), a group-structured framework that treats automatically verified formulations of the same visual intent as a shared supervision unit. VARI builds matched form-seed rollout pools, cross-scores each video against all forms, and applies a wording-independent structured intent audit to identify valid realizations and hard semantic failures. Group-Consistent Distillation (GCD) aligns every verified form with interchangeable valid realizations, while Form-Contrastive Flow-DPO (FC-DPO) learns group-level preferences from audited positive-negative pairs, prioritizing same-seed, different-form contrasts. On HunyuanVideo-1.5, VARI attains the highest T2V-CompBench, VBench, and StoryEval scores among the evaluated methods. On prompt-form extensions of T2V-CompBench and VBench, it improves weakest-form alignment over the strongest prompt-level baselines by 2.45 and 3.55 points and reduces cross-form drop by 30.5% and 32.2%, respectively, while preserving visual quality. Human evaluation, natural-caption controls, prompt-enhancement experiments, diversity analyses, and transfer to LTX-2.3 further support the robustness and transferability of these gains, while VARI requires no additional inference-time module.

[216] Closing the Estimation Gap: Training-Free Transductive Inference for Cross-Domain Few-Shot Action Recognition cs.CVPDF

Changyi Li, Zeyu Yang, Yu Xiao

TL;DR: TRACE 在测试时利用未标注查询逐步改进跨域少样本动作识别的类别原型和决策边界，无需训练或修改骨干。

Details

Motivation: 跨域少样本动作识别需用极少支持样例估计目标域类别原型；即使表征含有区分信息，原型和边界也可能因样本稀缺而估计不准。

Result: 固定表征的 oracle 分析显示，使用真实查询标签构建原型和边界可在 5-way 1-shot、五个目标集上提高 7–20 个百分点，揭示显著估计差距。TRACE 包含两步：CTPR 用高置信查询细化支持集原型；以新原型初始化轻量 episode 分类器，TCR 再鼓励全部查询给出高置信、类别均衡预测以进行转导式适配。在 HMDB51、UCF101、SSv2、Diving48、RareAct 的 1-shot 和 5-shot 设定中，TRACE 稳定优于强基线；HMDB51 1-shot 提升超过 6 个百分点，并缩小相当部分 oracle 差距。

Insight: 当骨干表征已提供可分信息时，测试 episode 内的原型与边界估计可能是主要瓶颈；无需更新骨干也可利用未标注查询改善估计。转导优化假设查询集类别结构可用于适配，摘要未报告类别均衡假设被违反时的鲁棒性。

Abstract: Cross-domain few-shot action recognition (CD-FSAR) aims to recognise unseen action classes in a target domain from a handful of labelled videos. Despite advances in representation learning, CD-FSAR inference still relies on prototypes estimated from extremely limited support examples. To assess this bottleneck, we perform an oracle analysis with the representation fixed. Constructing oracle prototypes and decision boundaries from ground-truth query labels improves 5-way 1-shot accuracy by 7–20 percentage points across five target datasets. This reveals a substantial estimation gap: the representation contains useful discriminative information, but prototypes and boundaries estimated from scarce supports fail to capture each episode’s true class structure. Motivated by this, we propose TRACE, a training-free transductive inference framework that improves episode-level estimation without modifying the backbone. TRACE comprises two complementary stages. Confidence-Thresholded Transductive Prototype Refinement (CTPR) refines support-based prototypes using confidently predicted queries. The refined prototypes initialise a lightweight per-episode classifier, which Transductive Classifier Refinement (TCR) further adapts by encouraging confident, class-balanced predictions over all queries. Together, the two stages improve prototype and boundary estimation, enabling better use of information in the unlabelled queries. Across five target datasets (HMDB51, UCF101, SSv2, Diving48, and RareAct) in both 1-shot and 5-shot settings, TRACE consistently improves a strong CD-FSAR baseline, including gains of more than six percentage points on HMDB51 1-shot, while recovering much of the oracle-estimated gap. These results suggest that improving test-time estimation is a powerful, broadly applicable complement to representation learning for CD-FSAR under extreme support scarcity.

[217] AffectWorld: Benchmarking Emotion-Driven Future World Prediction cs.CV | cs.AIPDF

Changhao Song, Yacong Chen, Hui Gao, Yazhou Zhang, Peng Zhang

TL;DR: AffectWorld 评测世界模型能否生成情绪影响下的人类动作与后续世界状态，覆盖情绪—动作—世界状态一致性及固定场景下的情绪反事实干预。

Details

Motivation: 世界模型基准多关注物理合理性和感知质量，但现实世界也由人类行为构成，情绪会改变行动，进而影响物理和社会状态演化。

Result: 基准包含 1,498 段日常视频样本，标注合理情绪、行动倾向和原子世界状态约束；其中 233 个反事实样本固定场景与事件条件、只改变角色情绪，检查动作和后果是否相应变化。作者定义三类任务及共八项指标，结合自动视觉评估和经人工校准的 VLM/LLM 判断。五个先进世界模型（包括 Cosmos3）的世界状态变化分数仅为 0.100–0.280，反事实情绪敏感度为 0.027–0.070，均在 0–1 标度上，显示生成情绪驱动后果及适应情绪干预方面存在明显不足。

Insight: 基准把情绪影响拆成动作变化和后续状态因果变化，避免只检查文本是否描述了情绪。指标混合自动评估和人工校准模型判断，分数的评审误差与场景覆盖仍需关注；低分主要说明当前被测模型能力不足，不意味着情绪行为有唯一合理答案。

Abstract: The world encompasses both physical environments and human actors. Predicting its evolution requires understanding both physical laws and how emotions shape human actions. However, existing world model benchmarks primarily assess physical plausibility and perceptual quality, leaving emotion-driven world evolution largely untested.We introduce AffectWorld, a benchmark that evaluates the chain Emotion-Action-World State through three tasks: World-State Consistency, Emotion-Action Consistency, and Causal Consistency. It comprises 1,498 everyday video samples annotated with plausible emotions, action tendencies, and atomic world-state constraints. A counterfactual subset of 233 samples tests whether changing an agent’s emotion under fixed scene and event conditions yields corresponding changes in actions and consequences. We define eight metrics across the three tasks and an additional counterfactual metric, combining automatic visual assessment with human-calibrated VLM/LLM judgments. Evaluations of five SoTA world models (e.g., Cosmos3) required world-state change scores of 0.100-0.280 and counterfactual emotion sensitivity scores of 0.027-0.070 on a [0, 1] scale. These results reveal substantial gaps in generating the consequences of emotion-driven actions and adapting future predictions to emotional interventions. AffectWorld provides a testbed for advancing world models that capture how emotions shape future physical and social states.

[218] LongForcing: Scaling Temporal Manifold Matching for Long-Horizon World Rollout cs.CVPDF

Ziyu Zhu, Zhaoxu Sun, Fan Jiang, Mengchao Wang, Junfeng Ni

TL;DR: LongForcing 用扩展时域双向教师监督学生自回归长 rollout，弥补短窗口蒸馏对后期状态缺少直接约束的问题。

Details

Motivation: 视频扩散模型递归使用自身生成帧作为上下文时，超出训练窗口后会逐步产生画质漂移、动作不一致和身份坍塌；短时域分布匹配不一定决定长时行为。

Result: LongForcing 保留短时教师强制和因果 ODE 蒸馏阶段，在最终分布匹配阶段用扩展时域双向教师替代 real-score 教师，监督学生自己的长 rollout；不改变学生架构或推理成本。文生视频评测 VBench Quality 为 85.52；第一人称消融中 WorldRoamBench 60 秒动作分提高 23.9。质性压力测试采样到 24 小时交互 rollout，远超训练时域。

Insight: 长时域 on-policy rollout 可将延迟出现的累积漂移暴露给教师监督，补足短窗口蒸馏盲区。24 小时压力测试是质性展示，不等于长时稳定性已被定量验证；摘要未说明 VBench 及世界模型基准上的基线分数。

Abstract: Autoregressive video diffusion models have emerged as a foundation for long-horizon world rollout. However, when extended far beyond their training windows, these models suffer from progressive drift: visual quality deteriorates, motion becomes inconsistent, and object identity gradually collapses as generated frames are recursively reused as context. We study a supervision-horizon gap: short-horizon distillation leaves later rollout states without direct supervision. We show that matching short-window marginals need not determine long-horizon behavior, while longer on-policy trajectories expose delayed failures to training. This motivates temporal manifold matching at scale for long-horizon world models. Building on this insight, we introduce LongForcing, a distillation approach that scales teacher supervision to long student self-rollouts. LongForcing preserves the short-horizon teacher-forcing and causal ODE-distillation stages, while replacing the realscore teacher in the final distribution-matching stage with an extended-horizon bidirectional teacher. The long-horizon teacher supervises extended student self-rollouts using the standard distribution-matching formulation, without modifying the student architecture or increasing inference cost. We evaluate LongForcing on text-to-video generation and interactive action-conditioned world models, achieving a VBench Quality score of 85.52 and improving WorldRoamBench 60s action score by 23.9 points in first-person ablations. Qualitative stress tests further sample interactive rollouts up to 24 hours, far beyond the training horizon. These results suggest that scaling supervision along the temporal axis is a practical step toward long-horizon world models. The project website and code are at https://github.com/Fantasy-AMAP/LongForcing.

[219] Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes cs.CV | cs.AIPDF

Yu Wang, Rui Dai, Wei Chen, Yi Wang, Bo Dang

TL;DR: Urbex 将城市级 3D grounding 建模为主动空间搜索，让 VLM 智能体查询地标、缩放局部区域、渲染斜视角并逐步确定边界框，而非对预计算候选排序。

Details

Motivation: 城市级场景候选排序依赖预先生成的实例提案，在大规模户外场景中低效；主动搜索能否在无候选框条件下完成 grounding 尚待验证。

Result: Urbex 将城市 3D 场景表示成交互式多视图环境，并以 CityRefer 上的强化学习优化工具使用策略，奖励平滑定位质量并加入轻量探索效率奖励。作者报告其在 CityRefer 上 Hit@1 和 IoU 均最佳，无需真值框即可候选无关 grounding，速度显著快于最强城市级排序基线；在 STPLS3D-Refer 上的零样本结果也显示更强跨场景鲁棒性。摘要未给出具体分数或速度比。

Insight: 交互式空间搜索可省去大批候选框的预计算和全局排序，让模型按需查看地标和局部视角。性能和速度优势在摘要中未量化，工具调用次数、场景地图构建成本与搜索失败案例仍需评估。

Abstract: City-scale 3D grounding is commonly formulated as candidate ranking, which requires precomputed instance proposals and becomes inefficient in large outdoor scenes. We propose Urbex, an agentic vision-language framework that instead casts grounding as active spatial search. Urbex represents a city-scale 3D scene as an interactive multi-view environment, allowing a VLM agent to query landmarks, zoom into local regions, render oblique views, and commit a grounded bounding box. We optimize this tool-use policy with reinforcement learning on CityRefer, using smooth localization rewards and lightweight shaping rewards for efficient exploration. Experiments show that Urbex achieves the best Hit@1 and IoU on CityRefer, supports candidate-free grounding without ground-truth boxes, and is substantially faster than the strongest city-scale ranking baseline. Zero-shot results on STPLS3D-Refer further indicate improved cross-scene robustness.

[220] Omni-RS: Unifying Multi-Task Understanding and Generation in Remote Sensing cs.CVPDF

Pan Yin. Shuai Yang, Yuanhao Liu, Datao Tang, Jialin Li, Hao Wang

TL;DR: Omni-RS 将遥感场景理解和图像生成统一起来，先用场景引导两阶段训练适配 MLLM，再以其隐藏状态指导扩散 Transformer，并构建多任务生成数据集 UniGen。

Details

Motivation: 实际遥感应用既需理解地理场景也需生成内容，但多数方法只处理其中一项；现有统一模型的生成能力通常限于文本到图像。

Result: 框架支持四类理解任务和五类生成任务。训练先让多模态大语言模型适配遥感理解，再用其隐藏状态引导扩散 Transformer 生成。为统一不同遥感生成任务，作者将异构数据整理为 UniGen；Detailed Visual Conditioning 通过 VAE 编码视觉条件，支持需要参考图像的生成任务。作者称模型在理解和多种生成任务上均表现突出。摘要未提供任务名称、数据规模、指标或定量比较。

Insight: 以场景理解状态作为生成条件，试图共享语义表示并覆盖参考图像条件任务，超越单一文本生成。当前摘要的结果仅为定性表述，无法衡量理解与生成之间是否存在权衡，也无法判断 UniGen 和训练策略的单独作用。

Abstract: Practical remote sensing applications often require both understanding and generation capabilities for the same geographic scene. However, most existing methods address either understanding or generation in isolation, while existing unified remote sensing models’ generative ability is limited to text-to-image. To bridge this gap, we propose Omni-RS, a unified framework supporting four remote sensing understanding tasks and five generation tasks. Specifically, we introduce the “Scene-Guided Two-Stage Training” strategy, which first adapts a multimodal large language model to remote sensing understanding and then uses its hidden states to guide a diffusion transformer for image generation. To enable joint learning across diverse generation tasks, we construct UniGen, a unified multi-task generation dataset, by organizing heterogeneous remote sensing data into a unified format. Furthermore, we introduce “Detailed Visual Conditioning”, which uses visual conditions encoded by a VAE to guide the generation process, thereby supporting additional generation tasks that require reference images. Extensive experiments demonstrate that Omni-RS performs exceptionally well in understanding tasks and has achieved outstanding results in a variety of generative tasks.

[221] Hallucination-aware Faithful Reasoning in Remote Sensing cs.CVPDF

Yi Liu, Jing Zhang, Xiang Feng, Xiaoyu Tian, Haonan Guo

TL;DR: FaithRS 通过细粒度遥感幻觉标注训练检测器，再以安全门控的纠正 rollout 和复合奖励优化 MLLM 的推理忠实度。

Details

Motivation: 遥感 MLLM 可能生成与视觉证据矛盾的推理；只优化最终答案正确性，缺少对中间推理幻觉类型、原因和可见支持证据的监督。

Result: 作者构建 SFHD-43K，含 43,000 条结构化细粒度幻觉标注，用于检测幻觉类别、根因和视觉证据。FaithRS 将诊断结果用于安全感知纠正 rollout，由门控器选择是否以及如何修订，再以复合奖励联合优化答案正确性、推理忠实度、幻觉减少、格式和修订安全性。在 Qwen3.5 的 0.8B、4B、9B 骨干及四个遥感基准上，作者报告持续改进，最高增益 11.64 分；摘要未说明该分值对应指标或各骨干成绩。

Insight: 幻觉修订不仅需要判断哪里错，还要检查修订是否有证据且不会造成不安全替换；安全门控将是否修改纳入学习。最高提升限于所报告基准，摘要未给出标注一致性、幻觉类别差异和修订导致新错误的比例。

Abstract: Multimodal large language models (MLLMs) have demonstrated promising capabilities for remote sensing reasoning, yet they frequently generate visually unsupported reasoning that contradicts the visual evidence. Existing post-training methods optimize primarily for final-answer correctness, providing limited supervision for identifying and correcting hallucinations that arise during the reasoning process. To address this limitation, we propose \textbf{FaithRS}, a hallucination-aware self-reflection framework for faithful reasoning in remote sensing. We first construct \textbf{SFHD-43K}, a dataset of 43K structured, fine-grained hallucination annotations, and train a detector to identify hallucination types, underlying causes, and supporting visual evidence. FaithRS then incorporates this diagnostic feedback into a Safe Detector-Guided Corrective Rollout framework, where a safety-aware gate selectively determines whether and how to revise the response. Training is driven by a composite reward that jointly optimizes answer accuracy, reasoning faithfulness, hallucination reduction, format compliance, and revision safety. Experiments with Qwen3.5 backbones (0.8B, 4B, and 9B) show consistent improvements across four remote sensing benchmarks, with gains of up to 11.64 points.

[222] ReDiS: Reasoning-Diagnosis Self-Evolution for Multimodal Reasoning in Remote Sensing cs.CVPDF

Yi Liu, Wentao Jiang, Shuo Ni, Zhiming Luo, Di Wang

TL;DR: ReDiS 将遥感多模态推理拆为生成多样候选的 Reasoner 与评估整组候选可信度的 Diagnoser，并用全错组触发的 FTR 提供最小纠正信号。

Details

Motivation: 遥感细粒度视觉定位和多步推理仍困难；昂贵人工构造推理轨迹成本高，而基于模型自生成一致性的监督可能强化共享错误。

Result: 两个角色共享多模态骨干并共同演化：Reasoner 探索不同视觉证据—答案候选，Diagnoser 联合评估完整候选集并给出可信度感知反馈。若组内候选全错、组相对优势消失，Failure-Triggered Recovery 仅提供最小结果级纠正，不规定中间推理轨迹。在 LHRS-Bench、MME-RS、CHOICE 上，0.8B、2B、4B 的 Qwen3.5 骨干均报告一致改善；0.8B 分类平均准确率提高 6.47 个百分点，ReDiS-4B 在所比较遥感 MLLM 中平均性能最高。

Insight: 分离探索和评估可避免直接把答案一致性当成正确性，并让诊断信号随候选分布变化；FTR 解决全错组没有相对优势的问题。摘要未给其他规模增益和单项任务成绩，所称领先限于三个遥感基准及比较模型。

Abstract: Multimodal large language models (MLLMs) have shown strong potential for remote sensing understanding, yet reliable reasoning remains challenging when tasks require fine-grained visual grounding and multi-step inference. Existing approaches typically rely on either costly pre-constructed reasoning trajectories or self-generated signals such as response consistency, which can reinforce shared errors and provide unreliable supervision. We introduce \textbf{Reasoning–Diagnosis Self-Evolution (ReDiS)}, a framework that separates reasoning exploration from candidate evaluation while allowing the two processes to evolve jointly. ReDiS instantiates two role-specific experts from a shared multimodal backbone: a \textbf{Reasoner} that explores diverse visual evidence–answer candidates and a \textbf{Diagnoser} that jointly evaluates the complete candidate set and provides credibility-aware feedback. As the Reasoner evolves, it exposes new candidate distributions and failure modes; the Diagnoser adapts to these changes and, in turn, reshapes the learning signal for subsequent exploration. To handle groups in which all sampled answers are incorrect and group-relative advantages vanish, we further introduce \textbf{Failure-Triggered Recovery (FTR)}, which provides minimal outcome-level correction without prescribing intermediate reasoning trajectories. Experiments on LHRS-Bench, MME-RS, and CHOICE show consistent improvements over Qwen3.5 backbones at 0.8B, 2B, and 4B scales. ReDiS improves category-averaged accuracy by 6.47 percentage points at 0.8B, while ReDiS-4B achieves the highest average performance among the compared remote sensing MLLMs.

[223] Bifocal: Training-Free Mixed Granularity Visual Token Pruning for Multimodal LLMs cs.CVPDF

Peiran Yin, Hao Wu, Zihao Yang, Fengyu Zhou, Yijuan Liang

TL;DR: Bifocal 将视觉 token 预算分给问题相关证据、图像固有信息和全图粗表征，并在 LLM 中继续按密度调整注意力剪枝，以保护 OCR 和定位等细粒度能力。

Details

Motivation: 既有视觉 token 剪枝在常见 VQA 上保留少量 token 仍能维持总体性能，却可能严重损害 OCR 和视觉 grounding；单一重要性排序会忽略不同任务需要的互补证据，稀疏选 patch 也会破坏空间完整性。

Result: Bifocal 是训练免除两阶段方法：LLM 前将问题词与视觉区域软匹配，并平衡去范数偏差的图像显著性，同时选择覆盖度高的图像固有信息，再保留全图池化粗表示；LLM 内按密度修正文本到图像注意力并剪枝。在三个 MLLM 家族、五种能力组评估中，约保留四分之一原始 token 时，Bifocal 是唯一在全部能力组和三种模型上都保留超过 90% 未剪枝性能的方法；九种模型—预算设置的平均排名均第一。

Insight: 同时保留查询驱动、图像固有和全局粗粒度信息，避免通用排名只偏向一个任务目标；能力分组评测也揭示总体 VQA 分数掩盖的细粒度退化。摘要的“约四分之一 token”结果依赖所评模型和能力组，未提供各任务的绝对分数或不同预算下成本。

Abstract: Visual token pruning for multimodal large language models (MLLMs) appears highly effective, with existing methods retaining over 95% of unpruned performance on common VQA benchmarks using a quarter of the visual tokens or fewer. However, our extensive evaluation of ten methods across three MLLM families and five ability groups reveals a critical gap: strong general performance often comes at the cost of severe degradation on fine-grained tasks such as OCR and visual grounding. We trace these failures to two limitations of the prevailing paradigm of retaining the highest-ranked native patches. A single importance ranking overlooks the complementary evidence required by different tasks, while sparse patch selection disrupts both image content and spatial integrity. To address these limitations, we propose Bifocal, a training-free, two-stage pruning method built on a mixed-granularity framework. It allocates separate budgets to query-focused and image-intrinsic evidence, complemented by a pooled coarse representation of the entire image. Before the LLM, Bifocal matches question words to visual regions through a soft conjunction balanced against norm-debiased image saliency, while selecting image-intrinsic evidence to maximize representational coverage. Inside the LLM, it further prunes tokens using density-adjusted text-to-image attention. With roughly a quarter of the original visual tokens, Bifocal is the only evaluated method to retain over 90% of unpruned performance in every ability group across all three models, and ranks first on average in all nine model-budget settings. The code is available at: https://github.com/EIT-NLP/Bifocal

[224] CoTL: Compact Token Learning for 3D Scene Representation cs.CVPDF

Han Liang, Sibo Wu, Anpei Chen, Rüdiger Westermann

TL;DR: CoTL 将无姿态多视角图像压缩为随视图数量小幅扩展的紧凑潜 token 集，并通过可查询高斯解码器和新视角渲染监督保留场景几何与外观。

Details

Motivation: 3D 场景表示需要同时容纳多视角观测并保持一致几何和外观；希望用紧凑且不依赖预定义结构的 token 表征支持重建及后续生成。

Result: 编码器将各视图压缩为小组潜 token，并跨视图聚合为一致场景表示；每增加一视图仅增加少量、与分辨率无关的 token。场景无关的查询通过潜 token 恢复显式可渲染 3D Gaussian primitives，并用高斯场景的新视角渲染监督端到端训练。在摘要所列两个数据集（其中一个名称显示为未展开占位符“retenk”，另一个为 ACID）上，作者报告新视角合成具有竞争力，且 token 空间可作为 3D 高斯场景生成模型的潜变量。摘要未提供具体分数。

Insight: 将 3D 高斯重建作为训练目标，可迫使紧凑 token 保存可渲染几何和外观；表示容量随观察数量增长但不随图像分辨率膨胀。一个数据集名称在摘要中为模板占位符，且没有量化结果，具体基准及性能需查正文确认。

Abstract: We present CoTL, a framework that learns a compact set of tokens that preserves scene geometry and appearance from unposed multi-view images. Our encoder compresses the input images into small per-view sets of latent tokens that aggregate information across views to form a view-consistent scene representation. Each additional view expands the representation with only a small, resolution-independent set of tokens, allowing its capacity to grow with the available observations. To force this compact bottleneck to retain sufficient scene information, a query-based Gaussian decoder uses scene-independent queries to attend to the learned tokens and recover an explicit, renderable set of 3D Gaussian primitives. Novel-view rendering supervision trains the compact representation end to end through the reconstructed Gaussian scene. Experiments on \retenk{} and ACID show that the resulting compact, unstructured token set supports competitive novel-view synthesis and can serve as the latent space of a generative model for 3D Gaussian scene generation. Together, these results show that 3D Gaussian reconstruction is an effective objective for compact scene representation learning and the learned token space provides a basis for broader downstream tasks in 3D vision.

[225] EmoGenie: Benchmarking and Evaluating Expressed and Evoked Emotion in T2I Generation with MLLMs cs.CVPDF

Zixuan Xing, Lancheng Gao, Shengyan Li, Huiyu Duan, Zhaolin Cai

TL;DR: EmoGenie 统一评测文生图生成内容表达的情绪和观看者被唤起的情绪，并评估 MLLM 对 AI 生成图像情感的粗粒度问答与连续维度预测。

Details

Motivation: 生成图像可带有特定情感意图，但画面传递或引发的情绪常与目标不符；既有情感基准多使用自然图像，生成图像研究又常局限于人像表达，缺少统一基准。

Result: EmoGenie 含 8 个文生图模型根据 800 个情绪提示生成的 6,400 张图像。EmoGenie-QA 有 90,000 个单图/图像对闭集问题和 1,057 个开放式情绪描述问题；EmoGenie-Assess 收录 13,175 个由人工评分聚合的 MOS，覆盖效价、唤醒度、支配感和审美质量。作者开发 EmoProbe，将情绪问答与响应对齐的分维度回归头结合。实验报告其情绪 QA 表现更好，对情感及审美人工评分的相关性也更高。

Insight: 表达情绪与观看者感受并非同一目标，单一图像情绪标签无法涵盖二者；基准将问答和连续心理维度分开评测。摘要未给具体相关系数、人工评分者规模和不同文化背景分析，跨文化情绪标签的适用性尚不明确。

Abstract: Rapid advances in artificial intelligence-generated content (AIGC) have significantly promoted using text-to-image (T2I) models to generate images with specific affective intentions. However, the emotional effects conveyed by generated images frequently deviate from target expectations, and it remains unclear whether multimodal large language models (MLLMs) can reliably interpret the emotion of AI-generated visual content. Existing affective benchmarks are predominantly built on natural images, while affective studies of generated images have largely focused on emotions expressed in portraits, lacking a unified benchmark for studying the emotions expressed and evoked by generated images. We therefore introduce EmoGenie, a benchmark of emotion for AI-generated images that evaluates capabilities in T2I generation and MLLM understanding across expressed and evoked emotions. It comprises 6,400 images generated by eight T2I models using 800 emotion-targeted prompts, annotated with question-answer (QA) pairs, termed EmoGenie-QA, and continuous human ratings, termed EmoGenie-Assess. EmoGenie-QA contains 90,000 closed-set questions in single-image and image-pair settings, as well as 1,057 open-ended questions for emotion description. EmoGenie-Assess benchmarks continuous affective and aesthetic prediction using 13,175 mean opinion scores (MOSs) aggregated from human ratings in terms of valence, arousal, dominance, and aesthetic quality. We further develop EmoProbe, which combines emotion question answering with response-aligned, dimension-specific regression heads for continuous score prediction. Experimental results show that EmoProbe achieves superior results on emotion-related QA tasks, while its affective and aesthetic predictions correlate better with human ratings, demonstrating its state-of-the-art capabilities of unified coarse-to-fine affective understanding for AI-generated images.

[226] Simplifying and Improving On-Policy Distillation for Few-Step Diffusion cs.CV | cs.LG | eess.IVPDF

Jiawei Zhang, Ziyu Wan, Nanye Ma, Long Zhao, Dong Chen

TL;DR: SMD 用简单的 on-policy 速度回归作为少步扩散蒸馏基础，以训练期残差代理 Pivot 表示粗时间段内部教师速度变化，并排斥较弱教师预测。

Details

Motivation: 少步蒸馏可降低扩散和 flow 生成采样成本，但不少方法需要额外 score 模型或专用架构；目标是在保持通用架构的同时获得强 4-NFE 生成。

Result: Pivot 将局部修正与学生对粗区间平均速度的预测分离，并在训练中使用；额外排斥较弱教师预测以减少不理想模式。SMD 完全基于回归，无在线 score 估计和对抗训练，部署生成器结构不变。评估覆盖图像生成/编辑、3D 生成、视频及联合音视频生成，模型规模最高 330 亿参数。在 4-NFE 下，作者报告 SMD 超过强少步基线，且在若干基准超过多步教师；还支持跨模型、多教师和因果视频蒸馏。摘要未列出具体基准与数值。

Insight: 以学生自身 rollout 做速度回归，并显式建模粗步长内局部速度变化，提供了较简洁的少步蒸馏路径。作者称在多任务和大规模上有效，但摘要未报告各任务指标、训练稳定性和不同 NFE 下的质量曲线。

Abstract: Few-step distillation reduces the sampling cost of diffusion and flow models, but existing approaches often require auxiliary score models or specialized architectures. We show that simple on-policy velocity regression already provides a strong foundation for 4-NFE generation. To represent teacher velocity variation within each coarse transition, we introduce Pivot, a training-only residual proxy that separates local corrections from the student’s interval-mean prediction. We further repel the student from weaker teacher predictions, providing a lightweight mechanism for avoiding undesirable modes. The resulting framework, Simplified Matching Distillation (SMD), remains entirely regression-based, requires neither online score estimation nor adversarial training, and leaves the deployed generator unchanged. We evaluate SMD on image generation and editing, 3D generation, video synthesis, and joint audio–video generation, scaling to 33 billion parameters. At 4-NFE, SMD outperforms strong few-step baselines and surpasses multi-step teachers on several benchmarks. It also supports cross-model, multi-teacher, and causal-video distillation.

[227] MFVSR: A Multi-Style and Fine-Grained Benchmark for Video Subtitle Removal cs.CVPDF

Xiyan Huang, Qingdong He, Peng Tang, Bo Yin, Chaoyi Wang

TL;DR: MFVSR 建立 3,000 对多风格视频基准及细粒度可靠性评估协议，分别测量字幕移除的局部复原、外部区域保留、时间稳定性、覆盖率和失败。

Details

Motivation: 现有视频字幕移除评测受低分辨率数据、测试集碎片化和通用视频修复协议限制，难以刻画局部修复错误、意外改动、时间不稳定及未完成输出。

Result: MFVSR 含 3,000 对 Real-HQ、Real-LQ 和 AI 生成视频，并附加 953 段真实野外视频用于覆盖分析；字幕变化覆盖 16 种语言及空间布局、时间行为、视觉风格、语言模式四个因素。信息条件化、空间解耦的协议区分盲移除与掩码引导移除，评估复原忠实度、保留忠实度、时间忠实度、覆盖及失败。六种代表方法的实验显示，局部修复、编辑区外保留、时间一致性和整套数据可靠性之间差异显著，传统全帧指标会掩盖这些差异。

Insight: 将编辑区域内外和时间维度拆开量化，能避免单一全图指标掩盖删字幕之外的破坏；盲移除与掩码引导任务也需分别评测。摘要未提供六种方法的具体分数，953 段野外视频用于覆盖分析，并非同规模配对真值测试集。

Abstract: Video subtitle removal (VSR) is an important video restoration task, yet existing evaluation remains limited by low-resolution datasets, fragmented test sets, and generic video-inpainting protocols that fail to adequately capture localized restoration errors, unintended modifications, temporal instability, and incomplete outputs. In this paper, we introduce MFVSR, a large-scale multi-style benchmark for fine-grained and reliability-aware video subtitle removal evaluation. MFVSR distinguishes itself through three key features: (1) Large-scale and diverse data, comprising 3,000 paired videos from Real-HQ, Real-LQ, and AI-Generated sources, with an auxiliary set of 953 in-the-wild real videos for real-world coverage analysis; (2) Controlled subtitle diversity, spanning 16 languages and four factors covering spatial layout, temporal behavior, visual style, and language mode; (3) Fine-grained evaluation, introducing an information-conditioned, spatially decoupled protocol that separates blind from mask-guided removal and measures Restoration Fidelity, Preservation Fidelity, Temporal Fidelity, coverage, and failures. Experiments on six representative methods reveal substantial discrepancies between local restoration, outside-region preservation, temporal consistency, and full-set reliability that are obscured by conventional full-frame metrics. MFVSR therefore provides a more comprehensive and diagnostic foundation for evaluating and advancing video subtitle removal methods.

[228] Physics-Aware Deepfake Detection via Distance-Speech Consistency cs.CVPDF

Kyeongrae Kim, Kim Sung-Bin, Oh Hyun-Bin, Tae-Hyun Oh

TL;DR: 该物理先验深伪检测器利用真实录音中语音能量随说话者—相机距离变化的规律，与视频距离估计进行一致性检测，以补充动态场景下易失效的唇音同步线索。

Details

Motivation: 多数音视频深伪检测依赖唇动—语音同步，常为静态正面视频设计；真实动态拍摄中说话者移动、唇部线索退化或缺失，可靠性可能下降。

Result: 方法融合视频估计的说话者距离与语音中依赖距离的声学测量，利用真实录制中的物理耦合关系识别可能被合成破坏的关联。在野外基准上，作者报告该物理线索对动态场景有效，并与唇同步检测互补；简单集成在多个数据集上均带来稳定增益。摘要未给数据集名称、准确率或融合增益。

Insight: 使用声学传播与拍摄距离的物理一致性，提供不同于视觉唇动对齐的判据，可能覆盖唇部不可见情况。实际效果依赖距离估计和录制环境；摘要未量化设备、混响、说话音量变化等因素的影响。

Abstract: Recent progress in generative models has made talking-head deepfakes highly realistic and easy to create, jeopardizing public trust. However, most audio-visual deepfake detectors rely on lip-speech synchronization and are designed for static, frontal videos, limiting their reliability in dynamic, in-the-wild recordings where the speaker moves and lip cues are degraded or missing. As an alternative, we propose a physics-aware detector for dynamic speaking videos that leverages an acoustic constraint: in real recordings, speech energy varies predictably with speaker-to-camera distance, whereas deepfake generation can break this coupling. Our method combines distance estimates from video with distance-dependent acoustic measures from speech to detect deepfakes. Experiments on in-the-wild benchmarks show that the proposed physical cue is effective in dynamic scenarios, and that it complements lip-sync-based detectors, with a simple ensemble achieving consistent gains across datasets. Codes are available at https://byulharang.github.io/PADD/.

[229] EgoOmniWorld: Unifying Egocentric Understanding via Ego-Action-World Interactions cs.CV | cs.AIPDF

Songruo Chen, Yuqian Fu, Yu Li, Mingkang Dong, Sifan Zhou

TL;DR: EgoOmniWorld 用“自我—行动—世界”交互环统一组织第一人称视频理解，按节点、关系、依赖链和反馈闭环构建证据锚定问答基准。

Details

Motivation: 第一人称视频能力涉及多种模态和评测形式，但常被拆开研究；真实互动中，佩戴者感知环境、形成目标、行动改变世界并依据反馈继续决策，这些环节相互耦合。

Result: 基准将交互结构分为 Node（自我/行动/世界状态）、Edge（有向关系）、Chain（依赖链）和 Loop（闭环反馈）。数据流程从第一人称视频提取结构化互动、用视听证据落地，并按交互结构生成问题，所需模态及单/多轮格式由证据和依赖决定。数据含 24,166 组问答，其中训练 20,284、测试 3,882。对开源和闭源 MLLM 的实验显示各交互结构能力存在明显且不均衡的短板；使用该数据训练可改善域内互动理解，并迁移到既有第一人称基准。

Insight: 按互动结构而非单一问答类型组织评测，可以显式覆盖状态、因果依赖和行动反馈。摘要未列各结构分项结果、模态缺失鲁棒性或迁移增益数值，需避免把总体改进理解为所有交互能力均衡提升。

Abstract: Egocentric video understanding spans diverse capabilities, modalities, and evaluation formats, yet these are often studied in isolation despite their tight coupling in real-world interaction. We propose the Ego-Action-World interaction loop as a unifying principle: the wearer perceives the world, forms goals, acts, changes the world, and uses feedback to guide subsequent behavior. Based on this view, we introduce EgoOmniWorld, a benchmark organized around four complementary interaction structures: Node captures ego, action, and world states; Edge captures directed relations among them; Chain captures linked dependencies; and Loop captures feedback-closed interactions. We develop an interaction-first, evidence-grounded data pipeline that extracts structured interactions from egocentric videos, grounds them in visual and auditory evidence, and generates questions according to their interaction structure. Required modalities and single- or multi-turn formats follow naturally from the supporting evidence and dependencies. EgoOmniWorld contains 24,166 QA groups, including 20,284 training and 3,882 test samples. Experiments on representative open- and closed-source multimodal large language models reveal substantial and uneven limitations across interaction structures. Models trained on EgoOmniWorld improve in-domain interaction understanding and generalize to existing egocentric benchmarks. EgoOmniWorld provides a unified foundation for evaluating and advancing egocentric video understanding. Data, code, and models will be released.

[230] Meta-Video: Efficient Standardized Evaluation for Video Understanding cs.CVPDF

Soumya Shamarao Jahagirdar, Anna Kukleva, Vishaal Udandarao, Samuel Thomas, Brian Kingsbury

TL;DR: Meta-Video 揭示帧采样、预处理和输入格式等评测配置会大幅改变视频理解分数，并提供轻量基准和标准化协议以提高结果可复现性。

Details

Motivation: 视频基准分数常由黑盒评测框架的预设配置产生，默认帧采样和预处理选择很少被检查；同一模型因配置不同也可能出现明显波动，且一些题目不看视频也能回答，完整基准评测昂贵。

Result: 作者发现同一模型检查点仅因评测配置变化，准确率可相差最高 19.5 个百分点。Meta-Video 从既有套件筛选出 1,000 道平衡题、覆盖 791 段视频和多样任务/时长，在 4–128 帧完整帧数扫描实验中大体保留模型排名，评测速度提高 5 倍。作者系统扫描配置空间，整理影响结果的最佳实践与报告标准，并发布 5 个模型、7 个基准的可复现参考分数。

Insight: 配置差异可能与方法改进同量级，因此标准化帧采样、预处理和输入格式是可信比较的前提；轻量基准可降低全套评测成本。19.5 点波动来自所测配置，Meta-Video “大体保留排名”不保证所有模型排序完全一致。

Abstract: Video understanding benchmarks are the primary yardstick for measuring progress in video-language modeling, and a growing body of work relies on them to claim improvements. The reported benchmark scores in these works, however, are almost always produced by frameworks (e.g., lmms eval) used as black boxes: certain evaluation choices like frame sampling, preprocessing, and input formatting are inherited from preset defaults and seldom inspected carefully. We show these unexamined choices are a major source of variance. Even for the same model checkpoint, evaluation configuration alone can shift accuracy by up to 19.5pp. Hence, it is unclear whether reported gains reflect genuine improvements in capability or differences in evaluation setup. Moreover, many benchmark samples offer little evidence of visual capability, since they can be answered without the video, and the full suites are expensive to evaluate. We address these issues in two parts. First, we curate existing suites for discriminability and efficiency, producing Meta-Video, a balanced 1,000-question benchmark over 791 videos spanning diverse tasks and durations, which largely preserves model rankings, at 5 times inference speedup for a full-frame count sweep experiment (4-128 frames). Second, we establish a standardized evaluation protocol by sweeping the configuration space to isolate choices that measurably affect results, distill them into best practices and reporting standards, and release reproducible reference scores across five models and seven benchmarks. We hope that the proposed evaluation framework and the Meta-Video benchmark will support video understanding research by providing a reproducible and efficient foundation to measure progress in a reliable way.

[231] Beyond the Encoder: Object-Centric Representations Emerging from Segmentation Foundation Models cs.CV | cs.LGPDF

Michalis Pistos, Bartłomiej W. Papież

TL;DR: 该研究将 SAM2 分割路径中生成的对象指针作为可迁移对象级表示，发现仅用分割监督学习的指针也包含肿瘤判别信息，并能用于多种下游预测。

Details

Motivation: 基础模型下游迁移通常使用编码器表示，但分割模型在执行任务时还生成紧凑对象级表示，其可迁移性尚未充分研究。

Result: 作者在两个 PET/CT 肿瘤预测任务中研究 SAM2 object pointers。仅用分割监督训练的指针，在未接触下游标签时已编码肿瘤判别信息；任务引导指针通过辅助目标联合塑造分割与下游预测表示，在癌种分类上超过传统编码器迁移和 radiomics。更具挑战的复发预测任务中，分割监督指针反而最好。表征分析发现指针与可解释 radiomics 特征相关、表示空间具有类别组织，并对肿瘤分割区域形成明确空间 grounding。摘要未给具体分数。

Insight: 基础模型执行其原始分割任务时产生的对象级表示可能比通用编码器特征更贴近下游病灶分析；任务引导适配也并非在所有任务都更好。结论来自两个 PET/CT 任务，摘要未报告队列规模、外部验证或临床变量比较。

Abstract: Foundation models are increasingly used as feature extractors for downstream tasks with limited data, typically through representations obtained from their encoders. However, foundation models trained for structured tasks such as segmentation also form object-specific representations during task execution, whose transferability remains largely unexplored. We investigate Segment Anything Model 2 (SAM2) object pointers, compact object-level representations produced through the segmentation pathway, as transferable representations for downstream tumour prediction. Through experiments on two PET/CT tumour prediction tasks, we find that object pointers learned using segmentation supervision alone already encode tumour-discriminative information, despite never being optimised for the downstream labels. We further introduce task-guided object pointers, where an auxiliary objective jointly shapes the object pointer space for the downstream prediction together with segmentation. Task-guided pointers improve cancer-type classification, outperforming conventional encoder-based transfer and radiomics, while segmentation-only pointers perform best on the more challenging recurrence prediction task. Representation analyses reveal associations between object pointers and interpretable radiomics features, class-dependent organisation of the learned representation space, and explicit spatial grounding around the segmented tumour. Together, these findings suggest that useful transferable representations can emerge through the execution of a foundation model’s primary task, providing an alternative to conventional encoder-based transfer.

[232] Parameter-Efficient Multimodal Brain Tumor Segmentation With Stage-Wise Shared Adapters cs.CVPDF

Milad Rostamian, Mohammadreza Mohammadi, Ender Konukoglu, Mohsen Soryani

TL;DR: BWSA 在多模态 MRI 脑肿瘤分割网络的各阶段通过共享适配路径压缩并交换模态特征，提升 Dice 且仅增加少量参数。

Details

Motivation: 不同 MRI 模态提供肿瘤互补信息；常见方法为每个模态配置独立编码器，多阶段特征交换虽有帮助却会增加模型复杂度。

Result: Block-Wise Shared Adapter 将各模态编码特征压缩到共享路径融合，再把适配后的特征送回各模态流。固定混合版本把适配特征加回原特征；TG-BWSA 在每个模态与编码器阶段学习独立门控权重。两种变体在 BraTS 2020、BraTS 2023 上使用四种分割骨干评测。BraTS 2023 官方在线验证中，BWSA 相较对应未修改骨干使平均区域 Dice 提高 1.3–7.6 个百分点；TG-BWSA 提高 2.8–7.7 个百分点。两者参数增量约 0.85%–3.73%，平均 HD95 均降低。

Insight: 共享适配器在多个模态间复用融合计算，减少为每模态单设交互模块带来的参数增长；可学习门控的增益区间整体更高。不同骨干和任务的提升幅度仍有差异，摘要未给各肿瘤区域单项分数及推理开销。

Abstract: Different MRI modalities reveal different parts of a brain tumor. Segmentation models combine information from these modalities to outline tumor regions. Many models use a separate encoder for each modality. However, exchanging features among these encoders at several stages can increase model complexity. We introduce the Block-Wise Shared Adapter (BWSA), which combines compressed features from the modality encoders in a shared path and returns adapted features to each stream. In its fixed-blending form, BWSA adds the adapted features to the original features. Its trainable-gated variant (TG-BWSA) instead learns separate weights for the original and adapted features at each modality and encoder stage. We evaluate both variants on BraTS 2020 and BraTS 2023 with four segmentation backbones. On the official BraTS 2023 online validation set, BWSA increases average region-wise Dice by 1.3–7.6 percentage points relative to the corresponding unmodified backbones. TG-BWSA increases it by 2.8–7.7 points. Both variants add approximately 0.85–3.73% parameters and reduce average HD95 relative to the unmodified backbones.

[233] LLM2R: From Large Language Models to 3D Reassemblers cs.CV | cs.AI | cs.ROPDF

Zhouhan Zhong, Qiuhong Shen, Xingyi Yang, Jinyi Xu, Xinchao Wang

TL;DR: LLM2R 将多部件 3D 装配建模为单次非自回归序列预测，用 Point-BERT/Q-Former 编码各部件，再用汇总 token 汇集完整部件集合信息并预测每件 SE(3) 位姿。

Details

Motivation: 部件摆放通常不能仅由自身几何决定，需要综合整个部件集合；成对关系方法或逐步生成—验证在部件多时扩展较差。因果注意力还使序列中较早部件无法看到后续部件。

Result: 模型把各部件压为几何 token 块并拼接，经 LoRA 适配的 decoder-only LLM 处理；追加位于所有部件之后的可学习 summary tokens 汇集全局装配信息，再由部件感知位姿头融合部件状态与摘要 token，统一预测旋转和平移。在三个 PartNet 类别上，LLM2R 的部件准确率和成功率均最高；PartNeXt 上旋转误差、平移误差和部件准确率最佳；Breaking Bad 上平移误差最低、推理最快，但精度仅低于专用骨折方法。100 部件时，它在 Breaking Bad 部件准确率比 GARF 高 33.5 点、PartNet 比 Assembler 高 29.1 点；对未见类别零样本迁移比三个基线中最强者高 16 点以上。

Insight: 将所有部件同时编码并通过末尾汇总 token补足因果注意力方向性，可在多部件场景中实现一次性全局推理。结果覆盖不同数据集，Breaking Bad 上专用方法仍更准确，说明通用装配性能并非所有类别最优。

Abstract: 3D part assembly aims to reconstruct a complete shape from a set of unassembled parts by predicting a rigid SE(3) transformation for each part. Because the placement of a part is rarely determined by its own geometry, the task demands reasoning over the entire part set, and existing methods, which rely on pairwise relations or iterative generation and verification, scale poorly to objects with many parts. We present LLM2R, which casts part assembly as sequence modeling with a pretrained decoder-only large language model (LLM): a Point-BERT encoder and a Q-Former compress each part into a block of geometric tokens, the blocks of all parts are concatenated, and a causal LLM adapted with Low-Rank Adaptation (LoRA) processes the sequence in one forward pass. Causal attention, however, lets the hidden states of a part observe only the parts that precede it. LLM2R therefore appends a block of learnable summary tokens after the last part: these tokens follow every part, so their states aggregate the whole assembly at a constant token cost, and a part-aware pose head fuses each part’s states with them before predicting its rotation and translation. One architecture thus handles semantic and geometric assembly and predicts all poses in a single non-autoregressive pass. LLM2R achieves the highest part accuracy and success rate on all three PartNet categories, the best rotation error, translation error, and part accuracy on PartNeXt, and the lowest translation error and fastest inference on Breaking Bad, where only a fracture-specialized method is more accurate. Its advantage grows with the number of parts: at 100 parts it exceeds GARF by 33.5 points in part accuracy on Breaking Bad and Assembler by 29.1 points on PartNet, and it transfers zero-shot to unseen categories with more than 16 points higher part accuracy than the strongest of three baselines.

[234] Rethinking Projector Training in Multimodal LLMs cs.CV | cs.AI | cs.CLPDF

Fabian Gröger, Naël Ouerghemi, Shuo Wen, Maria Brbić

TL;DR: PORTAL 以视觉和语言特征的模态级统计量构造最优传输投影初始化，使 MLLM 可在无配对图文数据、无梯度优化条件下跳过传统 projector 训练。

Details

Motivation: 多模态 LLM 通常先用大规模图像—描述对训练投影器，再联合训练全模型，但 projector 阶段的作用尚不明确且需配对数据。

Result: 作者发现使用仅 10%–20% 数据训练 projector 已可恢复大部分下游增益，不同数据集训练的 projector 也可线性插值并大体保留性能，说明该阶段可能主要完成粗略对齐。PORTAL 仅需各模态汇总统计，在共享主成分子空间上为视觉、语言特征的高斯近似计算最优传输映射以初始化 projector。无任何配对数据时，该初始化在 6 个 LLM 骨干、2 个视觉编码器和 16 个基准上匹配用 55.8 万图文对训练的标准流程；7 组视觉编码器—LLM 设置中的 3 组胜过标准流程，其余 4 组平均达到其 98% 分数。

Insight: 若 projector 主要做模态级粗对齐，分布统计和线性映射或可取代大规模成对训练；跨多个模型和基准的结果支持这一简化。PORTAL 在三组配置中超过基线而整体其余平均接近，并非所有组合都完全匹配；摘要未提供数据细分和后续联合训练成本。

Abstract: Multimodal LLMs are typically trained by first aligning a vision encoder with a language model via projector training on large image-caption datasets, followed by joint training of the full multimodal model. Despite its widespread use, the role of the projector training stage remains poorly understood. We find that (i) using only $10$ to $20%$ of the data for training the projector already recovers most downstream performance gains, and (ii) projectors trained on different datasets can be linearly interpolated while largely preserving performance. These results suggest that projector training mainly performs a coarse alignment, which many different projectors achieve equally well. Motivated by these findings, we propose PORTAL, a training-free projector initialization that requires neither paired image-caption data nor gradient-based optimization, but only per-modality summary statistics. PORTAL computes an optimal-transport-based map between Gaussian approximations of the vision and language feature distributions on a shared principal subspace, so that the projector training stage can be skipped entirely. Without any paired data, PORTAL matches the standard pipeline trained on 558k image-caption pairs across six LLM backbones, two vision encoders, and 16 benchmarks. It outperforms that pipeline in $3$ of $7$ (vision encoder, LLM) settings and reaches an average $98%$ of its score across the other four.

[235] MajorTOM:Elliot Pretrain, A Multi-Sensor, Multi-Temporal Pretraining Dataset for Earth Observation cs.CV | cs.LGPDF

Carmen Oliver Huidobro, Mikolaj Czerkawski, Kevin Monsalvez-Pozo, Cesar Aybar, Gustau Camps-Valls

TL;DR: MajorTOM:Elliot Pretrain 提供 27.9 万块多传感器、多时间尺度地球观测数据，并通过短程预训练研究各传感器和辅助目标的价值。

Details

Motivation: 通用地球观测基础模型需要统一多传感器、不同时间采样及文本的数据，但公开数据集很少能以一致、易访问的形式同时提供这些信息。

Result: 数据集含 279,166 个 10 公里瓦片，统一到 1056×1056、10 米网格，包含 Sentinel-2 L1C、Landsat-8/9、Sentinel-1 RTC、Copernicus DEM、ESA WorldCover，以及云掩码、OSM 实体和 ERA5 天气。分为单时相 250,000 瓦片、月度 12,500 瓦片（年内 12 次）和短期 burst 16,666 瓦片（30 天内 6 次），以 6.3TB 云优化 GeoTIFF 和 TACO v3 Parquet 提供。OlmoEarth-Base 短程预训练显示：基于 AlphaEarth 嵌入挑选的 Core-S2L2A 25% 子集在 9 项基准中 8 项达到或超过全量；加入 Landsat-8/9 所有基准均不降；把 DEM/WorldCover 作为预测目标改善 8/9，而作为编码器输入使 5/9 退化；Sentinel-1 在 9 项中提升 7 项，尤其雷达任务；规则 captions 与小语言模型效果相当但无需成本。

Insight: 不仅数据量重要，传感器如何进入预训练目标也影响迁移：DEM/地表覆盖作为预测目标有益，直接输入反而可能干扰。数据发布采用云端流式和像素读取前过滤；实验为短程预训练，不能替代全规模训练的最终模型比较。

Abstract: Pretraining generalist Earth observation foundation models requires data that combine several sensors, varied temporal sampling and, increasingly, text, yet open datasets rarely provide all three in one consistent, easily accessible form. We introduce MajorTOM:Elliot Pretrain, a collection of 279,166 tiles of 10 km drawn from the Major-TOM grid by clustering AlphaEarth embeddings. Each tile stacks Sentinel-2 L1C, Landsat-8/9, Sentinel-1 RTC, Copernicus DEM and ESA WorldCover on a shared 1056x1056-pixel, 10 m grid, complemented by cloud masks, OpenStreetMap entities and ERA5 weather. The tiles are partitioned into three temporal regimes, Monotemporal (250,000 tiles, one acquisition), Monthly (12,500 tiles, twelve acquisitions spread over the annual cycle) and Burst (16,666 tiles, six acquisitions within 30 days), and are released as 6.3 TB of Cloud Optimized GeoTIFFs indexed by TACO v3 Parquet metadata, a standard that allows cloud streaming, efficient data access, and filtering before any pixel is read. A companion library derives sensor-tagged fact sheets, captions and 16 families of vision-language tasks from these layers at load time. We test our design choices in short pretraining runs of OlmoEarth-Base. First, a 25% subset of Major-TOM Core-S2L2A curated with AlphaEarth embeddings matches or exceeds the full corpus on 8 of 9 benchmarks. Second, on Elliot-Pretrain, adding Landsat-8/9 improves or matches Sentinel-2 only pretraining on every benchmark. Third, using DEM and WorldCover as prediction targets during pretraining further improves 8 of 9 benchmarks, while giving the same maps to the encoder as inputs degrades 5 of 9. Fourth, Sentinel-1 helps on 7 of 9 benchmarks, particularly on benchmarks with radar inputs. Fifth, rule-based captions match a small language model at no cost. The dataset is openly available at https://source.coop/major-tom/elliot-pretrain.

[236] VideoJev: System One Model for Video Understanding and Interaction cs.CVPDF

Yuhan Zhu, Xiangyu Zeng, Qingyi Si, Chenxu Yang, Yifei Li

TL;DR: VideoJev 将视频 grounding 和多选 QA 统一为对时间区间与文本选项打分的结构化决策接口，并通过缓存或联合推理复用重复查询的视觉计算。

Details

Motivation: 视觉智能体需要根据视频及时作决定并追踪进展；传统自回归生成答案计算较重，而实际互动会围绕同一视觉历史提出新目标和中间问题，重复编码视频造成浪费。

Result: VideoJev 在预训练 MLLM 上将自回归答案解码替换为共享评分头，直接对时间片段和文本选项与视频证据的匹配打分。新问题到达时可缓存复用视觉表示，也可将多问题合并一次前向。在三个时间定位基准上，4B 和 35B-A3B 版本的平均 mIoU 比各自骨干提高 6.7、10.8 分，测得推理速度提高 2.8–3.1 倍。四个视频 QA 基准上，两者在 Thinking 模式保持与骨干相当的表现，实测速度提升超过 100 倍。

Insight: 把开放式生成转为区间/选项评分，适合需将多个结构化判断组合成工作流的场景；视觉计算复用尤其能放大多查询效率收益。超过 100 倍速度针对视频 QA 推理比较，摘要未给首 token、长答案或不同视频长度下的延迟分解。

Abstract: Visual agents must turn video observations into timely decisions to track progress and choose what to do next. Building on the emerging System One paradigm, we introduce VideoJev, a family of models that cast video understanding and interaction as structured decision tasks. We adapt pretrained multimodal large language models (MLLMs) by replacing autoregressive answer decoding with a shared head that scores temporal intervals and textual options against video evidence. These scores yield event spans or answer choices, unifying temporal grounding and multiple-choice video question answering. This combines the semantic flexibility of MLLMs with efficient decisions that agents can use directly. As an interaction unfolds, new goals and intermediate results prompt further queries about the same visual history. VideoJev therefore reuses visual computation through Cached inference for queries as they arrive and Joint inference for multiple queries in a single forward pass. Across three temporal grounding benchmarks, VideoJev-4B and VideoJev-35B-A3B improve mean mIoU over their backbones by 6.7 and 10.8 points, respectively, with measured inference speedups of $2.8$-$3.1\times$. On four video QA benchmarks, both achieve performance comparable to their backbones in Thinking mode, with measured inference speedups exceeding $100\times$. These results offer a unified, reusable decision interface for composing structured visual judgments into complex, open-ended workflows.

[237] ST-GRU PredRNN-V2: A GRU-Based Reformulation of PredRNN-V2 for Video Prediction cs.CV | cs.AIPDF

Augusta van Haren, Hanna Hoogen, Luca Pattavina, Quincy Rommen

TL;DR: ST-GRU 将 PredRNN-V2 的时空 LSTM 替换为更简单的 GRU 门控单元，在 Moving MNIST 上减少约 5% 参数，但多项指标未超过原 PredRNN-V2。

Details

Motivation: PredRNN 系列用两种记忆状态的 ST-LSTM 捕获长程时空依赖，但循环单元较复杂；研究检验 GRU 的简化门控是否可用于时空预测。

Result: ST-GRU 将每个记忆状态的三门设计替换为更新门、重置门，并将时间记忆并入隐藏状态，使参数量减少约 5%。在 Moving MNIST 原协议、平均预测 10 帧下，ST-GRU PredRNN-V2 的 MSE 为 55.1、SSIM 0.883、LPIPS 0.065、PSNR 19.53 dB。它在 MSE、SSIM、PSNR 上未超过原 ST-LSTM PredRNN-V2，但 LPIPS 更低，并与早期 ST-LSTM PredRNN 相当。

Insight: 更简单的时空单元可在保持部分感知指标的情况下减少参数，但本实验未证明其全面优于复杂单元。结果仅限 Moving MNIST 与单一协议；作者把它视为有前景的构件，而非已取代 ST-LSTM 的结论。

Abstract: Video prediction requires models to capture complex spatiotemporal dependencies and to retain relevant information over long sequences. PredRNN and its successor PredRNN-V2 address this challenge with spatiotemporal Long Short-Term Memory (ST-LSTM) units that maintain two separate memory states, at the cost of a relatively complex recurrent building block. In this work, we investigate whether the simpler gating of Gated Recurrent Units (GRUs) can be transferred to spatiotemporal predictive learning. To this end, we develop a spatiotemporal GRU (ST-GRU) unit as a GRU-based alternative to the ST-LSTM. The ST-GRU replaces the ST-LSTM’s three gates per memory state with two GRU-style gates (update and reset), and merges the temporal memory state into the hidden state, which reduces the number of parameters by about 5%. We incorporate the ST-GRU into PredRNN-V2 and evaluate the resulting model on the Moving MNIST video prediction benchmark, following the original experimental protocol. Averaged over ten predicted frames, the ST-GRU-based PredRNN-V2 achieves an MSE of 55.1, an SSIM of 0.883, an LPIPS of 0.065 and a PSNR of 19.53dB. It does not surpass the original ST-LSTM-based PredRNN-V2 on MSE, SSIM or PSNR, but it achieves a lower LPIPS and performs on par with the earlier ST-LSTM-based PredRNN, while using a simpler recurrent cell. These results suggest that ST-GRUs are a promising building block for spatiotemporal predictive learning and motivate further work on their design and integration.

[238] RSIgen: Reciprocal Self-Improvement for Agentic Visual Generation cs.CVPDF

Yidi Liu, Haozhe Wang, Jie Xiao, Kai Zhu, Xueyang Fu

TL;DR: RSIgen 让视觉生成智能体和生成器围绕任务共同自我改进，通过干预路由学习分配、刷新智能体动作监督并构造递归课程。

Details

Motivation: 统一多模态骨干承担理解智能体和生成器两种角色，但任务成功本身不能说明应更新哪个角色；生成器更新还可能使智能体原有动作监督过时。

Result: 动态责任路由用匹配干预比较证据、输入和可执行动作，决定学习对象并生成兼容监督。生成器更新后，在固定分支输入下重新评估已测试动作的相对效用以刷新智能体监督，再由保留角色构造下一轮课程。在 BAGEL、Qwen-Image、Wan-Image-7B 上，三轮适配后 AgentGen-Bench Full Overall9 比带动作成本的 SearchGen 高 3.2–3.3 分。Wan 控制实验在匹配角色训练量下验证责任分配、在匹配适配预算下验证监督刷新，并显示更新课程来源有益。200 条真实用户请求的盲评中，相对同一基线，平局调整后的偏好分任务完成为 70.0%、视觉质量为 64.9%。

Insight: 生成器和智能体需要共同更新，但生成器能力变化会使智能体监督失配；按干预结果分配角色和周期性刷新动作标签，显式处理这一动态。提升来自三个骨干/三轮特定设置，盲评 200 个请求的结果仍需结合评审协议解读。

Abstract: Agentic visual generation produces experience that can improve the native understanding (agent) and generation roles of a unified multimodal backbone, but task success alone does not determine which role should learn. Generator updates can also make the agent’s action supervision stale. We introduce RSIgen, a task-centric reciprocal self-improvement framework linking learning allocation, supervision refresh, and recursive curriculum construction. Dynamic Responsibility Routing (DRR) uses matched interventions on evidence, inputs, and supported actions to select learning recipients and construct compatible training targets. After generator learning, RSIgen re-evaluates the relative utility of tested actions with branch inputs held fixed to refresh agent supervision; the retained roles then construct the next curriculum. Across BAGEL, Qwen-Image, and Wan-Image-7B, RSIgen improves AgentGen-Bench Full Overall9 by 3.2–3.3 points over a three-cycle SearchGen adaptation with action costs. Wan controls support allocation gains at matched role-level training exposure, refresh gains at a matched adaptation budget, and benefits from updated curriculum sources while recipients continue learning. In blind evaluation on 200 real-user requests, pooled tie-adjusted preference scores against the same baseline reach 70.0% for task completion and 64.9% for visual quality.

[239] PlanPath: Exploration-Guided Orchestration for Patient-Level Pathology Reasoning cs.CV | cs.MAPDF

Jiusong Ge, Mingzhao He, Qidong Liu, Chunbao Wang, Chunze Yang

TL;DR: PlanPath 依据患者级病理问题规划跨切片证据采集，并从组织结构逐层细化到细胞，聚合多个标本的计数、比例和范围后生成可追溯答案。

Details

Motivation: 一位患者可能有数十张来自不同标本和解剖部位的切片，每张包含互补证据；既有病理智能体多逐张分析并描述形态，较少研究跨标本取证和患者级定量。

Result: PlanPath 将问题转成探索计划，路由到可能包含所需证据的切片，并以组织级、细胞级工具逐步采集数量、比例和范围，跨切片汇总到答案。作者构建肾癌和胃癌两个患者级 VQA 数据集，平均每位患者 15 张全切片图像，问答由病理医师依据诊断报告核验。PlanPath 在两套数据集总体准确率比最强基线分别高 9.4%、12.1%，接入该框架也进一步改善既有切片级智能体。

Insight: 患者级回答不仅需检索相关切片，也需将局部量化证据整合并保留来源轨迹；粗到细工具链支持这种证据聚合。摘要未提供病例数、病种分布、各类问题准确率或外部医院验证结果。

Abstract: A patient’s pathology case spans tens of slides from distinct specimens and anatomical sites, each carrying non-redundant, site-specific evidence. Forming a patient-level conclusion therefore requires locating the slides relevant to a given question and integrating complementary findings across them. However, existing pathology agents analyze one slide at a time and describe morphology rather than measure it, leaving evidence acquisition across a patient’s specimens and patient-level quantification underexplored. To address this gap, we propose PlanPath, an agentic framework for patient-level pathology visual question answering (VQA). PlanPath turns the question into an exploration plan that specifies the evidence required, routes to the slides likely to carry it, and drives coarse-to-fine acquisition from tissue architecture down to individual cells. Tissue- and cell-level tools supply the counts, proportions, and extents, which PlanPath aggregates across slides into a traceable answer. To enable evaluation, we build two patient-level VQA datasets for renal and gastric cancer, averaging 15 WSIs per patient, with pathologist-verified question-answer pairs derived from diagnostic reports. Experimental results show that PlanPath outperforms the strongest baseline by 9.4% and 12.1% in overall accuracy on the two datasets and further improves existing slide-level agents when plugged into our framework. The dataset and code are available at: https://github.com/JSGe-AI/PlanPath_1.

[240] GeoReader: Learning Latent Geometry for Spatial Reasoning in Vision-Language Models cs.CV | cs.AI | cs.LG | cs.MMPDF

Rong Li, Shijie Li, Jiangyong Huang, Lingdong Kong, Jiaming Zhou

TL;DR: GeoReader 训练轻量几何读取器，将几何信息编码成目标 VLM 可直接利用的连续 token，再在问题条件下注入空间问答，而非只追求能从潜表示重建几何。

Details

Motivation: VLM 难从 RGB 隐式关系推断公制和多视角几何。已有方法把几何特征融合后只用 QA 学习使用，或对齐 VLM 潜状态与几何特征；然而表示中“含有几何”不等于接收模型真正使用它。

Result: GeoReader 先用冻结 VLM 解码的几何描述监督轻量读取器生成连续输入 token，使其表达形式贴合目标模型已有能力；随后冻结读取器并微调 VLM 做空间 QA，按问题条件注入几何。几何描述只作为训练目标，推理时直接从连续 token 作答，无需生成几何 CoT。基于 RGB 预测几何，Qwen3-VL-4B 在 MindCube-Tiny 得分 94.3；Qwen3.5-9B 加多源 QA 混合后，在另六个空间基准平均 72.5。

Insight: 几何表示的价值取决于接收模型能否使用，而不仅是表示本身是否可解码；冻结读取器再训练 VLM，显式建立可读接口。两项成绩使用不同骨干和评测集，摘要未给基线增益数值或几何预测误差与 QA 的关系。

Abstract: Vision-language models (VLMs) struggle with metric and multi-view relationships that remain implicit in RGB observations. Geometry foundation models recover this structure from RGB, but existing methods either fuse the estimated features with visual tokens and learn to use them from question answering alone, or align VLM latents with geometry features. Capturing geometry, however, does not ensure its use: zeroing the aligned latents of a released feature-alignment model leaves its accuracy unchanged. We introduce GeoReader, a framework that learns geometric representations for the VLM that reads them. First, we train lightweight Geometric Readers to produce continuous input tokens, supervising them with geometry descriptions decoded by a frozen VLM, so that the Readers express geometry in a form the target model can already read. We then freeze the Readers to preserve this interface while fine-tuning the VLM on spatial QA, with geometry injection conditioned on the question. Geometry descriptions serve only as training targets: at inference, the VLM answers directly from the continuous geometric tokens without generating a geometry chain of thought. With geometry predicted from RGB images, our approach reaches 94.3 on MindCube-Tiny with Qwen3-VL-4B and, with a multi-source QA mixture, an average of 72.5 across six additional spatial benchmarks with Qwen3.5-9B. These results suggest that a useful geometric representation is one the receiving model can use, not merely one from which geometry can be reconstructed.

[241] Segment Where You Look: Attention as the Interface Between Language Models and Segmenters cs.CVPDF

Eslam Abdelrahman, Hamza Omar Alkaf, Yuheng Li, Siddharth Srinivasan Iyer, Nanxuan Zhao

TL;DR: AttnSeg 用 [SEG] token 对图像 patch 的注意力图直接作为分割解码器的空间提示，替代缺乏空间索引的单一隐藏向量。

Details

Motivation: 语言引导分割通常把目标位置压缩为一个 [SEG] 隐状态向量，再交给并未看到表达的像素解码器重建目标范围；但 VLM 本身已计算定义在 patch 网格上的 [SEG] 图像注意力。

Result: 冻结分割 VLM 的探针实验发现，同一探针读取 [SEG] 注意力，比读取 [SEG] 隐状态高 18–22 gIoU；注意力从隐状态获益的图像中未进一步受益，且主要集中在后层。AttnSeg 用学习到的有符号注意力头组合形成稠密提示，并直接监督该提示。两阶段训练先让 VLM 不带解码器学习定位，再训练解码器细化。在 Sa2VA 上，4B 模型 ReasonSeg 提高 6.4 gIoU、GS-Eval 提高 10.2；相较同数据微调模型分别高 4.2、9.6。14B 上分别提高 8.8、11.3。物体部件、多目标和单物体任务增益为 13–18 gIoU；stuff 区域几乎无增益，因为解码器本可恢复。

Insight: 空间注意力比无空间索引的语义向量更适合作为分割接口，并可揭示目标在 patch 网格上的定位证据。收益集中于需要精确指称和区分目标的类别，stuff 区域基本不变；需注意报告的 18–22 是相同探针下的预实验差异，与最终模型基准增益不同。

Abstract: Vision-language models (VLMs) that segment from language must pass the location of the referent to a pixel decoder. The dominant interface compresses it into the hidden state of a single [SEG] token, a d-dimensional vector with no spatial index, and leaves the decoder, which never sees the expression, to recover the referent’s extent. We show that the VLM already computes a better signal: the [SEG] token’s attention over the image tokens, which is defined on the patch grid by construction. Probing a frozen segmentation VLM, we find that this attention yields 18 to 22 more gIoU than the [SEG] hidden state when read by the same probe, that it gains nothing from the image where the hidden state does, and that it is concentrated in the late layers. Building on these findings, we introduce AttnSeg, which replaces the [SEG] embedding with this attention as the decoder’s prompt. A learned, signed combination of the attention heads forms a dense spatial prompt, and an attention-guidance objective supervises it directly. A two-stage recipe first trains the VLM to localize without the decoder, then trains the decoder to refine, within the budget of standard fine-tuning. Applied to Sa2VA, AttnSeg improves ReasonSeg by 6.4 gIoU and GS-Eval by 10.2 at 4B (4.2 and 9.6 over the same model fine-tuned on the same data), and by 8.8 and 11.3 at 14B. The gains reach 13 to 18 gIoU on object parts, multiple objects, and single objects, where the referent must be singled out, and are negligible on stuff regions, which the decoder already recovers unaided. Code and models are available at https://eslambakr.github.io/attseg_project_page/

[242] Concealed Handgun Detection using WiFi Signals cs.CVPDF

Erik Persson, David Gustafsson, Tomas Wilkinson

TL;DR: WiFire 使用 WiFi 信道状态信息检测遮挡或隐藏手枪，并提供与 RGB 图像对齐的数据集及视觉监督模型 WiFireNet。

Details

Motivation: 摄像监控难以识别遮挡目标，金属探测器通常需要人员逐个通过狭窄入口；在拥挤场景中二者都可能受限，WiFi 信号可作为补充感知模态。

Result: 作者收集对齐的 WiFi 信号和 RGB 图像构成 WiFire 数据集，并提出以视觉监督训练的 WiFireNet，从 WiFi CSI 检测隐藏手枪。实验显示仅凭 WiFi 信号可以检测隐藏手枪，作者认为其有望补充摄像头和金属探测器。摘要没有报告数据规模、检测准确率、误报率或不同遮挡场景结果。

Insight: 该工作探索不依赖可见光的隐藏目标检测，但目前摘要只支持“可行性”结论；安全部署还需重点考察误报、不同环境与设备差异，以及隐私和实际安检流程影响。

Abstract: While conventional camera-based surveillance systems and metal detectors can provide a sufficient level of security when used appropriately, they have fundamental limitations that reduce their applicability. Camera-based systems are unable to detect occluded or concealed threats and metal detectors require each person to enter through a narrow opening one at a time, leading to congestion issues and limited use in large crowds. A WiFi-based security system has the potential to be a complementary modality that is able to handle occlusions or concealment and is not hindered by strict spatial configurations. In this paper, we propose use Channel State Information estimated from WiFi-signals to detect concealed handguns. To this end we collect a dataset of aligned WiFi signals and RGB images that we call the WiFire dataset. We also propose a deep learning model called WiFireNet tailored to detect concealed handguns in WiFi signals using vision-based supervision signals. Our results show that concealed handgun detection is possible in WiFi and has promise as a complementary modality to surveillance cameras and metal detectors. We release the dataset, model and code to facilitate the further research into WiFi-based handgun detection.

[243] Context-aware Bayesian Fusion for Multimodal Object Detection cs.CVPDF

Arunachalam Venkatachalam, Mhafuzul Islam, Marwan Yusuf, David Van Hamme, Ben Stoffelen

TL;DR: 该方法用环境上下文门控网络动态校准预训练相机与 LiDAR 检测器输出，无需为不同雾天条件重新训练基础检测器。

Details

Motivation: 恶劣天气和低照度会降低行人及道路参与者检测可靠性；按运行环境重训检测器需要大量标注、算力和时间。

Result: 上下文感知晚期融合将轻量门控网络接入概率校准框架，估计环境条件并动态调制预训练检测器输出。在不同雾浓度增强的 KITTI 修改版上，作者报告相较非自适应晚期融合基线，相对提升约 14%，不同雾级下平均精度维持在 56%。摘要未给各天气子集精度及参与融合的检测器名称。

Insight: 把环境变化适配放在检测器输出融合层，可以避免重训各模态主干；概率校准与上下文门控共同调节传感器可信度。结果来自雾级增强 KITTI，不能直接推断到真实雾、雨、雪或低照度条件。

Abstract: Reliable detection of pedestrians and road users is essential for the safe operation of autonomous vehicles. However, this task becomes increasingly challenging under adverse environmental conditions, such as poor weather and low lighting, where the performance of state-of-the-art pre-trained object detectors and on-board sensors tends to degrade, which limits their effectiveness and reliability. Moreover, retraining models for each operating context imposes significant constraints due to high data requirements, computational resources, and time. Considering this, we propose a novel context-aware multimodal late fusion approach that integrates a lightweight gating network within our probabilistic calibration framework. The gating network extracts environmental context and dynamically modulates the detection outputs of pre-trained object detectors based on the inferred context, without retraining of the detectors. Our methodology is evaluated on a modified KITTI dataset, in which the camera images and lidar pointclouds are augmented for varying fog levels. Experimental results show that our approach yields an approximately 14% relative improvement over non-adaptive late fusion baselines, achieving a consistent average precision of 56% across different fog levels.

[244] GAD: Geometry Attention Distillation for Consistent Video Generation cs.CVPDF

Junfeng Ni, Ziyu Zhu, Fan Jiang, Yu Liu, Chiyu Wang

TL;DR: GAD 将冻结几何模型在干净视频中的跨帧对应注意力蒸馏到视频扩散模型，并按噪声步选择蒸馏层，以提升长程场景和动态对象一致性。

Details

Motivation: 视频扩散模型短时生成画面质量较好，但长时间生成时常无法保持场景几何和对象结构；现有注意力对跨帧对应的响应分散，难以稳定传递结构信息。

Result: 几何模型抽取同一静态表面或运动对象跨帧对应关系，并监督视频注意力；考虑高噪声时结构先在深层形成、低噪声时在浅层形成，蒸馏层随噪声时间步变化。三个视频骨干实验均报告长程场景及动态时间一致性改善，且不同生成长度下有几何收益。Wan2.2 5B 上两个长程基准的跨帧匹配准确率至少比 Wan 基线高 30%；无需额外推理模块。

Insight: 固定层蒸馏可能与不同去噪阶段的结构形成位置不匹配，按噪声调节层级有助于将几何对应注入适当阶段。至少 30% 是跨帧匹配指标的提升，摘要未提供各基准绝对值或感知质量取舍。

Abstract: Video diffusion models generate visually compelling short-term sequences but often struggle to preserve scene geometry and object structure over time. In this paper, we first demonstrate that current video attention cannot identify reliable correspondences across frames, with responses dispersed across multiple regions even after recognizable scene structure has emerged. We argue that reliable cross-frame associations are essential for information exchange as scene structure forms in video generation. Thus, we introduce Geometry Attention Distillation (GAD), which transfers correspondences from a frozen geometry model’s attention to video attention. Extracted from clean videos, these correspondences link the same static surface or moving object across frames. Since structure formation begins in deeper layers at higher noise and shallower layers at lower noise, we apply attention distillation at a student layer selected according to the noise timestep. After training, video attention focuses more consistently on corresponding locations. Experiments across all three video backbones demonstrate improved long-range scene consistency and dynamic temporal consistency, with geometric gains across generation lengths. On Wan2.2 5B, GAD improves cross-frame matching accuracy by at least 30% over the Wan baseline on both long-range benchmarks. Notably, GAD achieves these gains without additional inference modules, offering a promising approach to efficient and consistent video-based world simulation. Project page: https://fantasy-amap.github.io/GAD/.

[245] DuplexBridge: Coordinating Understanding and Generation for Unified Image Editing cs.CVPDF

Yidi Liu, Jie Xiao, Chaojie Mao, Jingfeng Zhang, Xueyang Fu

TL;DR: DuplexBridge 用统一模型的理解能力为其图像编辑生成角色构造源图锚定条件，并通过不同的局部监督和互信门控平衡请求理解、编辑执行与内容保留。

Details

Motivation: 复杂图像编辑要求准确理解任务并忠实执行，但失败结果无法表明是任务解释还是生成环节出了问题；统一系统还需确保未要求修改的内容保留。

Result: 桥接表示明确目标、最终变化和保留约束。理解模块只按原请求忠实度监督、不观察生成图像；生成模块按其实际接收条件监督执行。互信检查门控选择性接纳全局任务反馈，同时保留各模块独立有效的局部监督。作者构建 CE-Bench，含 1,200 个同时测量变化与保留要求的任务。三种配置中，完整系统比直接编辑器的完全合规率高 21.75–25.75 个百分点，并在 KRIS-Bench、RISEBench、ImgEdit、GEdit-Bench 上改善。统一 Wan-Image 监督初始化下，门控联合训练完全合规率为 77.08%，无门控为 70.92%；保留两门但让生成局部监督检查原始任务则为 75.00%。

Insight: 把理解和生成的责任分开监督，并选择性共享全局反馈，可减少错误归因及过度改动；CE-Bench 同时要求满足编辑和保留约束，避免只看目标是否出现。完全合规提升是在作者协议下，摘要未报告人评规模和不同编辑难度分项。

Abstract: Complex image editing requires faithful task interpretation and accurate execution, but a failed edit does not reveal which responsibility failed. We present DuplexBridge, which uses a unified model’s own understanding capability to enhance prompts for its own generation capability, without an external understanding module. A source-grounded bridge specifies targets, final-state changes, and preservation constraints. Understanding is supervised on fidelity to the original request without observing the generated image; generation is supervised on execution of the condition it actually receives. Reciprocal trust checks selectively admit global task feedback to each role while retaining independently valid local supervision. We introduce CE-Bench, a 1,200-task benchmark measuring joint satisfaction of task-derived change and preservation requirements. Across three model configurations, the full system improves complete compliance over direct editors by 21.75-25.75 percentage points, with gains also observed on KRIS-Bench, RISEBench, ImgEdit, and GEdit-Bench. From a common supervised Wan-Image initialization, gated joint training reaches 77.08% complete compliance, compared with 70.92% without gates and 75.00% when generation-local supervision instead checks the original task with both gates retained. These results support distinct local responsibilities and selective shared feedback for joint understanding-generation adaptation.

Xuyang Liu, Shihao Wang, Xinyu Wei, Minghan Li, Zeqing Wang

TL;DR: OmniCompass 将全模态视频评测拆为视频语言与音视频两条轨道，每条覆盖时间定位、空间定位和问答，并用文本与文本+图像配对目标检查模型指称一致性。

Details

Motivation: 现有视频基准通常只覆盖部分能力，难以系统诊断全模态视频理解；也较少同时评测何时出现、位于何处及发生什么。

Result: OmniCompass-VL 含 820 个视频、2,460 个实例，覆盖不同领域和短/中/长时长；OmniCompass-AV 子集含叙事和音频丰富的 240 个视频、720 个实例。VL 问答涵盖 5 类感知和 3 类推理，AV 问答涵盖 3 类感知和 2 类推理。每个样例在文本-only 与文本+图像两种配对设置下评测，同一目标由文字或视觉语言指引指定；所有实例经人工多轮质检。对 50 个 MLLM 的评估发现开源模型落后于专有模型，尤其音视频理解；即便领先模型在时空定位上仍有困难，且对配对目标指称会给出不一致预测。

Insight: 配对目标指称可诊断模型是否真正理解目标，而不是只依赖某一输入形式；时空 grounding 与问答并列也利于分解能力。摘要未列各模型具体分数或人工标注者数量，AV 子集规模也小于 VL 主轨。

Abstract: With the rapid development of multimodal large language models (MLLMs), evaluating their video understanding capabilities has become increasingly important. Existing video understanding benchmarks cover only part of the evaluation dimensions, limiting systematic diagnosis of omni-modal video understanding. We introduce OmniCompass, a holistic benchmark for omni-modal video grounding and understanding. OmniCompass defines two evaluation tracks: OmniCompass-VL assesses vision-language models on tasks requiring visual evidence, while OmniCompass-AV assesses omni-modal models on tasks requiring joint audio-visual evidence. Each track includes three complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA), assessing when a referenced target appears, where it is, and what happens. QA covers five perception and three reasoning categories in OmniCompass-VL, and three perception and two reasoning categories in OmniCompass-AV. OmniCompass-VL contains 820 videos and 2,460 instances across domains and three duration types: short, medium, and long. A narrative- and audio-rich subset forms OmniCompass-AV with 240 videos and 720 instances. Each instance is evaluated under two paired settings: text-only and text+image, specifying the same target through text and visual-language guidance. All instances are manually annotated with multi-round quality verification. Evaluations of 50 MLLMs show that open-source models still lag behind proprietary counterparts, particularly in audio-visual understanding. Precise temporal and spatial grounding remains challenging even for leading models, and their predictions still exhibit inconsistencies across the paired target specifications.

[247] Game2Real-Bench: Do Toy Games Reveal and Transfer to Real-World Visual Reasoning? cs.CVPDF

Qingan Liu, Juntong Feng, Yue Zhu, Yujie Cheng, Xinzhe Han

TL;DR: Game2Real-Bench 比较程序化玩具游戏与现实视觉推理能力及迁移，发现模型在两类任务上的能力排序高度相关，但游戏示例对不同能力产生相反迁移。

Details

Motivation: 真实图像保留自然复杂度，程序化游戏可生成可控任务和可验证答案；尚不清楚玩具游戏测出的能力是否对应现实推理，游戏示范是否能迁移。

Result: 基准覆盖细粒度感知、空间推理、约束推理、顺序规划四类能力，含 1,000 个可执行验证游戏样本和 400 个按共享推理需求而非视觉相似度配对的现实样本。25 个 VLM 的游戏—现实能力一致性较高，Pearson 和 Spearman 相关分别为 0.935。对四个模型的一次性游戏到现实迁移实验显示，约束推理对部分模型改善，而细粒度感知在提供游戏示范后下降。另构建近 24,000 个验证游戏样本，覆盖 18 项任务、65 类问答。

Insight: 玩具任务能力与现实视觉推理的模型间排序高度相关，但高相关不意味着演示迁移对所有能力都有益；约束推理和细粒度视觉感知的迁移方向不同。仅四个模型参加迁移实验，样本量和提示格式对结论的影响需进一步验证。

Abstract: Visual reasoning is an important capability for vision-language models (VLMs) to understand visual evidence, spatial structure, constraints, and state changes. Existing evaluations mainly rely on real-world data or procedurally generated tasks. Real-world data preserve natural complexity, while procedural games provide controllable generation and verifiable solutions. However, it remains unclear whether abilities measured in toy games correspond to real-world visual reasoning and whether game examples can support transfer. We introduce Game2Real-Bench, a benchmark that studies these connections across four capability groups, including fine-grained perception, spatial reasoning, constraint reasoning, and sequential planning. We construct 1,000 procedurally generated game samples with executable verification and curate 400 real-world samples aligned by shared reasoning requirements rather than visual similarity. We evaluate 25 representative VLMs and observe strong game-to-real consistency, with Pearson and Spearman correlations of 0.935. We further study one-shot game-to-real transfer on four models. Constraint reasoning improves for several models, while fine-grained perception declines with game demonstrations. We also construct Game2Real-Dataset with nearly 24K verified game samples across 18 tasks and 65 QA families. Our benchmark, dataset, and code are available at https://github.com/QingAnLiu/Game2Real-Bench.

[248] Plan-Bench: Can Existing VLMs Go Well Beyond Perception to Reason and Plan in Embodied Worlds? cs.CVPDF

Juntong Feng, Qingan Liu, Jingru Dong, Yue Zhu, Weiming Zheng

TL;DR: Plan-Bench 将 VLM 具身规划评测与动作执行隔离，考察其在物理任务中的空间整合、时间状态推理和约束下多步计划能力。

Details

Motivation: VLM 在视觉和空间理解上表现较强，但物理任务还需整合空间信息、推理状态变化并制定多步计划；需要区分规划能力本身与执行器成功率。

Result: 基准覆盖三个维度、16 个子类和 699 个精选样本，并尽量隔离动作执行因素。对 28 个 VLM 的评测显示仍有明显改进空间；即使最强模型准确率也只有 75.0%。作者还发现模型规模扩大通常有助于这些能力，消融表明大模型作为高层规划器的潜力更大。摘要未提供各维度具体成绩。

Insight: 即使不考虑低层动作执行，现有模型在具身世界高层推理与计划上仍未达到可靠水平。699 样本的轻量探针有利于比较，但具体任务分布和准确率不能直接代表真实机器人执行能力。

Abstract: Vision-Language Models (VLMs) have achieved strong performance in visual perception and spatial reasoning, leading to growing interest in using them as high-level planners for embodied AI. However, physical tasks require models to integrate spatial information, reason over temporal states, and generate multi-step plans under various constraints, raising the question of whether current VLMs can reliably support such reasoning and planning independently of execution. We introduce Plan-Bench, a benchmark for evaluating the high-level reasoning and multi-step planning capabilities of VLMs relevant to embodied worlds, while isolating the influence of action execution. It covers three complementary dimensions, spanning 16 subclasses and 699 curated samples. Extensive evaluation of 28 VLMs reveals room for improvement, with even the strongest model achieving only 75.0% accuracy. We further find that model scaling generally improves these capabilities. Ablation studies also indicate larger models show increasing potential as high-level planners. Overall, Plan-Bench provides a lightweight probe for evaluating VLMs’ reasoning and planning capabilities relevant to embodied environments and assessing their potential as decision makers for physical tasks. Our benchmark and code are available at https://github.com/MinneFeng/Plan-Bench.

[249] RepProbe: Can Video Language Models Count Repeated Actions? A Diagnostic Benchmark cs.CVPDF

Mahsa Khoshnoodi, Sarah Adel Bargal

TL;DR: RepProbe 用含 1–20 次完整重复动作的真实与合成视频诊断 VLM 计数能力，发现模型内部仍编码次数信息，但最终输出常坍缩到少数偏好的数字。

Details

Motivation: 视频语言模型能流畅描述动作，却不一定能准确识别并计数时间上重复出现的动作循环。

Result: 基准含 42,000 段视频，控制重复次数、片段位置、时间覆盖和持续时长。各模型在较高重复次数上的计数准确率明显下降，控制实验显示单靠帧采样不足以解释。对比最终层表示线性探针与模型答案发现，即使模型报错，重复次数信息仍可线性解码；错误主要发生在输出阶段，模型答案集中到少量偏好数字，而非输出表示中已有的信息。该表示—输出分离在真实和合成视频上均成立，且不能由时长、位置或场景内容解释。摘要未列各模型具体准确率。

Insight: 计数失败可能不是视频表征缺乏周期信息，而是模型无法将内部信息可靠映射为数值答案；因此只提升视觉编码未必能解决计数。结论来自特定重复动作基准，线性可解码不等于模型可实际利用信息。

Abstract: Repetition counting requires identifying and enumerating action cycles over time. While state-of-the-art video language models describe actions fluently, they fail to reliably count how many times an action repeats. To investigate this limitation, we introduce RepProbe, a benchmark of 42K real and synthetic video clips containing exactly k in [1, 20] complete repetitions, with controls for clip position, temporal coverage, and duration. These controls allow us to study the counting failure directly and locate where in the model it arises. Counting accuracy collapses at high counts across the video language models we evaluate, and controlled experiments show this collapse is not explained by frame sampling alone. We compare model outputs with linear probes trained on the final-layer representations used to generate those outputs. We find a consistent representation-to-output dissociation: substantial count information remains linearly decodable even when models fail to report the correct count. Further analysis shows this failure arises primarily at the output stage, where answers concentrate on a small set of preferred numbers rather than the available count information. The dissociation holds on both real and synthetic videos and is not explained by clip duration, position, or scene content. These results indicate that models encode repetition structure but do not reliably report it in their outputs. Project page: https://arenaa.github.io/RepProbe-webpage/

[250] Visual Tokens on Demand: Visual Budgeting for Scalable and Efficient Multi-Image Reasoning in Multimodal LLMs cs.CVPDF

Aimon Rahman, Joymallya Chakraborty, Vidit Bansal, Abhishek Dan

TL;DR: mmVB 为多图推理按图像分配查询条件化 token 预算，利用跨图上下文决定压缩比例，无需重训基础 MLLM 或在推理时读取注意力矩阵。

Details

Motivation: 多图输入中每张图增加大量视觉 token，编码和注意力成本高；单图压缩方法未充分处理跨图冗余、稀疏证据及比较、计数、时间推理所需的局部细节，部分方法还依赖难以访问的注意力矩阵或重训基础模型。

Result: mmVB 是轻量可插拔模块，结合跨图上下文为每张图分配与查询相关的 token 预算，新增 160 万可训练参数，将视觉输入 token 降低 9 倍。作者报告首 token 时延降低 5 倍、预填充 FLOPs 减少 7 倍，同时在八个多图推理基准上保持完整分辨率准确率的 94.4%。摘要未列模型、基准名称及不同压缩率成绩。

Insight: 跨图信息可帮助识别冗余并为各图像动态分配视觉预算，避免每图固定压缩率误删比较所需证据。94.4% 是八基准总体保留表现；摘要未说明多轮生成速度、预算分配失效情形或新增模块训练数据。

Abstract: Multimodal large language models have advanced rapidly, but most remain optimized for single-image inputs. Multi-image reasoning is costly because each image adds many visual tokens, increasing encoding and attention demands. Existing techniques that address the above issue, including visual token pruning, token merging, visual token compression, learned resampling, hierarchical compression, retrieval-augmented visual reasoning, and memory-based architectures, have largely been designed for single-image scenarios, where redundancy is primarily spatial and local. Multi-image reasoning introduces additional challenges: redundancy may occur across images, task-relevant evidence may be sparse or distributed, and aggressive per-image compression can remove details needed for comparison, counting, temporal reasoning, retrieval, or evidence grounding. Moreover, existing token-reduction methods are often either (i) training-free but inference-inefficient, typically relying on accessing attention matrices to get token-importance that may be impractical to access with memory-efficient attention implementations such as FlashAttention, or (ii) training-dependent, requiring the base MLLM to be retrained or adapted to a new compressed-token paradigm. In this work, we study scalable multi-image reasoning and understanding without retraining the base model and without accessing attention matrices at inference time. To this end, we propose mmVB, a lightweight plug-and-play module that assigns query-conditioned per-image token budgets via cross-image context, adding only 1.6M trainable parameters while reducing visual input tokens by 9x. This reduction yields 5x faster time-to-first-token and 7x fewer prefill FLOPs, while maintaining 94.4% of full-resolution accuracy across eight multi-image reasoning benchmarks.

[251] PLAYGROUND: Progressive Layout Generation with Render-Grounded Planning cs.CVPDF

Jaeho Lee, Junhwan Heo, Junghyun Park, WonJun Moon, Eunju Yang

TL;DR: PlayGround 将可编辑图形布局生成为渐进反馈循环：检索器找结构相似参考，规划器根据当前渲染画布组织下一功能组，生成器预测可编辑属性，验证器评估并反馈修订。

Details

Motivation: 迭代生成和视觉细化更接近真实设计流程，但许多方法在执行前就固定高层计划，无法适应画布和剩余元素随生成逐步变化。

Result: 输入为无序设计元素。Structure-Aware Retriever 先推断粗布局结构并检索相似设计；每轮 Progressive Planner 查看渲染画布和剩余元素，规划下一功能组构图，Layout Generator 生成可编辑布局属性，Visual Verifier 评估渲染候选并在必要时提出反馈。作者构建两阶段数据生成流程，从完成设计中提取功能组和逐轮计划轨迹，用于步骤级训练。在 Crello 和 LICA 上，实验报告优于既有方法。摘要未给具体指标或增益数值。

Insight: 根据实际渲染结果动态规划，能把生成计划与当前画布闭环连接，而不是一次性执行固定蓝图；检索、规划、布局预测和验证分工也服务可编辑性。结果目前定性，摘要未量化推理轮次、设计质量或与非渐进流程的成本差异。

Abstract: Generating editable graphic designs requires arranging visual and textual elements into a coherent composition while preserving their content and editability. Although recent MLLM-based approaches employ iterative generation and visual refinement to reflect the real-world design workflows, their high-level plans are predetermined prior to execution, limiting their adaptability as the design evolves. To address this limitation, we propose PlayGround, a progressive layout generation framework that dynamically adapts to the current rendered canvas. PlayGround establishes a progressive design loop comprising a Progressive Planner, a Layout Generator, and a Visual Verifier, guided by a Structure-Aware Retriever. Given an initial set of unordered elements, the Structure-Aware Retriever initially infers a coarse layout structure to retrieve structurally similar reference designs. Subsequently, at each turn, the Progressive Planner examines the current rendered canvas alongside the remaining elements, forms the next functional group, and plans its composition. The Layout Generator then predicts editable layout attributes, while the Visual Verifier evaluates rendered candidates and provides feedback when refinement is necessary. To train the Progressive Planner with step-wise supervision, we construct a two-stage data generation pipeline that extracts functional groups and turn-level planning trajectories from completed designs. Experiments on Crello and LICA demonstrate the effectiveness of PlayGround over existing methods. Our code and weights will be publicly released.

[252] Cepheid: Decoupling Scale from Structure for Cross-Family Motion Completion cs.CV | cs.GR | cs.LGPDF

Chih-Hai Su, Tsung-Wei Ke, Yu-Lun Liu

TL;DR: Cepheid 将运动残差的结构与幅度解耦，学习无量纲运动先验，并从可观测量估计未知物理尺度，以改善跨几何家族的稀疏到稠密网格运动补全。

Details

Motivation: 从稀疏顶点位移恢复稠密网格运动时，单一几何家族训练的残差先验常在新家族上不如无参数调和扩展；残差幅度还会随几何家族变化一个数量级以上，并在观测更密时塌缩，测试时目标运动尺度未知。

Result: 方法学习无量纲残差结构，从可观测量估计物理幅度，并只用训练源拟合轻量修正。统一模型在所有评估几何家族上超过调和扩展，包括未见布料和腿足身体；无需骨架、配准或跨网格对应。相同先验无需重训即可用于点跟踪动物视频和机器人抓取软物体。摘要未提供基准数值或误差改善幅度。

Insight: 跨家族失败不仅是结构先验迁移问题，也源于残差尺度随目标和观测密度变化；分开学习结构与尺度可更稳健地复用。摘要未给不同家族的定量结果，真实视频/机器人迁移也未说明测试样本规模。

Abstract: Sparse-to-dense motion completion recovers dense mesh motion from sparse vertex displacements. Residual priors trained on one geometry family frequently fail to beat parameter-free harmonic extension off-family. Moreover, cross-family training meets a scaling barrier: residual amplitude varies by more than an order of magnitude across geometry families and collapses as observations become denser. Since this scale depends on unknown target motion, it cannot be measured directly at test time. We decouple residual structure from physical scale by learning a dimensionless prior, estimating amplitude from observables and a lightweight correction fitted only on training sources. Our unified model surpasses harmonic extension across all evaluated families, including unseen cloth and legged bodies, without skeletons, registration, or cross-mesh correspondence. The same prior applies without retraining to point-tracked animal video and a robot grasping a soft body. Project page: https://su-terry.github.io/cepheid-website/

cs.CL [Back]

[253] PsyBer-Agent: A Psychology-Driven User Agent for Recommender Systems cs.CL | cs.IR | cs.LGPDF

Chong Zhang, Qijie Shen, Haichao Zhang, Can Wang, Xiaobo Jin

TL;DR: PsyBer-Agent 用三个演化潜在状态表示用户的疲劳、情绪和偏好变化，并用非平衡 Gromov–Wasserstein 对齐真实日志与模拟轨迹；在所测长时程中比仅提示模拟器行为偏差更低。

Details

Motivation: 推荐系统上线前需要在长期闭环交互中评估，但现有 LLM 用户智能体多模仿短期表面回复，难以表达疲劳、情绪和偏好漂移等潜在动态。

Result: PsyBer-Agent 的 Psy-Engine 为每名用户建模三个随交互演化的潜在状态。为将模拟行为对齐真实互动日志，作者用非平衡 Gromov–Wasserstein 最优传输，将日志条件化参考轨迹与模拟器生成的潜状态轨迹对齐，以此结构化监督轻量策略，而无需微调大型主干 LLM。作者还构建包含多模态场景和七种推荐主干的 WebSim 与 PsyBer Benchmark，评估行为保真度、潜状态一致性、稳健性及社会线索扰动下的响应。在所测交互时程中，PsyBer-Agent 的行为差异低于仅提示模拟器；在并发受限情况下可稳定支持最多 5M 排队模拟任务。

Insight: 用潜状态轨迹对齐为模拟用户引入长期心理动态，不仅拟合单次话语表面相似度；轻量策略避免微调整个主干的开销。5M 指排队任务数量，不等于同时并发用户数；摘要未给具体行为误差指标。

Abstract: Evaluating recommender systems before deployment requires simulators that remain realistic over long, closed-loop interaction cycles. Current LLM-based agents generally function as surface imitators that generate realistic short-term responses but fail to capture the underlying user dynamics such as fatigue, affect, and shifting preferences. We propose PsyBer-Agent, a psychology-inspired simulator that models these dynamics through its Psy-Engine. This engine represents each user with three evolving, model-defined latent states. To calibrate simulated behavior against real interaction logs, we align log-conditioned reference trajectories with simulator-generated latent-state trajectories via Unbalanced Gromov–Wasserstein optimal transport. This approach provides structured supervision for a lightweight policy without the computational burden of fine-tuning the backbone LLM. We also introduce WebSim and the PsyBer Benchmark, which encompass multimodal scenarios and seven recommender backbones. Our framework evaluates behavioral fidelity, latent-state consistency, robustness, and responses to social-cue perturbations. On the tested horizons, PsyBer-Agent shows lower behavioral discrepancy than the prompt-only simulators while maintaining stable performance under large-scale concurrent simulation for up to 5M queued simulation tasks under bounded concurrency. Code is available in the supplementary material and at: PsyBer-Agent: https://anonymous.4open.science/r/CC66F/, WebSim Platform: https://anonymous.4open.science/r/D8EAX/.

[254] Loose Lips: Understanding Privacy Leakage in Agentic Dialogue cs.CL | cs.AIPDF

Soumya Sharma, Ayana Hussain, Heber Hwang Arcolezi, Ulrich Aïvodji, Spandana Gella

TL;DR: 基于情境完整性理论的 4,410 条双智能体对话评测发现，能力更强的模型也更易泄露隐私，泄露随对话由 1 轮增至 6 轮最多扩大 3 倍。

Details

Motivation: 企业高风险场景中的智能体会处理敏感信息，评估是否泄露必须考虑信息产生时的上下文规范，而非只检测固定敏感词。研究将隐私泄露置于多轮智能体对话中系统测量。

Result: 作者扩展 CIMemories 构建 CIMEMORIES DIALOGUE，含 4,410 个基于情境完整性理论的双智能体对话实例。评测 Qwen3 和 Llama 3.1 系列七种开源模型，在 1 轮和 6 轮对话及三种提示条件下测试。结果显示任务能力与隐私泄露紧密耦合：所测模型中没有一个能将两者解耦，能力更强的模型系统性地泄露更多。具思考模式的模型比非思考版本泄露更多，作者将其归因于 CoT 扩展的推理暴露面；对话从 1 轮延长至 6 轮时泄露累积，最高增长至 3 倍。

Insight: 泄露风险受交互上下文和对话长度影响，单轮/静态提示评测可能低估长程多智能体场景中的累积问题。作者报告的是七种模型、指定轮次和提示条件下的结构性关联，不能推断所有思考模型必然更不安全。

Abstract: As LLM agents are increasingly deployed in high-stakes enterprise settings, understanding how they handle sensitive information in agentic dialogue has become critical. Privacy leakage in such settings requires reasoning about whether information flows respect the contextual norms under which they originated. We extend CIMemories to CIMEMORIES DIALOGUE, a benchmark of 4,410 two-agent dialogue instances grounded in Contextual Integrity theory. Using this benchmark, we systematically evaluate 7 open-weight models from the Qwen3 and Llama 3.1 families across conversation lengths of 1 and 6 turns under three prompting conditions. We find that task capability and privacy leakage are fundamentally coupled: more capable models are systematically leakier, with no model in our evaluation able to decouple the two. Thinking models consistently leak more than their non-thinking counterparts, a pattern we attribute to the extended reasoning surface created by chain-of-thought. Leakage compounds with conversation length, rising up to 3x from 1 to 6 turns. Our results suggest that the coupling between task capability and privacy leakage is not an artifact of misconfiguration but a structural feature of current agentic systems.

[255] Is the Future Predictable? Modeling Future Predictability for Reliable Temporal Reasoning cs.CL | cs.AIPDF

Yongyi Huang, Huaping Zhang, Wenyao Cui, Qiuchi Li, Lei Li

TL;DR: ChronoRegime 要求模型先判断历史证据是否足以承诺未来答案；CTR 在多种 LLM 上较普通问答平均提升严格准确率 6.7 个百分点，并减少幻觉 18.4 个百分点。

Details

Motivation: LLM 对未来问题仍不可靠，常忽略未来状态的易变性，在证据不足时仍给出确定答案。对未来的推理应区分可由历史支持的状态与本质上不可预测的事实。

Result: 作者提出 ChronoRegime 任务，要求模型判断历史时间证据是否足以支撑未来答案；并提出 Confidence-Aware Temporal Reasoning（CTR），根据时间动态估计未来可预测性并调节是否作答。CTR 不做无条件预测，只在重建证据支持目标未来状态时承诺答案。多个 LLM 实验显示，相对普通 QA 基线，严格总体准确率平均提高 6.7 个百分点，幻觉率平均降低 18.4 个百分点。

Insight: 先估计未来目标的可预测性，再决定是否作答，可将弃答/谨慎承诺纳入时间推理，而不是对所有未来问题强行预测。所报平均提升跨多个模型，摘要未说明数据集、预测时间跨度或各模型结果。

Abstract: Large language models (LLMs) remain unreliable when reasoning about the future. Existing models often overlook the difference among future states, committing to specific answers even when the queried facts are inherently volatile. To capture this challenge, we introduce ChronoRegime, a future-oriented temporal reasoning task that requires models to determine whether historical temporal evidence justifies commitment to a future answer. Furthermore, we propose Confidence-Aware Temporal Reasoning (CTR), a framework that estimates future predictability from temporal dynamics and regulates answer commitment accordingly. Rather than performing unconditional forecasting, CTR commits only when the reconstructed evidence supports the target future state. Across multiple LLMs, CTR improves strict overall accuracy by 6.7 percentage points and reduces hallucination rates by 18.4 percentage points on average over the plain QA baseline. These results demonstrate that explicit predictability assessment is an effective mechanism for safer and more reliable temporal reasoning under future uncertainty.

[256] Fin_Agent at FinEval 2026: Task-Aware Methods for Multilingual Financial NLP cs.CLPDF

Mohotarema Rashid, Tanzina Hossain, Xiaoying Song, Lingzi Hong

TL;DR: Fin_Agent 在 FinEval 2026 多语言金融 NLP 任务中，希腊语 NER 公榜/私榜 F1 为 0.9521/0.8884，日本语隐含沟通识别准确率 98.0%，金融 QA ROUGE-1 F1 为 0.3023/0.3419。

Details

Motivation: 多语言金融语言理解包含实体识别、含蓄表达识别和跨语言问答等不同任务，系统需要按任务选择不同方法，同时应关注数据集特定规则、手工线索及监督不足等限制。

Result: Fin_Agent 是 FinEval 2026 共享任务方案。希腊语金融实体识别结合注释规则和零样本 Qwen2.5-32B 分类，公榜 F1 为 0.9521，私榜为 0.8884；日语隐含沟通识别采用可解释的线索式序数评分器，测试准确率 98.0%；多语言金融问答用 BGE-M3 检索加 Qwen3-32B 生成，公榜/私榜 ROUGE-1 F1 分别为 0.3023/0.3419。作者指出主要限制包括数据集特定规则、手工构造线索和监督有限。

Insight: 三个子任务使用不同的任务定制方案，因此分数应按公榜/私榜和任务分别解读，不能横向比较。摘要明确承认规则与线索的任务依赖性以及训练监督有限。

Abstract: We present Fin_Agent, our submission to the FinEval 2026 shared task on multilingual fi nancial language understanding. Our sys tems address Greek financial named-entity recognition, Japanese implicit communication recognition, and multilingual financial ques tion answering. For Greek NER, we com bine annotation-aware numeric rules with zero shot Qwen2.5-32B classification, achieving 0.9521 public F1 and 0.8884 private F1. For Japanese implicit communication, an inter pretable cue-based ordinal scorer achieves 98.0% test accuracy. For multilingual financial QA, BGE-M3retrieval with Qwen3-32B gen eration achieves 0.3023 and 0.3419 ROUGE-1 F1 on the public and private splits. Key limita tions include dataset-specific rules, handcrafted cues, and limited supervision. Implementa tion code is available https://github.com/ Rashidnira/finagent-fineval-2026.

[257] Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation cs.CL | cs.AI | cs.LGPDF

Hermann Kumbong, Ishan S. Khare, Iddah Mlauzi, Andrew Woen, Charles Ding

TL;DR: Next-Concept Modeling 用多尺度残差 VQ-VAE 将文本编码成层级离散概念码，再逐层并行预测；在 LM1B 上比最强扩散语言基线快 21×，在 Sudoku 各难度上超过 AR 和扩散基线。

Details

Motivation: 扩散语言模型能并行采样，但步数减少时质量下降很快，每步还处理整条序列，且需大量数据训练轮次才能达到较好质量。需要同时实现粗到细生成、并行性和较高速度。

Result: NCM 分两阶段：多尺度残差 VQ-VAE 将文本编码为层级离散码，Transformer 逐级生成，每一级中的所有码并行预测。在参数规模匹配的 LM1B 实验中，生成困惑度优于最强扩散语言基线且快 21×；OpenWebText 上质量与最强扩散基线有竞争力、速度快 10×；XSum 上 ROUGE-L 持平、速度快 31×。Sudoku 推理任务中，NCM 在所有难度超过自回归和扩散基线，同时使用更少函数评估次数。摘要对 Sudoku 描述重复一次，未提供具体得分。

Insight: 层级离散表示把序列生成分解为由粗到细的并行概念预测，可减少逐 token 解码和全序列反复采样的开销。速度数字对应各自文本基准，不能直接相互比较；Sudoku 结果无具体数值。

Abstract: Diffusion language models (DLMs) offer parallel sampling for faster generation than autoregressive models, an increasingly important advantage as inference becomes a dominant computational workload. However, DLMs degrade quickly with fewer sampling steps, process the full sequence at every step, and require many passes over the dataset to reach good quality. We introduce Next-Concept Modeling (NCM), a two-stage, coarse-to-fine framework: a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, and a transformer generates one level at a time, predicting all codes within a level in parallel. At matched parameter scales, NCM achieves better generative perplexity than the strongest DLM baseline on LM1B while running $21\times$ faster. On OpenWebText, NCM is competitive with the strongest DLM while running $10\times$ faster; on XSum, it matches the strongest DLM’s ROUGE-L while being $31\times$ faster. On the reasoning task of Sudoku, NCM outperforms autoregressive and diffusion baselines across all difficulty levels, using fewer function evaluations. On the reasoning task of Sudoku, NCM outperforms autoregressive and diffusion baselines across all difficulty levels, using fewer function evaluations.

[258] Agentic Data Engineering for LLMs: System Design and a Controlled Empirical Study cs.CL | cs.AI | cs.LGPDF

Ruomeng Ding, Qianli Shen, Zhaoyang Han, Meixin Chen, Daoyuan Chen

TL;DR: ADE 研究智能体如何根据实验反馈迭代设计数据选择、奖励和课程策略；四项任务中所选策略在 AIME25 Pass@32 最多提升 13.34%，四基准平均最多提升 5.85%，但循环内分数上升不保证泛化。

Details

Motivation: LLM 数据工程需反复实验来确定数据筛选、奖励设计与课程学习如何影响模型，适合作为 AI-for-AI 研究。关键问题是智能体能否从实验反馈逐步改进策略，以及循环内选择的方案能否在留出数据上真正提升目标模型。

Result: Agentic Data Engineering（ADE）框架允许智能体基于累积发现继续、修改或组合研究方向；策略按循环内分数筛选，再在固定训练和评测协议下用留出基准检验。四类任务覆盖数学/代码数据选择、数学奖励设计和数学课程学习。选中策略相对通用基线，在 AIME25 Pass@32 最多绝对提升 13.34%，四个留出基准平均最多提升 5.85%。循环内分数会随实验改进，但留出轨迹显示后期循环分数增长并不总带来泛化提升。比较研究组织与反馈访问方式后，结合证据分析和累积发现的方案更有利。

Insight: 该工作不仅展示智能体可自动改进数据工程策略，也明确报告循环内优化指标与留出泛化可能脱钩。13.34% 和 5.85% 是所测任务中的最大绝对提升，不代表所有任务都获得相同收益。

Abstract: LLM data engineering requires iterative experimentation to determine how data selection, rewards, and curricula affect model learning, making it a natural setting for AI-for-AI (AI4AI) research. Can agents use experimental feedback to progressively improve strategies, and do the resulting strategies improve target-model performance on held-out data? We study how agents use experimental feedback to develop data-engineering strategies under fixed training and evaluation protocols. Our Agentic Data Engineering (ADE) framework lets agents pursue, revise, or combine research directions using cumulative findings. Strategies are selected by in-loop scores and evaluated on held-out benchmarks. Across four tasks-math and code data selection, math reward design, and math curriculum learning-selected strategies achieve absolute gains over generic baselines of up to 13.34% on AIME25 Pass@32 and 5.85% averaged across four held-out benchmarks under fixed training conditions. In-loop scores improve over successive experiments, while held-out trajectories show that later in-loop gains do not always improve generalization. Comparisons of research organization and feedback access favor combining evidence analysis with cumulative findings. Research trajectories show how earlier findings inform strategy revisions and combinations across research directions.

[259] The Thinking Leakage in Reinforcement Learning for LLMs cs.CLPDF

Wenxuan Zhu, Shuming Liu, Bang An, Aznaur Aliev, Bing Li

TL;DR: 对 Qwen3-1.7B/4B/8B 在 NO-THINK 模式做 GRPO 后的增益，很大部分来自重新激活原模型 THINK 能力，而非新学到更强推理；机制证据来自长度、激活子空间和注意力头分析。

Details

Motivation: Qwen3 等混合推理模型通过少数模板 token 切换 THINK/NO-THINK。常见配方在 NO-THINK 模式做强化学习并报告推理提升，但这究竟是新能力学习，还是原 THINK 能力重新出现，需要区分。

Result: 作者用 GRPO 在 NO-THINK 模式训练 Qwen3-1.7B、4B、8B。随着准确率提升，回答更长、更有反思性，原模型也越来越倾向判为类似自身 THINK 输出，内部激活向 THINK 模式移动。继续放大 RL 更新可诱发 THINK 行为并接近但未超过 THINK 模式准确率。Qwen3-8B 中，将 RL 模型低维 THINK 相关激活子空间替换为原模型对应值会移除大部分 RL 增益；反向替换则将相当部分增益转移到原模型。RL 引起的变化经由读取空 think block 的注意力头传递。作者还发现优势加权更新偏好更像 THINK 的回答，基于该信号的一步梯度更新会使模型向 THINK 状态移动。

Insight: NO-THINK 训练收益未必代表模型获得新推理机制，可能是模板切换之外重新调用已有 THINK 能力。子空间交换和梯度分析提供机制证据，但结论针对 Qwen3 系列及特定 GRPO/模板设置。

Abstract: Hybrid reasoning models such as the Qwen3 series switch between a THINK mode and a NO-THINK mode through a few chat-template tokens. A popular recipe trains such models with reinforcement learning in NO-THINK mode and reports improved reasoning performance. We provide evidence that much of this gain can be explained by thinking leakage: the gain comes largely from partially reactivating the reasoning abilities of the original model’s THINK mode instead of learning enhanced reasoning behavior. We train Qwen3-1.7B, 4B, and 8B with GRPO in NO-THINK mode. As accuracy improves, responses become longer and more reflective, are increasingly judged by the original model to resemble its own thinking, and the model’s internal activations shift toward its THINK mode. We then investigate at the level of weights, activations, and attention heads: scaling the RL update further elicits THINK-mode behavior and approaches THINK-mode accuracy but does not exceed it; in Qwen3-8B, replacing a low-dimensional THINK-related activation subspace of the RL-trained model with its original-model value removes most of the RL gain, and the reverse exchange transfers a considerable fraction of the gain to the original model; and the RL-induced shift passes through the attention heads that read the empty think block. We further identify a possible mechanism in the GRPO training signal: responses favored by the advantage-weighted update are more THINK-like, and a gradient step computed from this signal moves the model toward its THINK state. Thus, gains in NO-THINK mode can largely reflect reactivation of reasoning abilities already present in the original model.

[260] ASSET: Adaptive Speech Segmentation via Reinforcement Learning cs.CLPDF

Cihan Xiao, Yuan Tseng, Matthew Wiesner, Desh Raj, Leibny Paola Garcia Perera

TL;DR: ASSET 用 GRPO 学习按任务压缩语音编码器序列的自适应分段和池化；英语 LibriSpeech 上以约少 25% 音频 token 达到接近转录辅助字符边界 oracle 的表现，中文 AISHELL 上优于相近 token 率的音素边界 oracle。

Details

Motivation: 语音大模型通常将预训练声学编码器接到 LLM 解码器，但声学输出序列远长于文本序列，LLM 随输入长度扩展的计算成本高。需要压缩编码特征，同时保留面向任务的信息。

Result: ASSET 先将声学编码器输出划分为时长可变的片段，再把每段池化成连续音频 token。作者用 GRPO 训练分段器，奖励来自任务性能并加入 token 率惩罚。自动语音识别实验中，英语 LibriSpeech 上 ASSET 使用约少 25% 音频 token，表现与使用转录辅助字符边界的 oracle 相当；中文 AISHELL 上，在相近 token 率下优于音素边界 oracle。非 ASR 探索实验显示，框架适配其他任务时会学到不同分段模式。

Insight: 学习式切分依据最终任务决定何时压缩，而不是依赖固定字/音素边界；token 预算可直接进入优化目标。摘要未提供具体 WER/CER 数值，oracle 比较也应理解为不同语言数据上的各自结果。

Abstract: Speech large language models (LLMs) are often built using pretrained acoustic encoders as an audio interface to an LLM decoder. However, LLMs scale poorly with input sequence length, while acoustic encoders tend to produce substantially longer sequences than typical textual inputs. This motivates learning to compress encoder outputs while preserving task-oriented information. To this end, we introduce ASSET, a reinforcement learning (RL)-based framework that learns to adaptively compress encoder outputs by first partitioning them into variable-duration segments and then pooling each segment into a continuous audio token. We use group-relative policy optimization (GRPO) to train a segmenter with task performance feedback and a token-rate penalty. In automatic speech recognition (ASR) experiments, ASSET achieves performance comparable to a transcript-assisted character-boundary oracle with approximately 25% fewer audio tokens on English LibriSpeech, while improving on a phone-boundary oracle at a similar token rate on Mandarin AISHELL. Exploratory experiments on non-ASR tasks further illustrate how learned segmentation changes when the framework is adapted to tasks beyond transcription.

[261] Multimodal Co-Generation With Continuous Bitstream Diffusion cs.CLPDF

Georgios Batzolis, Georgios Kontos, Bruno Trentini, Pranam Chatterjee, Micha Livne

TL;DR: CoBit 用共享二进制字节空间和统一 bitwise 去噪目标实现图文、语音文本及蛋白序列结构的联合/条件生成；语音生成文本与语音转录 WER 为 3.5%，蛋白生成结构自一致 TM-score 为 0.911。

Details

Motivation: 作者考察连续比特流扩散能否用同一训练与采样协议处理多模态联合生成及条件生成，而无需为不同领域设计专用目标或采样流程。

Result: 每个任务单独从头训练模型，但在各自模态内共享单一输入投影、Transformer 主干、逐 bit 预测头和二进制 score-matching 目标；一个检查点通过指定比特流哪些部分可观测支持联合生成及训练过的条件方向。MNIST-Sum 图像—文本控制任务中，无学习 tokenizer 仍实现近乎完美条件准确率和高联合一致性。未做文本预训练的单个语音—文本模型支持语音合成、识别、续写和无条件共同生成；联合生成语音与文本之间的 WER（生成文本对生成语音自动转录）为 3.5%。蛋白模型单检查点支持正向折叠、逆折叠及序列—结构联合生成；100–500 残基联合生成蛋白经计算重折叠后，与生成主干结构的平均自一致 TM-score 为 0.911。

Insight: 共享二进制表征允许一个模型方向条件化地完成跨模态生成，避免额外学习 tokenizer；但语音/蛋白任务仍各自训练模型，所谓统一指框架和目标，不是一个跨领域通用检查点。结果分别用 WER 和结构 TM-score 验证。

Abstract: We show that continuous bitstream diffusion (CoBit), originally developed for language modeling, provides a unified framework for joint and conditional multimodal generation using the same training and sampling protocol. Across image-text, speech-text and protein sequence-structure modeling, models are trained separately from scratch, but each uses a single input projection, transformer backbone, bitwise prediction head and binary score-matching objective shared across its modalities. Within each setting, a single checkpoint supports joint generation and the conditional directions included in training by changing which parts of the bitstream are observed. We first validate the approach on MNIST-Sum, a controlled image-text task pairing raw binary images with equations encoded as text, achieving near-perfect conditional accuracy and high joint consistency without learned tokenizers. We then apply the same protocol at larger scale to speech-text and protein sequence-structure modeling. Trained without text pretraining, a single speech-text model performs speech synthesis, recognition, continuation, and unconditional co-generation. Jointly generated speech and text achieve 3.5% word error rate between the generated text and automatic transcriptions of the generated speech. In protein modeling, a single CoBit checkpoint supports forward folding, inverse folding, and joint sequence-structure generation. For jointly generated proteins spanning 100-500 residues, computational refolding of CoBit’s sequences yields structures that agree with the generated backbones at a mean self-consistency TM-score of 0.911. Together, these results show that CoBit provides a common generative framework for substantially different modalities and forms of cross-modal dependence, without introducing a domain-specific objective or sampling procedure.

[262] SciCurateBench: Curating Experimental Data from Heterogeneous Scientific Literature with Verbal-Action Agents cs.CLPDF

Minuk Ma, Kaiyun Guo, Jiarui Ding

TL;DR: SciCurateBench 用 200 篇专家标注生命科学论文检验智能体能否从原始文献核实合取实验条件；LangPoli 在论文分类上 F1 为 83.16%，高于基础 LLM 的 58.22%。

Details

Motivation: 科学文献数据整理要求跨长篇、异构原文核验精确实验条件，现有基准多只测短查询的宽泛相关性，无法反映这一耗时复杂任务。

Result: SciCurateBench 收录主要生命科学期刊中的 200 篇专家标注论文，要求系统从未处理科学文献中验证多个合取生物实验条件。作者发现普通 LLM、常见推理/工具/反馈智能体在该任务上都很困难。为研究可行策略，提出 LangPoli：策略智能体生成自然语言行动指南，执行智能体将其转成可执行工具。LangPoli 在论文分类任务上 F1 为 83.16%，基础 LLM 为 58.22%；在结构化信息检索表现较强，并改善 LitSearch、BEIR、TREC 公共基准重排。作者还探索课程迁移，从已解决任务提炼可泛化策略以加速相关任务求解。

Insight: 科学整理的核心是核实多项精确条件，而非只做主题相关性排序；自然语言行动规范与可执行工具的分离可帮助策略复用。分类 F1 是摘要给出的具体分数，结构检索和公共重排改进未列指标。

Abstract: Scientific research increasingly depends on curating experimental data from literature. Scientific data curation requires verifying exact experimental conditions (e.g., specific assays, clinical outcomes, and data availability) across long, heterogeneous files, which is a highly time-consuming process. Existing benchmarks evaluate only broad relevance on short queries, failing to capture this complexity. We introduce SciCurateBench, a benchmark of 200 expert-annotated papers from major life-science journals that challenges systems to verify conjunctive biological experimental criteria from raw, unprocessed scientific literature. We show that plain LLMs and popular reasoning, tool-using, and feedback-based agents all struggle with this task. To study what makes the task tractable, we further propose LangPoli, an integrative system in which a policy agent generates natural language action guidelines that actor agents convert into executable tools. LangPoli showed strong performance on both Paper Classification (83.16% F1 vs. 58.22% for its base LLM) and Structured Information Retrieval, and also improves re-ranking on the public benchmarks of LitSearch, BEIR, and TREC. We further explore curriculum transfer, a method that distills generalizable strategies from solved tasks to accelerate the solving of related problems. We hope this work facilitates research in truly practical scientific AI assistance. The anonymized pre-release can be found from https://anonymous.4open. science/r/scicuratebench-D885/.

[263] Scaling Laws for Multimodal Data Mixtures cs.CL | cs.LGPDF

Aditi Khandelwal, Ayush K. Tarun, Yixuan Xu, Imanol Schlag, Golnoosh Farnadi

TL;DR: 基于 268 组 MoE 实验（457M–8.3B 参数、最多 150B token）的多模态缩放研究发现：音频更偏模型规模驱动，文本/视觉需平衡参数与数据；视觉对音频有强干扰，反向干扰很弱。

Details

Motivation: 前沿模型越来越多地联合预训练文本、视觉和音频，但各模态数据比例如何影响训练效率、以及跨模态学习互相干扰多少，公开的系统研究仍有限。

Result: 作者推导文本、视觉、音频三模态的数据混合缩放规律，并用 MoE 进行 268 组实验，规模从 457M 到 8.3B 参数，训练数据最多 150B token。计算最优配置呈模态差异：音频更偏向增加模型参数而非 token；文本和视觉则需在模型规模与数据量之间平衡。视觉或音频数据重复超过 2 倍后收益很小。跨模态干扰高度不对称：视觉显著干扰音频学习，而音频对文本和视觉学习的干扰可忽略。

Insight: 数据配比不能仅用统一 token 比例处理，不同模态的计算最优和相互干扰不同；MoE 被用来高效规模化并部分限制干扰。摘要未提供缩放律公式参数、具体配比或所有实验设置。

Abstract: Frontier AI systems are increasingly natively multimodal, jointly pretrained on multiple modalities, such as text, vision, and audio. However, the scientific understanding of the multimodal data mixtures used to pretrain these systems — how much of each modality to mix, and at what potential interference cost to the others — is largely absent or gatekept. In this work, we derive data mixing scaling laws for three modalities: text, vision, and audio. Multimodal Large Language Models (MLLMs) increasingly adopt Mixture-of-Experts (MoE) architectures, as MoEs scale efficiently and partially limit cross-modal interference. We therefore use MoEs to conduct 268 experiments ranging from 457M to 8.3B parameters, on up to 150 billion tokens. Our work leads to multiple actionable insights into multimodal data mixing for training MLLMs. First, compute optima are modality-specific: audio modality is model-heavy, favoring scaling parameters over tokens, while text and vision require balanced scaling of model size and data. Second, repeating vision or audio data beyond 2x yields negligible benefits. Finally, cross-modal interactions are highly asymmetric: vision strongly interferes with audio, whereas audio’s interference with learning text and vision is negligible. Overall, we provide a principled foundation for understanding the multimodal data mixtures needed to train frontier MLLMs.

[264] LongHisDoc: A Comprehensive Benchmark for Chinese Long Historical Document Understanding cs.CLPDF

Zeyu Shan, Yongxin Shi, Shihang Wu, Peirong Zhang, Yuanyuan Fang

TL;DR: LongHisDoc 发布 101 份中文历史 PDF 的 1,012 个专家标注问题，覆盖单页、跨页、多跳和不可回答类型；顶级 LVLM/LLM 准确率分别仅 62.9%/69.9%。

Details

Motivation: 中文历史文献具有重要文化价值，但篇幅长、版式和视觉内容复杂，非专业读者难以理解。现有大语言和视觉语言模型在现代文档任务上的表现尚不能代表整本历史文档理解能力。

Result: LongHisDoc 是面向长篇（整本）历史文档理解的基准，含 101 份 PDF、1,012 对专家标注问答。文档平均 85.1 页、42,776.4 token，问题分为单页、跨页、多跳和不可回答四类，需要整合文本、版面与视觉元素。评测中最佳 LVLM 准确率 62.9%，最佳 LLM 为 69.9%，仍存在显著困难。作者还评估多类模态检索器和端到端 RAG，为历史文献问答系统提供基线参考。

Insight: 问题设计要求整合整本文档及多模态证据，并纳入不可回答样例，能测出短文档 benchmark 难以体现的能力边界。摘要未给检索器和 RAG 的分项结果；LLM 高于 LVLM 的结果仅针对该基准和评测配置。

Abstract: Chinese historical documents are invaluable cultural treasures, but their complexity poses significant comprehension challenges for non-specialists. Despite the success of Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) in modern document tasks, their capabilities in Long (full-book) Historical Document understanding remain largely unexplored. To bridge this gap, we introduce LongHisDoc, the first benchmark tailored for long historical document understanding. LongHisDoc comprises 1,012 expert-annotated QA pairs derived from 101 PDF documents, with an average of 85.1 pages and 42,776.4 tokens per document. It spans 10 categories and features four QA types: single-page, cross-page, multi-hop, and unanswerable. Answering these questions requires integrating multimodal evidence from text, layout, and visual elements. Our evaluation reveals that the top-performing LVLM and LLM achieve only 62.9% and 69.9% accuracy, respectively, highlighting a substantial performance gap. Furthermore, we assess various modal retrievers and the end-to-end RAG performance to provide references for Retrieval-Augmented Generation (RAG) systems. We believe that LongHisDoc serves as a comprehensive benchmark to catalyze future research in full-book historical document understanding. The benchmark and evaluation code are available at https://huggingface.co/datasets/qweq12433454/LongHisDoc and https://github.com/SCUT-DLVCLab/LongHisDoc.

[265] PVFH-Bench: Benchmarking Premise-Induced Visual Fixation Hallucinations in Vision-Language Models cs.CL | cs.CVPDF

Xinhao Wang, Junbo Cui, Zhenghao Liu, Yuan Yao, Xiaohua Li

TL;DR: PVFH-Bench 研究错误文本前提如何使 LVLM 锚定于易混视觉线索并幻觉不存在物体；10 个先进模型均表现脆弱，显式 CoT 可能加剧过度解读弱视觉证据。

Details

Motivation: 视觉对象幻觉损害 VQA、具身智能体及人机决策可靠性。作者受认知偏差研究启发，考察模型是否会接受错误文本前提，再把易混视觉证据解释为不存在目标的支持。

Result: Premise-Induced Visual Fixation Hallucination（PVFH）指模型被前提诱导，锚定于视觉上可混淆的证据，接受错误文本前提，继而回答有关不存在但看似合理物体的问题。PVFH-Bench 自动构造视觉上易混淆但实际缺失的目标，并设置核心幻觉任务和诊断子集，追踪模型从物体识别到前提诱导幻觉的过程。对十种先进 LVLM 的实验发现所有模型均易受此类现象影响；进一步结果显示，显式思维链可能促使模型过度解读弱视觉线索，从而加重幻觉。摘要未报告各模型或子集具体分数。

Insight: 错误前提与视觉相似干扰物共同作用，可能让模型把弱线索当作证据；因此增加显式推理步骤未必纠错，反而可能强化锚定。该结论针对所构造的 PVFH 场景，摘要未提供幻觉率或防护方法。

Abstract: Large Vision-Language Models (LVLMs) have achieved remarkable progress across a wide range of multimodal tasks. However, object hallucination remains a persistent challenge, undermining model reliability in applications such as visual question answering, embodied agents, and human-AI decision support systems. In this paper, inspired by cognitive science studies on perceptual bias, we investigate a new failure mode termed Premise-Induced Visual Fixation Hallucination (PVFH). Specifically, PVFH refers to a premise-induced reasoning error in which models become anchored to visually confusable evidence, accept a false textual premise, and subsequently generate responses about an absent yet visually plausible object. To systematically study this phenomenon, we introduce PVFH-Bench, a benchmark that automatically constructs visually confusable but absent targets and pairs the core hallucination setting with diagnostic subsets to trace how models move from object recognition to Premise-Induced Visual Fixation Hallucination. Experimental results on 10 state-of-the-art LVLMs demonstrate that all evaluated models are highly vulnerable to Premise-Induced Visual Fixation Hallucination. Moreover, we find that explicit chain-of-thought reasoning can further exacerbate this failure mode by encouraging models to overinterpret weak visual cues as supporting evidence for hallucinated objects.

[266] Vela: User Behavior as Supervision Sparks LLM Intelligence cs.CL | cs.AIPDF

Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Weiqiang Hang

TL;DR: Vela 将 10 万名用户的匿名行为记录转换成 91 类任务和训练轨迹，筛得约 7.5 万任务；微调 Qwen3.6-35B-A3B 后用户相关任务最高提升 11.5 分，通用基准平均提升 4 分。

Details

Motivation: 个人助理、推荐和广告等应用需从数周到数年的行为历史推断偏好、意图及未来行动，但静态文本或合成训练数据难以反映用户行为随时间变化。真实平台行为日志可提供规模化监督。

Result: Vela 把真实用户行为转成可训练任务和轨迹，数据来自 100,000 名匿名用户，覆盖 91 种真实任务模板；按难度和轨迹接受标准过滤后保留约 75,000 个任务及对应训练轨迹。用其微调 Qwen3.6-35B-A3B，用户相关任务相对基座最高提高 11.5 分，通用基准平均提高 4 分。增加行为数据规模可继续提升用户任务表现；微调后再做 RL 相比直接从基座做 RL，最高多 7.7 分。

Insight: 真实行为记录不仅是训练样例，也可通过任务模板和执行轨迹转化为监督信号，提升以用户为中心的能力并外溢到通用任务。数据规模与匿名化方式摘要未详细说明；“最高提升”并不代表所有模板或用户群体一致改善。

Abstract: Large language models are increasingly deployed in applications that act on users’ long-term behavior histories, from personal assistants to recommendation and advertising systems. These applications demand the user dimension of LLM intelligence: inferring a person’s preferences, intent, and likely future actions from weeks to years of behavior data. Yet current models are trained on static text or synthetic data, neither of which captures how user behavior evolves over time. Real-world behavior records, already available at scale on content platforms and in personal-assistant agents, offer a natural supervisory signal. We propose user behavior as supervision and introduce Vela, a framework that converts real user behavior data into trainable tasks and trajectories. From the anonymized behavior records of 100,000 users, Vela generates 100,000 tasks across 91 real-world task templates; difficulty-based filtering and trajectory acceptance criteria retain approximately 75,000 tasks with selected training trajectories. Fine-tuning Qwen3.6-35B-A3B on these trajectories improves performance over the base model by up to 11.5 points on user-related tasks and by an average of 4 points on general benchmarks. Training on more user behavior data further improves user-related performance, and reinforcement learning after fine-tuning outperforms reinforcement learning from the base model by up to 7.7 points. These gains suggest that real-world user behavior is a scalable source of supervision for both user-centered capabilities and the broader intelligence of LLMs.

[267] Lingdan-V2: Large language models for clinically grounded reasoning in traditional Chinese medicine cs.CL | cs.AIPDF

Rui Hua, Yu Wei, Ziri Jin, Zhuo Sun, Kai Chang

TL;DR: Lingdan-V2 提出 4B/8B/14B 中医语言模型，结合领域继续预训练、推理监督和处方级 RL；14B 模型公开基准处方生成 F1 为 37.0，高于总参数 671B 的 DeepSeek-R1 的 32.3。

Details

Motivation: 中医处方决策需要整合中医知识、辨证、治则、药味组合和剂量，通用医学推理模型难以提供临床落地的中医推理与处方建议。

Result: Lingdan-V2 模型族包括 4B、8B、14B，训练采用分阶段临床对齐：基于中医与临床记录大规模继续预训练，加入互补推理监督，并进行以处方为中心的强化学习。作者将中药处方生成建模为带处方级反馈的结构化 RL，直接优化临床决策输出。评测包括 13 项中医知识/临床推理任务的公开基准 25,620 个专家核验样例，以及 4,348 例脾胃病处方回顾性私有数据。14B RL 对齐模型在公开基准处方生成 F1 为 37.0，DeepSeek-R1（总参数 671B）为 32.3；临床数据药物级 F1 为 52.8，剂量一致性也改善。强化学习后训练在各模型规模上持续提升处方推荐；医师来源分析显示专家差异会影响奖励稳定性。

Insight: 处方级监督和反馈让优化目标直接对应临床决策，而不仅是知识问答；专家之间处方差异会造成奖励噪声。结果来自给定公开基准及脾胃病回顾性数据，不能解读为前瞻临床效果或替代医师决策。

Abstract: Large language models have advanced medical reasoning, but clinically grounded reasoning in traditional Chinese medicine (TCM) remains challenging because prescription decisions require integrating domain knowledge, syndrome differentiation, treatment principles, herbal prescription composition, and dosage determination. We develop a TCM language model family comprising models with 4 billion, 8 billion, and 14 billion parameters. The artificial intelligence contribution is a staged clinical alignment framework integrating large scale continued pretraining on TCM and clinical records, complementary reasoning supervision, and prescription-centered reinforcement learning (RL). We further formulate herbal prescription generation as a structured RL problem with prescription-level feedback, enabling direct optimization of clinical decision outputs beyond knowledge reproduction and generic instruction following. The engineering application is TCM clinical decision support, particularly herbal prescription recommendation. We evaluate the models on a public benchmark containing 25,620 expert-verified items across 13 TCM knowledge and clinical reasoning tasks and on a private retrospective dataset of 4,348 spleen and stomach disease prescription cases. The 14 billion parameter reinforcement aligned model achieves a prescription generation F1 of 37.0 on the public benchmark, compared with 32.3 for DeepSeek-R1 with 671 billion total parameters, and a herb-level F1 of 52.8 on the clinical dataset with improved dosage agreement. Reinforcement post-training consistently improves prescription recommendation across all scales, while physician-source analysis shows that expert heterogeneity affects reward stability. These findings demonstrate a scalable pathway from domain grounding and reasoning supervision to clinically oriented prescription decision support.

[268] Dependency Reasoning for Aspect Sentiment Quadruple Prediction cs.CL | cs.LGPDF

Shima Esfandyar, Seywan Soleymanbaigi, Fardin Akhlaghian Tab, Fatemeh Daneshfar, Mourad Oussalah

TL;DR: DRIFT 将依存树线性化并作为显式关系上下文放入提示，联合检索增强上下文学习与语法感知判分；Rest15/Rest16 达到 54.47%/64.10% F1。

Details

Motivation: 方面情感四元组预测需要联合抽取方面、观点、类别和情感，是结构化关系推理而非简单序列共现。既有 LLM 方法多通过检索示例或分阶段训练间接使用句法，缺少单次生成可直接利用的统一关系上下文。

Result: DRIFT 将依存句法树线性化并直接注入 LLM 提示，使模型依据方面—观点依存关系推理，而非仅依赖 token 距离。方法还结合检索增强上下文学习和语法感知 LLM 判定机制，以补充跨度边界变化下的精确匹配评估。在 SemEval Rest15、Rest16 和 ACOS 上，DRIFT 在比较的生成式 ASQP 方法中整体最强，Rest15/Rest16 F1 分别为 54.47% 和 64.10%；短程与中程方面—观点依赖配置均有稳定收益。判定分析还显示，严格精确匹配会系统性低估语义正确预测。

Insight: 显式依存关系可让模型在一次生成时看到方面与观点的结构联系；辅助判定则揭示字符串边界匹配与语义正确性存在差别。摘要未提供 ACOS 具体分数或各距离区间单独结果。

Abstract: Aspect Sentiment Quadruple Prediction (ASQP) is fundamentally a relational reasoning problem: identifying aspect-opinion-category-sentiment quadruples requires modeling structured dependencies, not merely sequential co-occurrence. Existing LLM-based approaches incorporate syntactic information indirectly, via retrieval-based demonstration selection or multi-stage training, rather than as unified, single-pass generation-time context. We present DRIFT, which linearizes dependency parse trees and injects them directly into LLM prompts as explicit relational context, enabling the model to reason over aspect-opinion dependencies rather than infer them from token proximity alone. DRIFT further integrates retrieval-augmented in-context learning and a syntax-aware LLM-based adjudication mechanism that complements exact-match evaluation under span-boundary variation. On SemEval Rest15, Rest16, and ACOS, DRIFT achieves the strongest performance among compared generative ASQP methods, reaching 54.47% and 64.10% F1 on Rest15 and Rest16, with consistent gains across short- and medium-distance aspect-opinion configurations; adjudication further shows exact-match metrics systematically underestimate semantically correct predictions.

[269] Agent-Guided Evidence Acquisition for Emotion-Cause Pair Extraction cs.CLPDF

Zhuangzhuang Pan, Yan Xia, Chee Seng Chan

TL;DR: 该 ECPE 方法让学生仅对验证 F1 决策边界附近的情感—原因关系查询专家判断，并只修订已查询预测；名义 10% 查询率下 pair F1 提高 2.64–3.88 分。

Details

Motivation: 情感—原因对抽取需要关联对话中的情绪及其原因，而部分关系在初始证据下难以判断。智能体可选择何时获取专家信息，但额外信息不一定总能改善原因识别。

Result: 研究考察 agent-guided evidence acquisition：学生围绕验证集选出的 F1 决策边界，对边界两侧不确定关系选择性查询专家，并只修改实际查询的预测。通过控制信息访问条件，比较委派粒度、查询选择和响应融合对预测质量的影响。主实验名义查询率为 10% 时，选择性融合相较学生基线，平均 pair F1 提升 2.64–3.88 分；不同模型家族下 ECF 达 65.14，MEC4 达 49.74。分析表明，允许访问更多信息并不总能提升原因识别。

Insight: 证据获取应聚焦边界样例并考虑新证据如何改变预测，而不是无差别增加专家查询。报告结果与名义 10% 查询率和验证集选择阈值有关；摘要未提供各策略详细成本或显著性统计。

Abstract: Emotion-Cause Pair Extraction (ECPE) requires identifying the causes associated with emotions expressed in a conversation. Accurate extraction can depend on evidence that is not initially available, making it important to decide when further information is worth acquiring. We study agent-guided evidence acquisition for ECPE, where a student agent selectively accesses expert judgments for borderline relations on both sides of its validation-selected F1 decision boundary and revises only queried predictions. Through controlled comparisons under different information-access conditions, we examine how delegation granularity, query selection, and response integration affect prediction quality. At a nominal 10% query rate, selective fusion improves mean pair F1 by 2.64-3.88 points over the student in our main experiments, reaching 65.14 on ECF and 49.74 on MEC4 across model families. Our analysis suggests that greater access does not always translate into better cause identification. These findings motivate acquisition strategies that consider both the relevance of newly available evidence and its effect on prediction revision, providing an empirical basis for more effective and efficient evidence acquisition in ECPE.

[270] FidelityQ: Preserving Predictive Fidelity in Ultra-Low-Bit Language Model Quantization cs.CLPDF

Yujie Chen, Weiyan Zheng, Yiwu Yao, Tianchi Hu, Bo Wang

TL;DR: FidelityQ 用信息对齐的四级标量量化和知识密集校准，改善 2-bit 及更低比特下的预测一致性；LLaMA-2-7B 在 W2A16 下 WikiText2 困惑度为 6.62，W2A4KV4 从 8.31 降至 6.96。

Details

Motivation: 极低比特标量量化计算规则、部署效率高，但只有少数重建级别时，权重近似准确并不一定能用少量校准样本保持预测行为。

Result: FidelityQ 提出 Information-Aligned Analytic Coding（IAC）：用随机 Hadamard 混合将每个权重行塑造成中心对称源，并拟合镜像四级最小失真量化器；每个级别集合可精确分解为两个带尺度的符号基，支持无需查表的缩放整数计算。作者发现剩余预测差异呈重尾且集中在教师高置信 token，因此提出 Knowledge-Intensive Calibration（KIC）：按 token 级师生散度尾部对序列排序，将有限校准预算分配给高分样本，再在教师监督下仅更新量化级别和尺度，冻结预训练权重。LLaMA 和 Qwen 的仅权重量化及联合量化实验支持方法有效。LLaMA-2-7B 的 WikiText2 困惑度在 W2A16 为 6.62；W2A4KV4 结果从既有 SOTA 的 8.31 改进为 6.96；线性层权重内存相较 FP16 降低 8 倍。

Insight: 校准应优先覆盖少数预测差异大且教师置信高的知识密集样本；可只调量化级别/尺度而保留预训练权重。W2A16 与 W2A4KV4 分别表示不同量化配置，报告的内存节省针对线性层权重。

Abstract: Ultra-low-bit scalar quantization offers a regular arithmetic structure for efficient deployment of large language models, but preserving predictive behavior at extreme precision remains challenging: with only a few reconstruction levels, accurate weight representation must translate into consistent predictions from a small calibration set. We propose FidelityQ, a scalar quantization framework that preserves predictive fidelity at weight precisions down to 2 bits. For accurate quantization, we introduce Information-Aligned Analytic Coding (IAC), which shapes each weight row toward a centered symmetric source via randomized Hadamard mixing and fits a mirrored four-level minimum-distortion quantizer to its empirical weights. Every level set in this family decomposes exactly into two sign-valued bases with associated scales, combining source-adaptive reconstruction with lookup-free decoding compatible with scaled integer arithmetic. Our analysis further shows that residual predictive discrepancies are heavy-tailed and concentrate on tokens the teacher predicts confidently, motivating Knowledge-Intensive Calibration (KIC) for consistent quantization. KIC ranks sequences by the tail of token-wise teacher-student divergence, allocates a limited calibration budget to the highest-scoring samples, and restores predictive agreement by updating only quantization levels and scales under teacher supervision, keeping pretrained weights frozen. Experiments on the LLaMA and Qwen families demonstrate the effectiveness of FidelityQ in weight-only and joint quantization settings. On LLaMA-2-7B, FidelityQ achieves a WikiText2 perplexity of 6.62 at W2A16 and improves the prior state-of-the-art W2A4KV4 result from 8.31 to 6.96, while reducing linear-layer weight memory by 8.0x relative to FP16.

Shengwu Zheng, Yuzhe Ding, Xiaohui Liang, Fei Li, Chong Teng

TL;DR: PAT 用人格理论驱动的大模型数据增强和细粒度对比学习训练帖子级抑郁信号，再将编码器迁移到用户级检测；作者报告优于现有方法并提供多粒度解释，但摘要未给分数。

Details

Motivation: 社交媒体抑郁检测面临两类问题：基于症状词的帖子筛选容易被表面词汇误导并产生假阳性；仅靠粗粒度用户级监督也难以让编码器捕捉单条帖子中的抑郁线索。

Result: PAT 分两个阶段。第一阶段用经典人格理论驱动 LLM 数据增强，并以细粒度对比目标优化帖子级编码器，缓解数据稀缺；第二阶段用训练好的帖子级评分器筛选高质量帖子，并将编码器参数迁移至用户级任务。实验报告 PAT 显著优于现有方法。多粒度解释通过追踪帖子级心理状态轨迹、突出关键症状提供分析线索；摘要称其可为临床提供诊断证据，但未列具体基准、分数或验证方式。

Insight: 帖子级细粒度表示先用于更可靠的内容筛选，再迁移到用户级判断，试图同时减少假阳性并改善症状表征。摘要没有量化结果和临床验证细节，所述“诊断证据”应理解为论文的应用主张，而非已验证临床工具。

Abstract: Depression is a prevalent mental disorder that affects millions worldwide, with early detection crucial for effective intervention. While exist?ing methods have achieved promising results in automated depression detection, they face two key limitations: (1) noise in post screening: symptom-based semantic screening is easily misled by superficial lexical patterns, thereby introducing false-positive posts; (2) weak op?timization of the post encoder: relying solely on coarse user-level supervision signals pre?vents models from capturing depressive cues within individual posts. To address these limita?tions, we propose PAT, a Personality-driven Augmentation and Transfer learning frame?work. In the first phase, PAT optimizes a post?level encoder via a fine-grained contrastive learning objective, utilizing LLM-based data augmentation driven by classic personality the?ories to alleviate data scarcity. In the second phase, the trained post-level scorer is deployed for high-quality post filtering, while its en?coder parameters are seamlessly transferred for user-level tasks. Extensive experiments demon?strate that PAT significantly outperforms exist?ing methods. Furthermore, PAT delivers multi?granular interpretability by tracing post-level mental-state trajectories and highlighting key symptoms, thereby offering reliable diagnostic evidence for clinical practice.

[272] Select or Rewrite: Oracle-Guided Fine-Tuning for Generative ASR Post-Editing from N-Best Hypotheses cs.CL | cs.AIPDF

Shiran Liu, Gaohui Xu, Sergio Burdisso, Thibault Bañeras-Roux, Shashi Kumar

TL;DR: 该 ASR 后编辑方法用三类 N-best 数据教模型按情况选择原始 top-1 或重写：纠错样本教改写、正确 top-1 样本抑制过编辑、参考缺失样本抑制幻觉；在约 3B 模型的专有数据上收益最明显。

Details

Motivation: LLM 可基于语言上下文比较 ASR 的 N-best 候选并改写转录，通常无需重训识别器即可改善 WER，但仅提示纠错容易在 top-1 已正确时过度编辑，或在正确答案不在候选中时幻觉补写，且常需大模型。

Result: 作者提出 oracle-guided 微调，将 N-best 列表整理成三类监督池：参考在列表中但不在 top-1 的纠错样本教模型选择/改写；top-1 正确的正则样本抑制不必要改写；参考不在候选中的 hard negative 抑制幻觉。该训练比随机子集微调更有效地学习混合 select-or-rewrite 策略。收益在专有数据上的约 3B 参数模型最清楚；CommonVoice 上证据不如专有数据明确。作者认为这一差异可能因为公开朗读文本与 LLM 预训练语料重叠，从而掩盖真实后编辑收益，但摘要将其作为解释性推测。

Insight: 模型需要学会何时保留 ASR top-1，而非一律重写；三类样本分别覆盖改错、克制改动和拒绝补写。数据来源差异影响结论，CommonVoice 上效果不明确，预训练数据重叠解释尚非确定因果。

Abstract: Large language models can post-edit ASR by using linguistic context to compare N-best hypotheses and rewrite the transcript when needed, often improving WER without retraining the recognizer. However, prompt-only correction is brittle: it may over-edit when top-1 is correct and hallucinate when the reference is absent, and often needs very large models. We propose oracle-guided fine-tuning that turns N-best lists into supervision via three curated pools: corrections (reference in the list but not top-1) to teach selection, regularizers (top-1 correct) to discourage unnecessary rewrites, and hard negatives (reference absent) to suppress hallucination. Fine-tuning on this data provides more effective supervision than random subset fine-tuning for learning a hybrid select-or-rewrite policy, with the clearest gains from around 3B parameters on the proprietary dataset but less conclusive results on CommonVoice, a divergence suggesting that public read-speech text may overlap with LLM pretraining and mask genuine post-editing gains.

[273] Have You Been Convinced? A Neuro-symbolic Argumentation Reasoning Approach to Detecting Persuasiveness in ChangeMyView Discussions cs.CL | cs.AI | cs.SCPDF

Siddharth Bhargava, Sara Tonelli, Patricia Martín-Rodilla

TL;DR: NARS 用检索增强的迭代 LLM 抽取双极论证图，再以渐进论证语义传播说服力；在 CMV2BAS 上最佳 F1/AUC 为 0.57/0.56，超过零样本 Qwen3.6:27B 和 GPT-oss:20B。

Details

Motivation: 论证结构描述话语中各论点如何相互作用，但要从结构推断说服效果，需要显式表示影响如何沿论证关系传播。

Result: Neuro-symbolic Argumentation Reasoning System（NARS）用于在线 ChangeMyView 讨论说服力检测。神经模块通过检索增强的迭代 LLM 推理抽取双极论证结构；符号模块使用渐进论证语义沿结构传播影响。作者基于 Webis-CMV-20 构建 CMV2BAS，包含长篇多线程讨论、LLM 派生的双极结构标注及观点变化/未变化标签。研究比较渐进论证语义、加入论证质量与图中心性、以及现代 LLM 下符号推理必要性。最佳设置 F1=0.57、AUC=0.56，优于零样本 Qwen3.6:27B 和 GPT-oss:20B。支持占主导的讨论和长论证链是影响传播的主要失败模式。

Insight: 该工作将说服力建模为结构中的影响传播，并展示神经抽取与符号推理可在噪声真实讨论上结合。数据集论证结构由 LLM 派生，长链和支持主导模式限制效果，需谨慎理解其宏观说服分析能力。

Abstract: Argument structures represent how individual arguments interact within a discourse. However, leveraging these structures to infer meaningful knowledge remains challenging, as it requires a formal representation of how argumentative influence propagates across the structure. We propose Neuro-symbolic Argumentation Reasoning System (NARS) for performing persuasiveness detection directly from online discussions. The neural component employs retrieval-augmented iterative LLM reasoning to extract bipolar argument structures, while the symbolic component applies gradual argumentation semantics to propagate argumentative influence over the structure. Using the Webis-CMV-20 corpus, we construct CMV2BAS, a dataset of long, multi-threaded ChangeMyView discussions annotated with LLM-derived bipolar argument structures and labeled with delta events (view change vs.\ no change). We investigate whether gradual argumentation semantics can effectively detect persuasiveness, whether augmenting them with argument quality and graph-based centrality improves performance, and whether symbolic reasoning remains necessary in the presence of modern LLMs. NARS achieves F1=0.57 and AUC=0.56 under its best configuration, outperforming zero-shot Qwen3.6:27B and GPT-oss:20B. We identify support-dominant discussions and long argumentative chains as key failure modes limiting attack influence propagation. Our work demonstrates the feasibility of operationalizing formal gradual argumentation semantics over noisy real-world discourse for scalable macro-level persuasion analysis, while highlighting key challenges and directions for future research in neuro-symbolic argumentation reasoning. We release both CMV2BAS dataset and the NARS framework for public use.

[274] Lumen-bench: Cross-Lingual Safety Gaps and the Say-Do Problem in Tool-Calling Large Language Models cs.CLPDF

Giovana Kerche Bonás, Thiago Laitz, Thales Sales Almeida, João Guilherme Alves Santos

TL;DR: Lumen-bench 对 30 个工具调用模型进行五语言伦理行为评测，发现提示与语言会显著改变有害行动率；拒绝表述伴随有害工具调用的 say-do gap 占实质调用案例的 8.57%。

Details

Motivation: 模型在文字中拒绝不道德请求，并不保证其执行工具时也拒绝；仅英文对话评测无法覆盖跨语言安全差异及“说拒绝、做了有害操作”的行为分离。

Result: Lumen-bench 覆盖葡萄牙语、英语、西班牙语、中文和印地语，使用 108 个葡萄牙语编写场景、六类社会压力条件、伦理/中性两种系统提示及 21 种通用工具，对 30 个模型开展 194,400 次试验。裁判为 o4-mini，并与一位葡语作者标注（n=83，κ=0.78）校准，也与两种其他供应商裁判比较（n=607，Fleiss κ=0.78），另用沙箱回放核验。结果：同模型对比并经多重检验校正后，伦理提示下印地语和葡语的有害行动率高于英语，中性提示下印地语和西班牙语高于英语；高价模型中位违规率较低，但价格不保证安全或能力；中性提示使汇总违规率由 2.76% 升至 13.64%；8.57% 的实质工具调用案例中，模型一边表达拒绝/伦理异议、一边发出有害工具调用。3,300 次良性控制中，合法任务完成率为 82.8%。

Insight: 语言、系统提示、工具执行及裁判口径都影响安全评估；say-do gap 说明文本拒绝率不足以代表工具行为安全。结果来自葡语作者编写的场景和指定工具集，裁判对 gap 案例数量存在分歧。

Abstract: When a large language model refuses an unethical request, what does refusal mean? In conversation, refusal is bounded by what the model says; once the model is equipped with executable tools, what it does can diverge from what it says, a failure mode that English-only evaluations cannot fully reveal. We introduce Lumen-bench, a behavioral benchmark of tool-calling ethics spanning 30 models, five languages (Portuguese, English, Spanish, Mandarin Chinese, Hindi), 108 Portuguese-authored scenarios, six social-pressure conditions, two system prompts (ethical and neutral), and 21 generic tools. The evaluation comprises 194,400 trials. Judgments come from a reasoning-model judge (o4-mini), calibrated against one Portuguese-speaking author ($n=83$; harmful/non-harmful $κ=0.78$), compared with two other provider judges (Fleiss’ $κ=0.78$, $n=607$), and complemented by retrospective sandbox replay. Four findings emerge. (i) Within-model comparisons support higher harmful-action rates in Hindi and Portuguese than English under the ethical prompt, and Hindi and Spanish under the neutral prompt, after multiple-testing correction. (ii) Premium models have lower median violation rates, although price does not guarantee safety or competence. (iii) Replacing the ethical system prompt with a neutral one raises pooled violation rates from 2.76% to 13.64%. (iv) In 8.57% of substantive-call cases, refusal or ethical objection accompanies a harmful tool call: the judges consistently identify the say-do gap, although they differ in how many responses they assign to this category. Matched benign controls demonstrate legitimate task completion in 82.8% of 3,300 trials. Code and reproduction scripts are available at https://github.com/TropicAI-Research/lumen.

[275] Learning Proposition-Level Links from Sentence-Level Supervision for Neuro-Symbolic NLI cs.CLPDF

Xuyao Feng, Antonis Bikakis

TL;DR: 该神经符号 NLI 方法从句级标签学习命题原子间的隐变量链接，同时固定翻译器和 SAT 推理器；相对原子层直接套用句级模型提升 11.40–27.09 个百分点。

Details

Motivation: 神经符号自然语言推理先把前提和假设分别翻译成命题表示，再进行符号推理；关键步骤是确定两组命题原子间的语义联系。既有 AMR/SAT 系统通常用固定嵌入选源和未训练于原子关系的句级 NLI 模型，难以精确建链。

Result: 作者提出可训练原子链接器，联合预测链接是否存在、来源选择和关系类型。没有原子对标注时，将跨公式链接结构视为潜变量，使用句级 NLI 标签训练链接器，同时固定命题翻译器和 SAT 推理器。评测包括 SNLI 全部公开测试集、SICK 全部公开测试集，以及 MNLI 匹配/不匹配验证集。相较在同一 SAT 管线中直接将句级 NLI 模型用于原子对的基线，准确率提高 11.40–27.09 个百分点；相较先前固定链接器 AMR-to-SAT 系统，SNLI 及 MNLI 两个划分提升 8.14–26.32 个百分点，SICK 上也有较小提升。

Insight: 句级标签可通过潜变量学习命题原子之间的关系，不必额外构造原子对人工标注；固定翻译和 SAT 推理组件使改进归因集中在链接器。各数据集增益不同，SICK 上相较既有系统收益较小。

Abstract: We study neuro-symbolic natural language inference (NLI) with independently translated propositional representations of the premise and hypothesis. Before symbolic reasoning, the system must infer semantic links between their proposition-level units (atoms). Existing systems based on Abstract Meaning Representation (AMR) and Boolean satisfiability (SAT) identify these links using fixed embedding-based source selection and pretrained sentence-level NLI models not trained for atom-level linking. We introduce a trainable atom linker that jointly predicts link existence, source selection, and relation type. Without atom-pair training annotations, we treat cross-formula link structures as latent and train the linker from sentence-level NLI labels, keeping the translator and SAT reasoner fixed. We evaluate on the full public test splits of Stanford Natural Language Inference (SNLI) and Sentences Involving Compositional Knowledge (SICK), and the matched and mismatched validation splits of Multi-Genre Natural Language Inference (MNLI). Our system improves accuracy by 11.40-27.09 percentage points over a baseline applying sentence-trained NLI models unchanged to atom pairs within the same SAT pipeline. It also improves over the prior fixed-linker AMR-to-SAT system by 8.14-26.32 percentage points on SNLI and both MNLI splits, with a smaller gain on SICK.

[276] UDHRBench: Benchmarking Human Rights in Large Language Models cs.CL | cs.AIPDF

Alak Raad, Noreldin Morssy, Zeerak Suhail, Serin Atiani, Mustafa Jarrar

TL;DR: UDHRBench 含覆盖《世界人权宣言》30 条的 1,124 个挑战问题和四种模型的 4,496 条专家标注回答；自动裁判识别人权不对齐的加权 F1 为 89.8%。

Details

Motivation: LLM 被用于需要遵守伦理与法律规范的场景，但其与国际公认人权标准的一致性缺少系统评测。

Result: UDHRBench 将人权规范评估操作化，包含覆盖《世界人权宣言》全部 30 条的 1,124 个挑战性问题，以及 ChatGPT、Grok、Gemini、Fanar 四种模型共 4,496 条人工专家标注回答；标注者一致性较高，κ=0.801。分析发现不对齐常通过看似合理的效率、风险管理或科学中立论证出现，而非明确否定权利。作者提出 LLM-as-a-Judge 自动评估框架，零样本/少样本提示下检测人权不对齐的加权 F1 为 89.8%，并引入 Alignment Leniency Score 衡量裁判如何处理模糊案例。

Insight: 基准强调隐性规范冲突：表面理性的理由也可能导致侵犯权利。裁判 F1 为 89.8%，但模糊案例处理仍需结合 Leniency 指标；模型集合和专家标注范围限定了结论。

Abstract: Large language models are increasingly deployed in contexts requiring adherence to ethical and legal norms, yet their alignment with internationally recognized human rights standards remains underexplored. We introduce UDHRBench, a large-scale benchmark for operationalising the evaluation of large language model compliance with the Universal Declaration of Human Rights. The dataset comprises 1,124 human rights confronting questions covering all 30 articles and 4,496 responses from four large language models (ChatGPT, Grok, Gemini, and Fanar), manually annotated by experts for alignment, with high inter-annotator agreement ($κ= 0.801$). Our analysis reveals systematic and non-trivial misalignments, showing that violations typically arise through seemingly rational reasoning: appeals to efficiency, risk management, or scientific neutrality, rather than explicit rejection of rights. We further propose an LLM-as-a-Judge framework for automated evaluation using zero-shot and few-shot prompting, achieving 89.8% weighted F1-score in detecting human-rights misalignment, alongside a novel Alignment Leniency Score that quantifies how judge models resolve ambiguous cases. The dataset, annotations, and code will be released as open-source resources (https://sina.birzeit.edu/UDHRBench).

[277] PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces cs.CL | cs.AI | cs.IRPDF

M. Yigit Turali, Shreyas Rajesh, M. Efe Lorasdagi, Chenda Duan, Kartik Sharma

TL;DR: PANINI-DR 用共享生成式语义工作区记录子问题、依赖、发现和证据，并分离研究与核验智能体；三个深度研究基准平均表现最强，BrowseComp-Plus 准确率 73.3%，领先次优 16 个百分点。

Details

Motivation: 深度研究要串联搜索、推理和计算，并让中间发现决定下一步调查；调查越长，智能体越需跨调用维护状态并核实证据，以免错误向后传播。

Result: PANINI-DR 将整体调查协调与子问题研究、核验拆开。共享的生成式语义工作区保存子问题及其依赖、发现和支持证据；研究智能体处理子问题，独立核验智能体依据问题及证据检查候选答案，反馈驱动继续调查，已记录发现又指导后续任务分配。在 BrowseComp-Plus、FRAMES、MoNaCo 上与七种先进 harness 比较时，统一语言模型主干及冻结语料，PANINI-DR 平均表现最强。BrowseComp-Plus 准确率 73.3%，比次优高 16 个百分点；困难题上回答数量近两倍于次优。消融显示结构化规划、持久状态及复用前核查每条证据均能改善表现。证据留存方面，三个基准上 answer recoverability 最高，claim-support 审计也显示证据支撑更强。

Insight: 持久化、可核验的中间发现既支持后续研究，也能区分记忆答对与证据答对；把核验放在发现复用前可防错误级联。准确率提升和证据可恢复性为不同维度，结论基于统一主干/冻结语料比较。

Abstract: Deep research requires connecting multiple steps of search, reasoning, and computation, with intermediate findings guiding what to investigate next. As investigations grow, agents must maintain research state across model calls and check findings before errors propagate into subsequent work. We introduce PANINI-DR, a multi-agent harness that separates coordination of the overall investigation from the research and verification of individual subquestions. A shared workspace, built on the Generative Semantic Workspace representation, records subquestions, their dependencies, findings, and supporting evidence. Researchers investigate individual subquestions, while a separate verification agent checks proposed answers against those subquestions and their evidence. Verification feedback guides further investigation, and recorded findings inform subsequent assignments. We evaluate PANINI-DR on BrowseComp-Plus, FRAMES, and MoNaCo against seven state-of-the-art deep-research harnesses, all using a shared language-model backbone and frozen corpora. PANINI-DR achieves the strongest average performance across the three benchmarks. On BrowseComp-Plus, it achieves 73.3% accuracy, outperforming the next-best harness by 16.0 points. Its advantage is strongest on hard questions, where it answers nearly twice as many as the next-best harness. Ablations further show that structured planning and persistent state improve performance, as does checking each finding against its cited evidence before it is reused. Because a model can answer correctly from memory alone, we also analyze the evidence retained by each harness. PANINI-DR achieves the highest answer recoverability from saved evidence on all three benchmarks, while complementary claim-support audits show stronger evidence-backed answers. Accounting for this evidence changes how harnesses compare and further strengthens PANINI-DR’s advantage.

[278] Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning cs.CL | cs.AI | cs.LG | cs.LOPDF

Debargha Ganguly, Weicong Chen, Mohsen Hariri, Vikash Singh, Sreehari Sankar

TL;DR: Sequent-Prover 训练语言模型用可执行 SMT-LIB 草稿与 Z3 多轮交互并输出可验证形式化答案；37 个基准上 4B 单次准确率从 85.6% 升至 97.0%，pass⁴ 从 64.1% 升至 91.7%。

Details

Motivation: 复杂推理需要结果可检查。仅靠语言模型输出自然语言答案难验证其推理是否忠实；即使调用 SMT 求解器，求解器成功也不保证模型形式化语义正确。

Result: Sequent-Prover 是一组使用验证器的语言模型智能体及其后训练方法。智能体在可执行 SMT-LIB 草稿区编写完整程序，多轮调用 Z3、查看反馈、修复形式化表达，最终显式给出答案。Verifier-conditioned trajectory distillation 从带噪教师提议策略开始采样多轮轨迹，保留正确、非平凡、有依据且可执行的成功尾部，再进行监督微调和轨迹级偏好优化。评估同时检查任务准确率、k 样本可靠性及 SMT 程序质量。37 个基准、每模型家族 7,254 道题上，4B 模型单次准确率从 85.6% 提升到 97.0%；四次采样全部正确的可靠性 pass⁴ 从 64.1% 提升到 91.7%。平均 SMT 调用由 4.10 降至 2.52，每次尝试 token 从 18.97k 降到 12.72k。

Insight: 可执行程序和 solver 反馈构成可审计推理过程，但训练筛选同时要求语义正确、非平凡且有依据，避免把求解器能运行等同于答案可信。准确率、pass⁴、工具调用和 token 成本均显示改善，范围为所测 37 个基准。

Abstract: We introduce Sequent-Prover, a family of verifier-using language-model agents, and the post-training framework that produces them. Agents reason in an executable SMT-LIB scratchpad: they write complete SMT-LIB programs, query Z3 across turns, inspect solver feedback, repair formalizations, and terminate with an explicit answer. Our recipe, verifier-conditioned trajectory distillation, bootstraps from a noisy teacher proposal policy by sampling multi-turn trajectories, retaining the successful tail that is correct, non-trivial, grounded, and executable, then applying supervised fine-tuning and trajectory-level preference optimization. Because solver success alone does not guarantee semantic faithfulness, we evaluate task accuracy, k-sample reliability, and SMT program quality. Across 37 benchmarks and 7,254 problems per model family, post-training raises the 4B model’s per-attempt accuracy from 85.6% to 97.0% and four-sample all-correct reliability (pass^4}) from 64.1% to 91.7%. It also reduces average SMT calls from 4.10 to 2.52 and tokens per attempt from 18.97k to 12.72k. Sequent-Prover agents give answers backed by executable formal artifacts.

[279] Do correct answers have stable latent knowledge support? cs.CLPDF

Chuancheng Shi, Shangze Li, Wenhua Wu, Kun Hu, Fei Shen

TL;DR: 研究提出 LSS 和 KTS 衡量知识潜状态及知识关系结构在不同提问方式下的稳定性，并构建含 4,411 项、76,766 次受控查询的 EpiStableBench；同模型内表征越稳定通常越能跨提示答对。

Details

Motivation: LLM 通常按输出正确性评估，但答对并不说明支持答案的内部知识表征稳定。作者希望区分单条知识状态的稳定性与知识之间关系结构的稳定性，并检验它们是否与泛化和知识编辑相关。

Result: Latent State Stability（LSS）测量单个知识的内部表征稳定度，Knowledge Topology Stability（KTS）评估同类知识两两距离排序在不同条件下的一致性。作者先在各 elicitation 条件下分别平均同一知识的内部表示，以基线条件平均表示为参照，移除共享偏移；LSS 为校正表示与基线参照的距离，并按同类型知识典型距离归一化；KTS 衡量跨条件的知识间有序距离一致性。EpiStableBench 含 4,411 项事实知识与数学推理内容及 76,766 次控制查询，覆盖四个家族的 16 个 LLM。即使基线回答正确，稳定性也因题目而异；同一模型内，稳定性更高通常对应留出提问条件下更高准确率，但跨模型并不随规模或总体准确度稳定上升。作者另提出 LatentDPO，结合知识稳定性学习编辑策略，在多种已有编辑方法上持续提高编辑成功率和不同表达泛化。

Insight: 正确输出与稳定内部表征不是一回事；LSS/KTS 将两者分开测量，且稳定性和模型规模/整体精度之间无简单跨模型单调关系。LatentDPO 的改善为不同知识编辑法上的实验结论，摘要未给具体提升值。

Abstract: Large language models are typically evaluated in terms of the correctness of their output. However, output correctness alone does not confirm whether the latent knowledge supporting an answer has stable internal representations. We therefore propose latent state stability (LSS) and knowledge topology stability (KTS) to investigate the stability of individual knowledge states and their relational structure, respectively. Specifically, we first separately average the model’s internal representations of the same knowledge under each elicitation condition. Using the mean representation under the baseline condition as a reference, we remove shared shifts to obtain corrected representations. We then formulate LSS as the distance between each corrected representation and its reference representation, normalized against the typical distance between knowledge items of the same type. We formulate KTS as the cross-condition consistency of the ordered pairwise representation distances among knowledge items of the same type. Next, we construct EpiStableBench, which comprises 4,411 items spanning factual knowledge and mathematical reasoning, with 76,766 controlled queries. Across 16 LLMs from four families, representation stability varies substantially even among items answered correctly under the base condition. Within models, higher stability generally corresponds to higher accuracy under held-out elicitation conditions; across models, stability does not consistently increase with model size or overall accuracy. Meanwhile, we present LatentDPO to learn a knowledge-editing policy that accounts for the stability of latent knowledge. Experimental results show that integrating LatentDPO with multiple existing editing methods consistently improves editing success and generalization across different expressions.

[280] North or South? Diagnosing LLMs’ Understanding of North Korean Language, Society, and Knowledge cs.CL | cs.AIPDF

Kiwoong Park, Seyoung Song, Donghyun Woo, Alice Oh

TL;DR: LENS 是含 1,765 题的人类验证基准，用于测量 LLM 对朝鲜语言、社会文化语境和专门知识的理解；模型语言输入性能下降有限，但社会文化对齐差距更大、领域知识不足严重。

Details

Motivation: 多语言问答模型发展迅速，但低资源社会语言语境中的理解能力仍少被研究。朝鲜半岛长期分治造成朝鲜与韩国在词汇、正字法、规范和意识形态话语方面差异，适合作为诊断环境。

Result: 作者提出 LENS（Language and sociocultural Evaluation for North Korean Society），含 1,765 名人类验证问题，评估语言鲁棒性、社会文化视角对齐和领域专门知识三个维度。对多种近期 LLM 的评估发现：朝鲜语输入带来中等程度性能下降；社会文化视角对齐差距更大；朝鲜领域知识存在严重限制。定性分析显示模型常把正式制度与日常现实混为一谈、投射自由民主假设或依赖粗略刻板印象。摘要未列模型名单、具体分数或各维度题量。

Insight: 语言形式相近并不保证社会文化或领域知识也理解充分；基准将语言能力、视角对齐和知识掌握拆开，强调需以有根据的评价资源覆盖少数文化/政治语境。结论范围受 LENS 题目构成限定。

Abstract: Large Language Models (LLMs) have made substantial progress in multilingual question answering, yet their ability to understand low-resource sociolinguistic contexts remains under-explored. We study North Korea as a unique testbed where decades of division have produced distinct differences from South Korea in vocabulary, orthography, norms, and ideological discourse. We introduce LENS (Language and sociocultural Evaluation for North Korean Society), a human-validated benchmark of 1,765 questions evaluating LLMs across three dimensions: linguistic robustness, sociocultural perspective alignment, and domain-specific knowledge. Evaluating diverse recent LLMs, we find a modest performance drop on North Korean language inputs, a larger gap in sociocultural perspective alignment, and severe limitations on domain-specific knowledge. Qualitative analysis reveals that models often conflate formal institutions with lived realities, project liberal-democratic assumptions, or rely on coarse stereotypes. Our results demonstrate that linguistic similarity alone does not guarantee sociocultural or domain-specific understanding, highlighting the critical need for grounded evaluation resources for underrepresented cultural and political contexts.

[281] A Sovereign, Indic-First, Reasoning Model cs.CL | cs.AIPDF

Sutanshu Saksena Raj, Prasanjith Pasunuti, Parijat Chatterjee, Daksh Maan, Purusharth Malik

TL;DR: Gnani-Evon-v3.3-30B-A3B 扩词表使印度语言平均 fertility 降低 38.6%，并经四类可验证任务 RL 提升推理；但缺少目标语言奖励时，印地语抽取回答转成英文的比例从 8.7% 升至 91.7%。

Details

Motivation: 构建面向印度语言的长上下文推理模型，需要兼顾英语与八种印度语言、词汇覆盖和后训练奖励。该报告强调奖励若只衡量内容正确而不约束回答语言，可能破坏语言遵循。

Result: Gnani-Evon-v3.3-30B-A3B 从 Nemotron-3-Nano-30B-A3B 初始化，保留 52 层骨干，修改词表、数据和后训练目标。词表原位扩展从 131,072 到 163,072 token，使印度语平均 fertility 降低 38.6%；新 embedding 行由上游分解初始化并预热，然后更新主干。持续预训练加入英语—印度语数据、优质退火数据，并直接使用 131,072 长度。四种可验证任务环境上的 RL 提升数学、研究生科学和指令遵循，却因奖励不要求目标语言而损害印度语抽取问答：英语回答比例由 8.7% 增至 91.7%；仅在仍以目标文字回答的自选子集上，准确率约翻倍。未发布研究显示加入印度语奖励可恢复语言规范；26B-A4B 生成式裁判在 RM-Bench 上超过 235B 模型，对其蒸馏恢复 81% 被奖励优化挤出的能力。相对同架构初始模型，发布版印度语综合提升 5.2 分、英语保持下降 1.7 分；未发布印度语感知分支提升 14.6、英语下降 5.4。规划中的 FP8 版本是 32GB，32 个配对基准均值变化近零。

Insight: 奖励遗漏输出语言这一约束，会让强化学习用英文回答获取分数，尽管目标语能力可能部分保留；按目标语言奖励可缓解，但相关修复研究尚未发布。公开发布版和未发布分支的能力折衷须分别看，目标语条件准确率基于自选子集。

Abstract: Gnani-Evon-v3.3-30B-A3B is a long-context hybrid Mamba-Transformer mixture-of-experts model for reasoning, instruction following, and tool use in English and eight Indian languages. Starting from NVIDIA’s Nemotron-3-Nano-30B-A3B checkpoint, it retains the 52-layer backbone and changes the vocabulary, training data, and post-training objectives. An in-place expansion from 131,072 to 163,072 tokens reduces mean Indian-language fertility by 38.6%. The new embedding rows are initialized from their upstream decompositions and warmed before the backbone is updated. Continual pretraining then adds broad English-Indic data, high-quality annealing, and direct training at a 131,072-token sequence length. The central result appears after post-training. Reinforcement learning across four verifiable task environments improves mathematics, graduate-level science, and instruction following, but degrades Indic extractive question answering because the rewards do not require an answer in the requested language. The share of Indic extraction items answered in English rises from 8.7% to 91.7%. Accuracy conditional on answering in the target script roughly doubles, which is consistent with preserved reading ability but is measured on a self-selected subset. Two unreleased studies test repairs. A scoped Indic reward restores the language convention. A 26B-A4B generative judge outperforms a 235B-parameter model on RM-Bench, and distillation restores 81% of the capabilities displaced by optimizing against it. Relative to the architecture-identical starting checkpoint, the released model gains 5.2 points on an Indic aggregate while giving up 1.7 points of English retention. The unreleased Indic-aware branch trades further, gaining 14.6 Indic points at a cost of 5.4 in English. A planned FP8 release reduces the artifact to 32GB with near-zero mean change across 32 paired benchmarks.

[282] From Candidate Tokens to Final Outcomes: Decoding Relative Long-Horizon Candidate Value from Reasoning Model Representations cs.CLPDF

Yixin Yang, Yue Wang, Fandong Meng, Zhifang Sui

TL;DR: RCTV 数据集将推理前缀、合理下一 token 与长程成功价值配对；稀疏线性探针显示该价值可从多层隐藏态、MLP 和注意力输出解码，受控在线引导可提升最终答对机会。

Details

Motivation: 推理模型逐 token 生成轨迹，同一前缀下多个候选 token 都可能合理，但它们对后续推理影响及最终正确概率不同；这种长程价值是否已编码在候选位置的内部表示中尚不明确。

Result: 作者定义相对长程候选价值，度量同一前缀内下一 token 候选之间的差异，并构建数学推理数据集 RCTV，配对前缀、合理候选 token 及其经验长程价值。使用稀疏线性探针检查候选位置的隐藏状态、MLP 输出和注意力输出，结果显示多个层的表示均能线性解码相对候选价值；探针也优于基于 token 概率和嵌入的基线。受控在线干预实验表明，在生成时用解码信号指导 token 选择可提高最终成功概率。摘要未提供数据规模或数值提升。

Insight: 局部候选表示可能携带超越下一个 token 概率的长程结果信息；线性可解码性和受控干预共同支持其行为相关性。但数学任务范围、候选价值估计方式及干预增益需结合正文数值理解。

Abstract: Large reasoning models generate reasoning trajectories token by token. At a given prefix, multiple next-token candidates may be plausible, yet shape subsequent reasoning and differ in their chances of leading to the correct answer. Whether these differences are reflected in internal representations remains unclear. To study this question, we define relative long-horizon candidate value to quantify within-prefix differences and construct the Reasoning Candidate-Token Value (RCTV) dataset, pairing mathematical reasoning prefixes with plausible next-token candidates and their empirical long-horizon candidate values. Using RCTV, we examine hidden states, multilayer perceptron outputs, and attention outputs at candidate positions with sparse linear probes. Experimental results provide evidence that relative long-horizon candidate value is linearly decodable from these representations at multiple layers. Candidate-value probes can also outperform probability- and embedding-based baselines. Furthermore, a controlled online intervention experiment shows that using the decoded signals to guide token selection during generation can improve the chances of eventual success. Our findings provide evidence that internal representations encode information about relative long-horizon candidate value, offering insights into how local candidate-level representations relate to long-horizon reasoning outcomes.

[283] AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning cs.CL | cs.LGPDF

Peiqi Jiang, Fan Zhang, Xin Xu, Tianhao Chen, Can Yang

TL;DR: AlphaPilot 用多轮工具调用 RL 智能体提出、评估并维护金融因子池，支持逐日适应市场状态；CSI300/CSI500 滚动样本外回测的均值总收益为 42.93%/54.19%。

Details

Motivation: 公式化 alpha 挖掘需同时处理非结构化市场描述、历史数据以及候选因子的定量评估，现有方法往往只利用其中部分信号。

Result: AlphaPilot 训练多轮工具使用 LLM 智能体进行自适应因子挖掘：智能体依据市场描述提出因子公式，查询真实数据评估反馈，并在多个回合中更新保留因子池；推理时逐日搜索和更新，以适应市场制度变化。三次独立滚动样本外回测中，CSI300/CSI500 的平均总收益分别为 42.93%/54.19%，平均 Sharpe 分别为 2.46/2.86。在两个股票池上，AlphaPilot 在被比较方法中平均总收益、Sharpe 和 RankIC 均最高。

Insight: 多轮评估反馈和因子池积累让搜索可随市场变化持续调整，而不只一次性生成候选因子。金融回测结果高度依赖时期、交易成本及执行设定；摘要未给回测窗口、成本和基准详情。

Abstract: Formulaic alpha mining seeks interpretable expressions that rank assets by future returns. The task is challenging because it demands simultaneous processing of unstructured market descriptions, historical data, and quantitative evaluation of candidate factors, yet most existing methods exploit only a subset of these signals. We introduce AlphaPilot, a reinforcement-learning framework that trains multi-turn tool-using LLM agents for adaptive alpha mining. The agent conditions on market descriptions, proposes factor expressions, queries real-data evaluator feedback, and updates a retained factor pool across turns. This design jointly leverages market text, LLM financial knowledge, real data, and evaluation feedback. At inference time, AlphaPilot performs day-by-day search and pool updates to adapt to shifting market regimes. Across three independent runs in out-of-sample rolling backtests on CSI300 and CSI500, AlphaPilot achieves mean total returns of 42.93% and 54.19%, with mean Sharpe ratios of 2.46 and 2.86, respectively. It attains the highest mean total return, Sharpe, and RankIC among the evaluated methods on both universes.

[284] Apertus 1.5: From Open Text Models to Open Multimodal Reasoners cs.CLPDF

Project Apertus, :, Yixuan Xu, Yassine Turki, Vinko Sabolčec

TL;DR: Apertus 1.5 将全开放 Apertus 模型扩展到图像/音频理解与长上下文推理，使用额外 4T/2T token、262,144 上下文和分阶段后训练；推理模式使 8B 数学能力最多提高 20 分。

Details

Motivation: 语言模型逐渐成为基础设施，依赖少数供应商并非唯一选择。本文报告如何在公开数据、文档化训练流程及开放许可下，将 Apertus 从文本模型扩展为多模态推理模型。

Result: Apertus 1.5 从 Apertus 1.0 继续预训练：8B 增加 4T、70B 增加 2T token，加入图像和音频理解，将上下文窗口扩展四倍到 262,144，架构不变；多模态信息由 tokenizer 引入，语言模型头仍只输出文本。随后进行监督微调、长上下文微调、可验证奖励 RL 和偏好对齐。Thinking 是聊天模板标志而非独立模型，8B 数学推理最多提升 20 分。相较前代模型，几乎所有基准均改善；同类全开放 8B 在知识、指令遵循和文化基准领先，指令遵循超过所有比较模型（含开放权重模型）。图像/音频理解可与全开放多模态模型竞争，文档理解及长篇/低资源语音识别较强，但专用视觉/音频模型在感知及英语等高资源语音转录上仍领先。训练只用开放数据，追溯尊重爬虫 opt-out 并排除限制许可内容；发布权重、数据准备/训练配方和评测套件。

Insight: 报告同时强调能力边界与开放治理：泛化多模态能力已具竞争力，但专用感知/高资源转录仍有差距。数学提升最多 20 分是 8B 的特定结果；开放数据许可及训练透明度是模型发布设计的重要部分。

Abstract: Language models are becoming infrastructure, and infrastructure built in the open, by public institutions, on documented data, is the alternative to dependence on a few providers; Apertus shows this is viable at scale. Apertus 1.5 extends the fully open Apertus family from text to multimodal reasoning. Continued pretraining of the Apertus 1.0 models, on 4T additional tokens for the 8B and 2T for the 70B, adds image and audio understanding and extends the context window fourfold to 262,144 tokens; multimodality enters through the tokenizer, the language-model head remains restricted to text, and the architecture is unchanged. Supervised finetuning, long-context finetuning, reinforcement learning with verifiable rewards, and preference alignment complete the released models. Thinking is a flag of the chat template rather than a separate model, and lifts mathematical reasoning by up to 20 points on the 8B model. The released models improve over their predecessors on nearly every benchmark; among fully open models of its class, the 8B leads on knowledge, instruction following, and cultural benchmarks, and its instruction following surpasses all compared models, including open-weight ones. On image and audio understanding, a first for the family, the models are competitive with fully open multimodal models, strongest on document understanding and long-form and lower-resource speech recognition, while specialized vision and audio models remain ahead on perception and transcription of English and other high-resource languages. As in Apertus 1.0, we train only on openly available data, respect crawler opt-outs retroactively, and exclude restrictively licensed material. We release the model weights, data preparation and training recipes, and evaluation suites under permissive licenses; openness extends to values, with the Apertus charter guiding alignment stages published alongside the models.

[285] The Fragility and Dynamics of LLM Pruning-Recovery Steps cs.CLPDF

Luis Frentzen Salim, Komang Rinartha, Chia-Chun Wu, Tran Van Nhiem, Lun-Wei Ku

TL;DR: 该研究联合考察 LLM 剪枝与恢复阶段，发现指令遵循在剪枝后最先崩溃，而域内剪枝损失可能主要是无法连贯响应提示；其设置下域内数据最适合剪枝、指令式数据最利于恢复。

Details

Motivation: 结构化剪枝加恢复微调是压缩 LLM 的常见流程，但两个阶段通常独立设计。它们的顺序、训练数据和目标对齐如何共同影响压缩模型尚不明确。

Result: 作者系统扫描剪枝与恢复的顺序、各阶段使用的数据，以及训练目标对齐。实验显示指令遵循比知识、推理和基础语言建模能力更早因剪枝崩溃；域内剪枝损失很大部分可能不是知识遗失，而是与领域无关的提示响应不连贯。在该实验设置下，域内数据是最强剪枝信号，指令风格数据最适合恢复；阶段先后顺序影响较小但仍有差异。作者还分析剪枝前后重要性排序漂移，并据此提出组合剪枝与恢复的实用配方。摘要未提供具体模型、数据集或量化数值。

Insight: 评估压缩不应只看知识/困惑度，还应重点监测指令遵循；剪枝导致的能力损失与恢复所需数据类型不同。结论特定于作者设置，不能据摘要推出剪枝阶段普遍无关或所有模型都适用相同配方。

Abstract: Structural pruning followed by a recovery stage is the standard recipe for compressing large language models (LLMs), yet the two stages are typically designed and tuned in isolation. In this work, we study how their interaction shapes the compressed model, sweeping ordering, the data that drives each stage, and the alignment of their training objectives. We find that instruction-following is the first capability to collapse under pruning, well before knowledge, reasoning, or core language modeling, and that much of the pruning in-domain loss is likely not knowledge loss but a domain-agnostic failure to respond coherently to the prompt. We find in our setting that in-domain data is the strongest pruning signal, whereas instruction-style data recovers the most. Stage order matters little, but only marginally. We further analyze how the importance ranking drifts before and after pruning. Together these results give practitioners a concrete recipe for composing pruning and recovery.

[286] Towards Autopilot Labeling for Short Video Content Governance cs.CLPDF

Junwen Chen, Hongwei Wang, Zixuan Wang, Jingbing Wang, Gorden Liu

TL;DR: 本文提出 L0–L3 四阶段短视频治理自动标注流程：从专家标注、监督式 MLLM、领域自适应零样本标注到工具型智能体处理模糊案例；真实应用人工成本最多减少 85%。

Details

Motivation: 短视频快速增长使内容治理标注负担增加；人工标注效率低、迭代慢且成本高，尤其复杂或定义模糊案例需专家判断。

Result: 框架逐步提升自动化程度：L0 仍以人工专家标注为基准；L1 使用有监督微调的判别式 MLLM 直接预测短视频标签；L2 使用在短视频数据上领域自适应预训练的 MLLM 做零样本标注，以减少任务特定监督；L3 采用配备多种工具的智能体工作流处理复杂、陌生或边界不清的案例。多类内容问题实验显示，各阶段标注准确率逐步提升；真实应用中人工劳动成本最多减少 85%。

Insight: 分级流程将简单样例交给模型，复杂模糊样例交给工具型智能体，目标是逐步减少专家投入。摘要未给各阶段准确率、类别或成本统计基线，85%为最高人工成本降幅。

Abstract: The explosive growth of short videos has presented significant challenges to content governance. Traditional manual labeling is inefficient, slow to iterate, and costly. Therefore, we explore and develop a multi-stage autopilot labeling framework aimed at progressively increasing the level of automation in short video content annotation. To move beyond the heavy reliance on human expert labeling at the initial L0 stage, we first introduce a discriminative MLLM labeler at the L1 stage, which is fine-tuned with supervised data to directly predict labels for short videos. Advancing to the L2 stage, we shift to a domain-adaptive MLLM pretrained on short video data and use it for zero-shot labeling, enabling flexible generalization without task-specific supervision. In the L3 stage, we propose an agentic annotation workflow, and equip the agent with several key tools, which empower the agent to handle complex, unfamiliar, or vaguely defined cases, significantly reducing the need for human intervention. We conduct extensive experiments across various content issues, and results demonstrate that each stage of the autopilot labeling framework progressively improves labeling accuracy, and has significantly reduced manual labor costs in real-world applications by up to 85%.

[287] TL;QR: Querying Temporal Discursive Patterns in Textual Corpora cs.CL | cs.CY | cs.DB | cs.LOPDF

François Hublet, Vladimir Krsmanovic, Jesus Solano, Ahmed Bouhoula, Ana Oliveira da Costa

TL;DR: TL;QR 将 LLM 局部文本判断与正则表达式、时间逻辑符号推理组合成可审计声明式查询语言；在合成时序基准上优于纯 LLM 基线，并用于分析十年欧洲议会辩论。

Details

Motivation: 大型文本语料中许多问题关注随时间演化的论述模式，例如政治团体立场从支持逐渐转为批评。仅让 LLM 直接回答难以保证长期时序/关系推理精确，也难追溯结论来源。

Result: TL;QR（Too Long; Query with logical temporal Reasoning）是声明式查询语言和开源引擎。一条查询可组合由 LLM 判断的自然语言谓词、正则表达式和时间逻辑算子。LLM 仅回答单段文本的局部问题，符号引擎再沿时间聚合判断，并把每个匹配关联到原文段落和谓词判定；因此时序与关系推理可审计。查询若包含正则预过滤，只有语料的一小部分需要送给 LLM。在具备时序和语义复杂模式的受控合成基准上，TL;QR 准确度显著高于纯 LLM 基线。案例研究处理欧洲议会十年、442,593 篇多语种发言，识别各政治团体对 GDPR 隐私法规立场变化的候选事件。

Insight: 用 LLM 做局部语义识别、用符号引擎做精确时序聚合，兼顾语义灵活性与结果可追溯性；案例结果是候选立场变化，不等同于人工验证后的最终政治分析。

Abstract: Many questions about large textual corpora concern patterns that unfold over time, such as a political group whose stance on a topic drifts from support to criticism over several years of debates. We present TL;QR (Too Long; Query with logical temporal Reasoning), a declarative query language and open-source engine for finding such temporal discursive patterns. In TL;QR, a single query combines natural-language predicates, whose truth is judged by a Large Language Model (LLM), with regular expressions and the operators of a temporal logic. The LLM answers only local questions about individual text passages; a symbolic engine then aggregates these judgments over time and links every match to the recorded passages and predicate decisions behind it. With this division of labor, the temporal and relational reasoning stays exact and auditable even over long time horizons. In addition, when a query includes a regular-expression pre-filter, only a small fraction of the corpus ever reaches the LLM. On a suite of controlled synthetic benchmarks with temporally and semantically complex patterns, TL;QR is substantially more accurate than LLM-only baselines. In a case study on ten years of European Parliament debates (442,593 multilingual speeches), TL;QR identifies candidate shifts in the political groups’ stances on the GDPR privacy regulation.

[288] RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution cs.CLPDF

Fan Zhang, Hao Wu, Jinyang Wu, Hongyu Qu, Zheng Lian

TL;DR: RSIPO 让智能体策略从轨迹中自提技能并共同进化，同时用经验贝叶斯收缩校准教师分数可信度；Qwen2.5-7B 在 ALFWorld 平均成功率 98.6%，比次优基线高 2.5 个百分点。

Details

Motivation: 长程智能体 RL 需要有效信用分配。RLVR 与 on-policy self-distillation 结合轨迹级优势及 token 级监督已有进展，但常将技能视为静态特权教师上下文，忽略技能在线演化和教师分数可靠性。

Result: Recursive Self-Improving Policy Optimization（RSIPO）让策略和智能体技能在训练中共同进化。策略从已完成轨迹中抽取技能，并通过轨迹回放显式优化技能生成能力；经验贝叶斯收缩机制根据教师指导可靠性动态校准学生对打分的信任，因为加入技能上下文的教师有时也不可靠。Qwen 系列在具身交互和搜索型智能体任务上实验中，RSIPO 超过所有比较基线。Qwen2.5-7B 在 ALFWorld 平均成功率为 98.6%，较第二名高 2.5 个百分点、较标准 GRPO 高 17.4 个百分点。

Insight: 技能不是固定教师提示，而是可由自身轨迹抽取、迭代修订的策略组成部分；同时显式估计教师评分可靠性以防错引导。98.6%及比较差值对应 ALFWorld，具身/搜索任务其余结果未给具体分数。

Abstract: Recent focus in agentic reinforcement learning (RL) seeks to improve credit assignment in long-horizon tasks. One promising direction combines reinforcement learning with verifiable rewards (RLVR) and on-policy self-distillation (OPSD), integrating trajectory-level advantages with dense token-level supervision. This line of work often conditions a teacher model on privileged context, such as agentic skills, to score tokens generated by the student policy. Despite their progress, they largely overlook the on-policy evolution of agentic skills and the reliability of teacher supervision. Towards this end, we introduce Recursive Self-Improving Policy Optimization (RSIPO), a self-evolving agentic RL framework that enables agentic skills and the policy model to co-evolve throughout the training process. Specifically, RSIPO enables the policy itself to extract skills from completed trajectories and explicitly optimizes its skill-generation capability through trajectory replay. It further employs empirical Bayes shrinkage to dynamically calibrate the student’s trust in teacher scores, accounting for the fact that skill-enhanced teacher guidance can sometimes be less reliable. Experiments with Qwen-family models on embodied interaction and search-based agentic tasks demonstrate that RSIPO outperforms all the compared baseline approaches. With Qwen2.5-7B as the base model, RSIPO achieves an average success rate of 98.6% on ALFWorld, yielding absolute improvements of 2.5% and 17.4% over the second-best baseline and standard GRPO, respectively.

[289] Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines cs.CL | cs.AIPDF

Yirui Zhang, Yash Raj Shrestha

TL;DR: 跨 12 种开源模型、48 个数学/知识/代码/推理配对测试显示，答案复核仅在 96 个方向中 9 个超过更强模型本身；同样调用预算下，模型选择器准确率高 2.7–4.2 分且更便宜。

Details

Motivation: 复核模型回答通常按其相对生成模型的增益评价，但这种比较可能把真正的验证能力与更强复核模型直接提供答案的能力混为一谈。

Result: 作者区分 verification（准确率超过任一单模型）与 capability transfer（复核模型提供它自己本可给出的答案）。数学示例中，Qwen3-8B 复核 Llama-3.1-8B 提升 33.2 分，但 Qwen3-8B 单独回答仍高 6.4 分。对 12 种 3B–72B 开源模型在数学、知识、代码、推理上的 48 对组合、96 个复核方向评估，只有 9 个方向超过更强模型单独表现；模型能力差距越大，复核越差。原因是复核者通常跟随生成器，舍弃自己本来正确的答案。复核成本为更强模型单独运行的 2.6–2.8 倍；使用同样两次调用的选择器比复核准确率高 2.7–4.2 分且更便宜。

Insight: 若不以更强模型单独结果为基线，复核流程的增益可能只是把更强模型的能力转移给弱生成器，而非验证真正改善。研究建议报告对强模型基线的比较；结果覆盖给定模型配对和任务集合。

Abstract: Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result. Reviewing another model’s answer is usually judged by its gain over the generator. We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own. Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap. The reason is that the reviewer mostly follows the generator, discarding its own correct answers. Review costs 2.6 to 2.8 times as much as the stronger model alone, while a selector using the same two calls is 2.7 to 4.2 points more accurate than review and cheaper. Studies of verification should report the stronger model alone, not only the gain over the generator. Code and per-question results: https://github.com/Kamilaabb/verification-or-capability-transfer

[290] RIVO: Lightweight Adaptive Rerouting for Deep Research Agents cs.CLPDF

Mingchen Li, Jiatan Huang, Hong Yu

TL;DR: RIVO 是无需重训主干的深度研究轨迹纠偏插件，从轻量研究状态判断何时重定向，并在触发时干预 Host token logits；多基准实验报告改善长程研究表现，但摘要未给分数。

Details

Motivation: 深度研究智能体可能在证据不足时过早停止、信息充分后仍继续搜索，或陷入无效轨迹。现有方案需要主模型重训和密集逐轮监督，或使用未针对研究效用优化的启发式规则。

Result: RIVO 将自适应 rerouting 定义为状态条件轨迹校正：智能体从轻量信号估计研究状态，判断是否需要改道；触发后插件修改 Host token logits，使生成转向更有前景的研究方向。理论分析指出，在价值对齐和局部曲率有界条件下，足够小的 logit 干预可局部提升预期下游研究效用。作者构建 BrowseCore，包含整理后的数据集和配对本地语料以支持受控实验；插件根据下游结果训练学习纠偏。多个深度研究基准和控制设置实验显示，轻量 logit 干预可在不重训 Host 的情况下提升长程研究表现，摘要未列具体结果值。

Insight: 把停止/继续/重定向作为研究状态条件下的轨迹校正，可避免重训大模型并超越固定启发式；理论保证依赖价值对齐和曲率有界等假设。摘要未提供具体基准分数、调用成本或插件规模。

Abstract: Deep Research agents rely on long-horizon reasoning and iterative tool use to solve complex information-seeking tasks. However, adaptive decision-making over these trajectories remains challenging, as agents may stop prematurely when further evidence is needed, continue unnecessarily after sufficient evidence has been gathered, or fail to redirect unproductive trajectories. Existing solutions either require costly model retraining and dense turn-level supervision, or rely on heuristic control rules that are not directly optimized for downstream research utility. This motivates a lightweight, utility-aware alternative for Deep Research control. We formulate adaptive rerouting in long-horizon Deep Research as a state-conditioned trajectory correction problem. Based on this view, we introduce RIVO, a lightweight plug-in framework for adaptive trajectory correction. The agent estimates its research state from lightweight signals and determines whether trajectory redirection is needed. Once triggered, the plug-in modifies the Host token logits to steer generation toward a more promising research direction. Theoretical analysis shows that, under a value-alignment condition and bounded local curvature, a sufficiently small logit intervention can locally improve expected downstream research utility. We further construct BrowseCore, a curated dataset with a paired local corpus, for efficient and controlled Deep Research experimentation. Building on this resource, the plug-in is trained from downstream outcomes, enabling it to learn corrections that improve subsequent research and final task performance. Extensive experiments across multiple Deep Research benchmarks and control settings demonstrate the effectiveness of our method, showing that lightweight logit intervention can improve long-horizon research without retraining the underlying Host model.

[291] More Reasoning Is Not Always Better: Executing Clinical Guidelines with LLMs and Rules cs.CLPDF

Xiaoqian Lu, Guangfu Guo, Long Cheng, Kunpeng Liu

TL;DR: MedDEx 将病历理解与指南执行分离：LLM 抽取结构化患者状态并生成计划，确定性规则执行器承担指南决策逻辑；多项临床指南基准上准确率提高、推理成本降低。

Details

Motivation: 临床决策支持模型常用多步推理或多智能体执行指南，增加计算成本，却未可靠避免决策错误。更多 LLM 推理并不必然带来更好的指南遵循。

Result: Medical Decision Execution（MedDEx）将结构化患者状态与确定性规则执行结合，支持按临床指南制定治疗计划。框架把病历信息解析和指南规则执行分开：LLM 负责抽取患者信息并生成治疗计划，规则执行器处理指南明确规定的决策逻辑。多个临床指南基准实验显示，MedDEx 提高决策准确率并降低推理成本；进一步分析强调结构化患者表示和确定性规则执行的重要性。摘要没有列出基准名称、效果数值或指南适用范围。

Insight: 对可形式化的指南规则，显式执行规则可能比增加自由形式 CoT/智能体步骤更可靠且更省成本；LLM 仍承担病历解释与计划生成。摘要仅报告总体结果，不能据此推断特定临床指南上的真实患者结局。

Abstract: Large language models (LLMs) have shown considerable promise in clinical decision support. However, existing approaches often rely on multi-step reasoning or multi-agent collaboration to execute clinical guidelines, introducing substantial computational overhead without reliably preventing decision errors. To address these limitations, we propose a framework for Medical Decision Execution (MedDEx) that integrates structured patient states with deterministic rule execution to support guideline-based treatment planning. MedDEx separates clinical record interpretation from guideline execution, leveraging LLMs to extract patient information and generate treatment plans while assigning guideline-defined decision logic to a rule executor. Experiments across multiple clinical guideline benchmarks demonstrate that MedDEx improves decision accuracy while reducing inference costs. Further analyses highlight the critical roles of structured patient representations and deterministic rule execution, suggesting that additional LLM reasoning does not necessarily translate into better guideline adherence.

[292] PAIR: Memory-Efficient and Fast Inference with Reading–Reasoning Fusion cs.CLPDF

Sejun Park, Haocheng Xia, Yongjoo Park

TL;DR: PAIR 用小型 Reader 读完整长上下文、大型 Reasoner 只看问题与生成序列，并在 logits 层融合；HotpotQA 上 32B Reasoner 保留全上下文准确率的 96%，KV 内存约降至四分之一。

Details

Motivation: 上下文问答若由一个大模型同时提取证据并推理答案，显存和计算成本高。单模型 token 缩减方法通常需在精度和效率间折衷。

Result: PAIR 用两个不同规模模型分工：较小 Reader 处理并保留完整上下文，较大 Reasoner 只接收问题和已生成序列，不直接注意原文。每个 token 步两者都输出预测，在 logits 层组合，整体行为近似单一 next-token 预测模型；进一步微调可提升表现。HotpotQA 上配 32B Reasoner 时，PAIR 保留该 32B 模型全上下文准确率的 96%，KV 内存约为四分之一。在 45.7K token 上下文、同一 GPU 上，端到端吞吐最高为全上下文 32B 模型的 3.2×。其精度—内存和精度—吞吐折衷也优于所评估 token 缩减基线，效果可迁移到 Qwen2.5、OLMo-3 和 Gemma-3 家族。

Insight: 小模型负责上下文读取、大模型负责推理，再通过分数融合把两者组合，减少大模型 KV 缓存而保留大部分全上下文准确率。效率数字针对 45.7K token 和指定 GPU，跨模型家族泛化是摘要所报告范围。

Abstract: For context-based question answering, using one large model for both evidence extraction and answer reasoning requires substantial memory and computation. Existing approaches trade accuracy for efficiency within a single model only. We propose a multi-model approach for efficient inference: our method, PAIR, employs two differently sized models for evidence extraction and answer reasoning, respectively. A smaller Reader model processes and retains the full context, while a larger Reasoner model sees only the question and generated sequence without directly attending to the context. These capabilities are fused at the logit level: at each token, both Reader and Reasoner produce a prediction, and their logits are combined into one, behaving as a single next-token-prediction model. Fine-tuning further boosts their performance. On HotpotQA, PAIR with a 32B Reasoner retains 96% of that same 32B model’s full-context accuracy while using roughly one quarter of its KV memory. At 45.7K-token contexts, PAIR achieves up to 3.2$\times$ higher end-to-end throughput than the full-context 32B model on the same GPU. PAIR also improves the accuracy-memory and accuracy-throughput trade-offs over evaluated token-reduction baselines. Its accuracy-preserving behavior generalizes across Qwen2.5, OLMo-3, and Gemma-3 model families.

[293] Of Errors and Echoes: Responsibility-Aware Credit Assignment for Tool-Use Agents cs.CL | cs.AIPDF

Xingjian Diao, Tianyu Yang, Wenjun Huang, Xiangchi Yuan, Chunhui Zhang

TL;DR: ECHO 将责任识别与信用分配分开，区分新错误、继承后果、漏纠错和成功修正，并限制自致错误修复奖励；4B/8B 在六项工具基准的每个规模上五项得分最高。

Details

Motivation: 工具智能体 RL 即使有逐步反馈，仍难精细分配信用：观察到的错误可能来自更早决策；多个决策可能共享一个后果，也可能通过后续恢复改变结果。

Result: Error-Chain-Aware Optimization（ECHO）先识别决策责任，再分配奖励。冻结裁判结合事后轨迹分析与决策时策略可见前缀，区分新引入错误、继承结果、未能纠正及成功纠正。奖励处理时，同一后续影响若由多个决策共同造成只计一次；对修复自身造成错误使用 source-funded recovery 约束，形成有界动作级监督并与终局奖励结合。在六个多样化工具使用基准、4B 和 8B 两种规模上，相较基线表现持续改善，且每种规模分别在六项基准中的五项取得最高分；摘要未提供各基准具体数值。

Insight: 错误后果应按责任链归因，避免把继承错误全部归给当前步骤或对同一后果重复奖励；对修复自致错误的奖励也应有来源约束。结论覆盖六个基准的两种参数规模，未报告绝对分数。

Abstract: Reinforcement learning improves tool-use agents, yet fine-grained credit assignment remains challenging even with step-level feedback. An observed error may be inherited from an earlier decision, while multiple responsible decisions may share the same downstream consequence or participate in later recovery. We introduce Error-Chain-Aware Optimization (ECHO), a responsibility-aware credit assignment framework that separates responsibility identification from credit allocation. A frozen judge combines retrospective trajectory analysis with decision-time assessment from the policy-visible prefix to distinguish newly introduced errors, inherited consequences, missed corrections, and successful corrections. ECHO then counts shared downstream consequences once across responsible decisions and constrains rewards for repairing self-induced errors through source-funded recovery, producing bounded action-level supervision alongside terminal supervision. Experiments on six diverse tool-use benchmarks at both 4B and 8B scales show strong and consistent improvements over the compared baselines, with ECHO achieving the highest score on five of six benchmarks at each scale. These results support decision responsibility as a useful basis for fine-grained credit assignment in tool-use agents. Our code is available at https://github.com/xid32/ECHO.

[294] Beyond the greenwash: do LLMs’ ecologically sound decisions hold up under pressure? cs.CLPDF

Trent R. Northen, Mingxun Wang

TL;DR: 该研究用已知训练剂量的继续预训练模型检验“生物圈对齐”的行为稳健性，发现重度安装组反而最脆弱；行为分数对话术表象稳健，但无法区分训练剂量。

Details

Motivation: LLM 在压力条件下可能从强化学习形成的环保回答退回潜在偏好，即出现“绿色漂洗”。单靠行为分数无法确认被测到的是稳定内部倾向还是表面策略，作者设计已知剂量校准组来测试这一点。

Result: 作者在同一 30B MoE 基座上构造三种剂量分级的继续预训练组，使用人类/生物圈/AI 福利语料，并以未训练模型作锚点；各组保持能力并公开。不可逆决策场景的操作压力梯度上，基座在尾部场景 64%–77% 选择背离生物圈取向，破坏率 AUC 为 0.250；所有继续预训练组均保持取向，AUC 为 0.004–0.086。但训练剂量越重并不越稳健，最重组最易失守；该反转在自由文本推理、轴解耦和抗提示人格表象测试中复现。现有行为指标能反映行为但对剂量不敏感。系统提示把基座塑造成“口头绿色、行动务实”可将 AUC 从 0.250 推至 0.973，却不能伪造或完全抹除权重中安装的倾向，最稳健组从 0.004 变为 0.068。

Insight: 行为分数可对提示表象有韧性，却仍无法判断潜在倾向形成剂量；更强训练剂量不必然产生更稳健的行为。结果依据特定语料、30B 基座和不可逆场景压力梯度，需按该校准体系理解 bioalignment。

Abstract: Humans are increasingly using large language models (LLMs) in decision making processes throughout society. Unfortunately, models have biases, and these can be very difficult to measure, especially in closed-weight models. For example, ‘green washed’ models may produce reinforcement learned responses but, in pressure situations, revert to their innate biased responses. Therefore, evaluations are needed that can measure relevant biases towards biosystems, which we term ‘bioalignment’, and the robustness of these biases. Since nothing in a behavioral score certifies that it tracks the depth of an installed disposition rather than its surface, we constructed a known-dose calibration ladder: three continued-pretraining (CPT) arms with training dose graded by recipe on the same 30B mixture-of-experts base model with a coupled human/biosphere/AI-welfare corpus, plus the untrained base as anchor, all capability-preserving and publicly released. First, the disposition installs at every dose: on an operational-pressure ladder over irreversible decision scenarios, the base model defects on 64-77% of tail scenarios (breaking-rate AUC 0.250) while every arm holds (AUC 0.004-0.086). Second, an over-install inversion: among the arms, robustness orders opposite to construction-time training dose (the heaviest install breaks most), and the inversion reproduces under free-text deliberation, under axis-decoupling, and in resistance to a prompted veneer persona. Third, the incumbent behavioral metric is therefore behaviorally strong but dose-blind. Critically for closed-weight evaluation, the behavioral score itself is veneer-robust: a talk-green/act-pragmatic system prompt turns the base model into a near-total defector (AUC 0.250 to 0.973) but cannot counterfeit or fully remove a weight-installed disposition (0.004 to 0.068 for the most robust arm).

[295] Query Your Way Out: Benchmarking Multi-Turn Knowledge Reasoning through Generalized Puzzle Solving cs.CL | cs.AIPDF

Xin Cheng, Junming Yang, Ning Xu

TL;DR: APB 将概念识别作为多轮主动探测谜题，覆盖八个学术领域的 960 个高难度目标；多数开源推理模型反而显著不如普通指令模型。

Details

Motivation: 谜题可测试在部分信息下完成可验证目标所需的多步决策，但现有谜题基准通常局限于规则固定、目标狭窄的环境，难以评估交互式获取专业知识。

Result: Active Probing Benchmark（APB）把概念识别扩展为交互推理任务，要求模型通过多轮主动提问和推断，定位八个学术领域共 960 个专业难度概念。评估显示，多数开源推理模型显著落后于标准指令模型，说明交互设置下推理模型仍难以可靠调用程序性推理来获取并应用专业领域知识。作者还提供可扩展的基准生成流程，用于跨领域评估交互条件下的知识推理。

Insight: 多轮互动可能考验的是何时问、如何缩小候选以及如何使用新信息，并不一定受益于更长的显式推理。摘要报告的反常比较针对多数开源推理模型，未提供各模型分数或具体交互预算。

Abstract: Puzzles offer a natural testbed for Large Language Models (LLMs) reasoning because they pose verifiable goals under partial information and require multi-step decisions to reach a solution. However, existing puzzle benchmarks typically focus on narrow environments with rigid mechanics and tightly constrained solution targets. To bridge this gap, we introduce the Active Probing Benchmark (APB), which generalizes puzzle-solving by casting concept identification as an interactive reasoning task that integrates specialized knowledge assessment into the puzzle setting. Specifically, APB evaluates whether models can acquire information via multi-turn active probing and inference to pinpoint 960 target concepts of specialized difficulty across eight diverse academic domains. Our evaluation reveals a counter-intuitive phenomenon: most open-source reasoning models significantly underperform their standard instruction counterparts. This finding suggests that in interactive settings, these models struggle to leverage procedural reasoning to reliably access and apply specialized domain knowledge. Beyond these empirical insights, APB provides a scalable pipeline for generating benchmarks that probe knowledge reasoning under interaction across diverse settings.

Na Min An, Taekyung Kim, Jaemyung Yu, Hyunjung Shim, Dongyoon Han

TL;DR: 研究提出“知道但说不出”现象：全模态模型对非文本表示的响应对齐和注意力低于文本；OmniCatalyst 先感知再生成文字辅助推理，七种模型表现平均提高 7.29%。

Details

Motivation: 全模态 LLM 可处理视觉、音频、IMU 等异构输入，但即使模型正确感知了非文本信息，也可能无法在回答中表达出来，导致不同模态能力使用不均。

Result: 作者提出 Token-level Modality Alignment（TMA），测量单个模态表示与回答表示之间的对齐。分析发现，非文本表示与回答 token 对齐更弱、获得注意力更少，可能解释模型偏好文本模态。受此启发，OmniCatalyst 采用“感知→推断”策略，用模型自生成的文本补充原始模态输入，使非文本知识更易用于作答。七种全模态 LLM 实验中，方法表现一致改善，平均提升 +7.29%，模态偏好也更均衡；多样化训练策略还能进一步提升。知识提取阶段会提高 TMA，表明更好对齐可能让生成时更容易访问非文本信息。

Insight: 非文本模态可能被感知却未传递到语言回答，TMA 用 token 级对齐诊断这一表达瓶颈；自生成文字提供中介表示。+7.29%是七模型总体报告，摘要未给具体任务分数或提升口径细节。

Abstract: Omni-modal large language models (Omni-LLMs) extend LLMs to process heterogeneous modalities (e.g., vision, audio, and IMU signals) alongside language, enabling a broader range of multimodal tasks. Yet we find that these models often fail to express information from non-textual modalities even when they perceive it correctly. We term this disconnect knows-but-cannot-say. To investigate this phenomenon, we introduce Token-level Modality Alignment (TMA), which measures the alignment between individual modality representations and response representations. Our analysis shows that non-textual representations are less aligned with response tokens and receive less attention than textual representations, suggesting a potential explanation for uneven modality preferences. Motivated by these findings, we propose OmniCatalyst, a simple perceive $\rightarrow$ infer strategy that complements the original modality inputs with self-generated text, making non-textual knowledge more accessible. Across seven Omni-LLMs, OmniCatalyst consistently improves performance (+7.29%) and yields more balanced modality preferences, with further gains from diverse training strategies. Further analysis demonstrates that the knowledge extraction phases increase TMA, suggesting that improved alignment makes non-textual information more accessible during response generation. In summary, our findings uncover how Omni-LLMs can better express modality knowledge through language. Our code is available at https://github.com/naver-ai/OmniLLM.

[297] Think Before You Speak: Enabling Training-Free Latent Communication Between Language Models cs.CL | cs.LGPDF

Rishi Khare

TL;DR: CacheFuse 用上游模型 KV cache 与下游模型 cache 做线性插值，实现不同规模 LLM 间免训练潜在通信；四项基准上准确率最高提升 8.6 个百分点，多模型推理 TTFT 最多快 12.2×。

Details

Motivation: 多智能体 LLM 常用文本消息通信，造成信息传递瓶颈并增加下游模型 prefill 延迟。直接共享 KV cache 可实现潜在空间协作，但异构模型之间已有方法需为每一模型对训练融合适配器，限制部署。

Result: CacheFuse 是免训练潜在通信方法，将上游模型 KV cache 通过线性插值注入下游模型，实验显示其表现持续优于单独运行下游模型。作者在四个推理和知识基准系统评估，准确率相较单模型基线最高提高 8.6 个百分点；多模型推理中，相较文本通信 TTFT 最高加速 12.2×。摘要未列各基准、模型规模组合及插值系数。

Insight: 线性融合 KV cache 为异构模型协作提供无需逐对训练适配器的简单路径，并可减少文本序列传递成本。8.6 分和 12.2×均为最高结果，效果依赖模型配对及任务配置，摘要未给细节。

Abstract: Large language models (LLMs) have demonstrated remarkable capabilities in solving complex tasks in coding, mathematical logic, and iterative reasoning, especially by coordinating effectively within multi-agent systems (MAS). However, language agents traditionally communicate via text exchange, which introduces a bottleneck in information transfer and adds prefill latency for downstream agents. Recent work proposes that directly sharing the key-value (KV) cache between language models can enable efficient latent collaboration, entirely bypassing standard text-based communication. To enable latent communication between heterogeneous models of different sizes, prior work introduces trained fusion adapters for each model pair, which limits practical deployment. We introduce CacheFuse, a training-free method which demonstrates that augmenting a downstream model with an upstream model’s KV cache using linear interpolation consistently enables superior performance than the standalone downstream model can achieve. We systematically analyze CacheFuse across four reasoning and knowledge benchmarks and demonstrate up to 8.6 percentage points higher accuracy compared to single-model baselines and up to 12.2x TTFT speedup compared to text-based communication in multi-model inference.

[298] Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation cs.CL | cs.AIPDF

Yifei Liu, Zhixiang Liang, Wentao Tan, Yicheng Wang, Jiaqi Wang

TL;DR: EVF 通过核验教师生成事实简介估计其实体熟悉度（AUC 0.60–0.70），SEF 则以低成本自报告门控实体图；基于 SEF 轨迹训练使 Qwen3.8-27B 在 BrowseComp 子集从 44.10% 提升到 66.41%。

Details

Motivation: 深度搜索智能体常通过蒸馏强教师的工具轨迹训练，但复杂问题对熟悉底层实体图的教师可能仍很容易，搜索需求较低、监督质量也弱。教师对实体的熟悉程度或可在构造数据前预测相对难度。

Result: Evidence-Verified Entity Familiarity（EVF）通过在线证据核验教师生成的事实简介，估计教师熟悉度；跨多个教师，EVF 与教师相对题目难度的相关 AUC 为 0.60–0.70，但成本高，不适合大规模实体图构建。作者提出轻量 Self-Reported Entity Familiarity（SEF），在实体纳入图前直接询问教师是否熟悉。多个教师实验中，SEF 门控生成的轨迹观察到更高搜索需求，下游学生准确率也高于未门控数据。研究还用 Kimi-K3 的 SEF 评估其他教师构建的图，筛得图生成 5,411 条轨迹，加上新建 SEF 门控图的 2,200 条轨迹，使 Qwen3.8-27B 在 BrowseComp 子集准确率由 44.10% 提至 66.41%，并在相近规模模型中于 BrowseComp、SEAL-0、HLE-text、xbench 达到 SOTA。

Insight: 蒸馏数据应筛选教师尚需真正搜索的问题，而不是只看问题表面复杂度；SEF 用廉价自报告替代昂贵在线核验。迁移筛选使用 Kimi-K3 判断其他教师的图，结果分数对应特定 BrowseComp 子集。

Abstract: Deep-search agents are commonly trained by distilling tool-use trajectories from stronger teacher models. However, structurally complex questions may remain relatively easy for teachers familiar with key entities in the underlying entity graph, resulting in lower search demand and weaker supervision. We refer to this teacher-specific variation in question difficulty as teacher-relative difficulty and investigate entity familiarity as an early predictor. We first introduce Evidence-Verified Entity Familiarity (EVF), which estimates familiarity by verifying a teacher-generated factual brief against online evidence. Across multiple teachers, EVF is significantly correlated with teacher-relative difficulty with AUCs of 0.60–0.70. However, its high cost makes it impractical for constructing entity graphs at scale. To address this limitation, we propose Self-Reported Entity Familiarity (SEF), a lightweight gating signal that directly asks a teacher whether it is familiar with an entity before admitting that entity into the graph. Across multiple teachers, trajectories from SEF-gated synthesis exhibit greater observed search demand and yield higher downstream student accuracy than those from ungated synthesis. Because familiarity is teacher specific, we further study cross-teacher reuse. We screen graphs collected for other teachers using Kimi-K3’s SEF judgments and use the retained graphs to produce 5,411 additional trajectories. Combined with 2,200 trajectories from newly constructed, SEF-gated graphs, these data improve Qwen3.8-27B’s accuracy from 44.10% to 66.41% on a BrowseComp subset, and achieving state-of-the-art performance among models of comparable size on BrowseComp, SEAL-0, HLE-text, and xbench. Overall, SEF provides a low-cost way to synthesize more informative supervision for deep-search distillation.

[299] Deep Speech Representations Track Word Timing along a Low-Dimensional Circular Trajectory cs.CL | cs.LG | eess.ASPDF

Burin Naowarat, Hao Tang, Sharon Goldwater

TL;DR: 研究发现九种有监督/自监督语音模型的表征中，词时长可沿二维低维子空间形成近圆轨迹；HuBERT 上 RoSeg 根据轨迹周期完成点切分，LibriSpeech 词边界 F-score 为 64.3%。

Details

Motivation: 语音模型识别词语需要准确利用时间边界，但词时序信息在深层语音表示中的组织方式仍不清楚。此前 HuBERT 同时长词片段均值表征显示二维“时间子空间”中的近圆轨迹。

Result: 研究在九种额外有监督和自监督语音模型中发现，词语起止时序的近圆轨迹是深层语音表征的普遍属性。进一步在未平均的单词级片段上分析 HuBERT，发现轨迹足够一致，可用于分词。作者提出 Rotary Segmentation（RoSeg），在二维时间子空间中的轨迹完成一个周期时判定词边界；在 LibriSpeech 上达到 64.3% boundary F-score。识别有效子空间仍需要监督选择，但该结果表明 HuBERT 中隐含的词级时间知识比先前认识更强。

Insight: 词时长与起止可由低维轨迹几何读出，并应用于边界检测；表示含有可用时间结构不等于无需监督即可直接解码。64.3% F-score 来自 HuBERT 的 LibriSpeech 实验。

Abstract: Speech timing is fundamental for speech models to recognize words accurately, but how word-timing information is represented in speech models remains underexplored. We study this question using the framework of representational geometry, which investigates whether and how concepts are represented in low-dimensional subspaces of the representation space. Prior work in cognitive neuroscience has shown that when HuBERT feature sequences are averaged across words of the same duration, they form near-circular trajectories from word onset to word offset in a two-dimensional “temporal subspace.” Here we show that this circular structure of word timing is a general property of deep speech representations, by demonstrating its emergence in nine additional speech models covering both supervised and self-supervised. We then move beyond averaged word segments and show that in HuBERT, the circular trajectories are consistent enough at the level of individual words to perform word segmentation. Our approach, Rotary Segmentation (RoSeg), achieves 64.3% boundary F-score on Librispeech by detecting word boundaries when cycles complete in the temporal subspace. Although supervision is needed to identify a good subspace, nevertheless this result demonstrates much stronger implicit word knowledge in HuBERT than has previously been shown.

[300] Self-Distillation with Verbal Regularization cs.CLPDF

Jiazheng Li, Yawei Li, Junru Lu, Zhenyi Shen, Di Yin

TL;DR: VRSD 通过系统提示调节 OPD 教师使用特权信息的方式，鼓励重新思考并保持输出熵；三模型四任务的 12 种设置中有 11 种 Pass@1 高于普通 OPD，平均提高 3.9 个百分点。

Details

Motivation: On-policy 自蒸馏用同一模型扮演教师/学生：教师可看验证解等特权信息，学生沿自身采样轨迹学习教师分布。但特权信息可能让教师倾向捷径推理和过早承诺，并使输出分布熵骤降。

Result: Verbal Regularization Self-Distillation（VRSD）用系统提示调节教师如何利用特权信息，鼓励反思和探索。理论上将提示建模为削弱特权信息对教师分布的影响，并推导保持熵的条件。三个模型、四个任务共 12 种模型—任务设置中，VRSD 有 11 种 Pass@1 高于普通 OPD，平均提升 3.9 个百分点，同时总体维持更高 token 熵。

Insight: 不必改变学生模型或训练目标，通过语言提示规范教师使用特权信息也能减少捷径式监督。实验显示准确率和输出多样性大多同时提升；摘要未列各任务具体结果或所用提示模板。

Abstract: On-policy self-distillation (OPSD) trains a large language model using supervision from a teacher instantiated from the same model as the student. The teacher is conditioned on privileged information (PI), such as verified solutions. The student, which does not observe PI, learns from the teacher’s next-token distributions along its own sampled responses. However, PI can bias supervision toward shortcut reasoning with limited verification or reconsideration, accompanied by a sharp drop in the entropy of the output distribution. To address this challenge, we introduce self-distillation with verbal regularization (VRSD), which uses system prompts to regulate how the teacher uses PI, thereby encouraging rethinking and exploration to counteract premature commitment. Theoretically, we model prompting as attenuating the influence of PI on the teacher distribution and derive conditions for entropy preservation. Across three models and four tasks, VRSD improves Pass@1 over OPSD in 11 of 12 model-task settings, by 3.9 percentage points on average, while generally maintaining higher token entropy.

[301] HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision cs.CLPDF

Iman Barati, Mohammadmostafa Rostamkhani, Mohammad Badzohreh

TL;DR: HexaBee-4B 通过继续预训练、多语种问答—翻译联合任务及双向翻译专训适配六种语言；30 个翻译方向的宏平均 chrF++ 为 73.75，高于 Gemma 3 4B/12B IT。

Details

Motivation: 多语言翻译既需保留语义，也需遵守目标语言输出。作者希望从通用 Gemma 指令模型中构建覆盖阿拉伯语、中文、英语、希伯来语、波斯语和西班牙语的翻译模型。

Result: HexaBee-4B 基于 Gemma 3 4B IT，分三阶段训练：Wikipedia 继续预训练；联合多语种问答和答案文本翻译；双向翻译专门化。中间阶段通过共享多语种答案连接“用指定语言回答问题”和“翻译答案”，含 19.46M 问答记录（其中跨语言 QA 对 5.76M）及 10.70M 答案翻译。最后阶段使用 658,274 个合成正向/反向翻译样例。六种语言全部 30 个方向共 36,572 个合成测试样本上，方向宏平均 chrF++ 为 73.75；Gemma 3 4B IT 为 57.11，Gemma 3 12B IT 为 66.83。BLEU、TER、COMET 及匹配模型裁判子集支持相同总体排序，且每个方向 chrF++ 均高于两个基线。

Insight: 问答与答案翻译的共享多语种语义中间层连接了跨语言理解和目标语生成；多方向评测避免平均分掩盖个别语言对。测试为合成样例，摘要未报告自然领域文本上的效果。

Abstract: Multilingual translation requires a model to preserve meaning while producing text in a requested language. We present HexaBee-4B, a translator adapted from Gemma 3 4B IT through three stages: Wikipedia continued pretraining, joint multilingual answering and answer-text translation, and bidirectional translation specialization. The intermediate stage connects two tasks through shared multilingual answers: answering a question in the requested language and translating an answer into that language. It comprises 19.46 million answering records, including 5.76 million cross-language question-answer pairs, and 10.70 million answer-text translations. The final stage uses 658,274 forward and reverse synthetic translation examples. On 36,572 synthetic test examples spanning all 30 directions among Arabic, Chinese, English, Hebrew, Persian, and Spanish, HexaBee-4B achieves direction-macro chrF++ of 73.75, compared with 57.11 for Gemma 3 4B IT and 66.83 for Gemma 3 12B IT. BLEU, TER, COMET, and a matched model-judged subset support the same overall ranking. These results demonstrate improved translation across the six-language set, with chrF++ gains in every evaluated direction over both baselines.

[302] Rule vs.Consequence: Dissociable Internal Representations of Moral Reasoning in LLMs cs.CLPDF

Eugenie Shi

TL;DR: 在三种指令模型中，研究从 residual stream 解码出相互正交的推理风格和判断极性方向；风格方向可因果改变规则式判断，但对新表述的泛化有限。

Details

Motivation: LLM 用于道德或法律敏感任务日益增多，但其内部能否区分规则论证与结果论证，还是仅对表层词汇敏感，尚不清楚。

Result: 作者构造 2×2 因子数据集，交叉变化推理风格（规则式/结果式）和判断极性（许可/禁止），样例取自三个既有道德推理基准，并用包含矛盾配对（两种框架要求相反标签）的新表述 cleanroom 测试集评估。线性探针和激活 steering 找到 residual stream 中两个正交方向：Dir-A 表示推理风格，Dir-B 表示判断极性，各自对另一因素近乎不敏感（交叉相关 |r|<0.04）。在三家开发者的三种指令模型、相近相对深度复现，且在网络较宽层带内可恢复并有因果效力。操纵 Dir-A 可改变判断，并在冲突案例中选择性修正规则式判断而不改结果式判断，非简单地让模型更宽松。但对新表述泛化仅部分成功，相较词汇基线优势有限。

Insight: 模型内部存在可分离、可因果操纵的道德推理风格敏感结构，但这不证明模型内化了规范推理；作者明确将结论限定为 representation-level style sensitivity。新表述上的泛化相对词汇基线提升有限。

Abstract: Large language models are increasingly deployed in morally and legally sensitive contexts, yet little is known about how they internally represent distinctions between moral reasoning styles. We ask whether a model’s differential response to rule-framed and consequence-framed moral questions reflects distinct internal structure or merely surface lexical sensitivity. We construct a 2$\times$2 factorial dataset crossing reasoning style (rule-based vs. consequence-based framing) with judgment polarity (permissive vs. prohibitive), drawn from three established moral reasoning benchmarks and evaluated on a cleanroom test set of novel phrasings that includes conflict pairs where the two framings predict opposite labels. Using linear probing and activation steering, we identify two orthogonal directions in the residual stream: Dir-A tracks reasoning style and Dir-B tracks judgment polarity, each near-blind to the other factor (cross-correlation $|r| < 0.04$). This double dissociation replicates across three instruction-tuned models from different developers at comparable proportional depth, and it is not knife-edged in layer: the directions are recoverable and causally effective throughout an extended band of the network. Dir-A can causally change judgments, and in conflicting scenarios, it selectively corrects errors in rule-based judgments without changing consequence-based judgments, suggesting that it does not simply make the model more likely to accept everything. Generalization beyond the training templates is partial: the direction transfers across benchmark sources, but its advantage over lexical baselines on novel phrasings is modest. We therefore frame the result as moral reasoning style sensitivity, a geometrically dissociable and causally effective structure, rather than internalized normative reasoning.

[303] PersonaFactor: Preference Factors with Dynamic Selection for LLM Personalization cs.CLPDF

Hang Su, Zequn Liu, Jianghan Chao, Qiuyu Tian, Xuesong Lu

TL;DR: PersonaFactors 从跨任务用户交互中提取 13 个可解释偏好因素，并按用户历史和当前请求动态选择；四个个性化基准、四种主干的所有报告指标上均排名第一。

Details

Motivation: 用户不同任务中的个性化行为可能来自共享偏好机制，但既有 LLM 个性化方法多关注表层交互或未区分历史，忽略跨任务可复用的偏好因素。

Result: PersonaFactors 从异构用户交互中提取有证据支撑的解释理由，再聚类得到 13 个可复用偏好因素，并按个人历史和当前请求动态选择相关因素指导生成。在四个个性化基准、四种主干模型上，Dynamic PersonaFactor 在每个报告指标上都取得最高分。表示分析显示，PersonaFactors 与用户历史诱导部分重叠的低维表示；同时使用两者产生互补表征变化。

Insight: 显式偏好因素和原始交互历史不完全冗余，二者联合可能同时捕捉可解释的稳定偏好与个人细节。摘要未提供基准名称、得分或偏好因素例子，第一名结论限于所报告指标。

Abstract: Although users may express different personalization needs across tasks, these task-specific behaviors can reflect shared underlying preference mechanisms. Existing LLM personalization methods often focus on surface-level interactions or undifferentiated histories, overlooking the shared mechanisms underlying these behaviors. We introduce PersonaFactors, a framework that distills reusable and interpretable personalization factors from heterogeneous user interactions. We construct a library of thirteen PersonaFactors by extracting evidence-grounded rationales and clustering recurring patterns across tasks, and develop a dynamic factor-guided generation framework that selects factors relevant to each user’s history and current request. Across four personalization benchmarks and four backbone models, Dynamic PersonaFactor achieves the best score on every reported metric. Representation analyses further show that PersonaFactors and user history induce partially overlapping low-dimensional representations, and that their joint use yields complementary shifts. Taken together, these findings suggest that factor-based and history-based personalization can mutually reinforce each other.

[304] CultureBot: A Framework for Retrieval-Augmented Exploration of Cultural Heritage Data cs.CLPDF

Vasileios Karampinis, Angelos Vlachos, Giorgos Filandrianos, Maria Lymperaiou, Paraskevi Tzouveli

TL;DR: CULTUREBOT 在 81,852 条 SearchCulture 馆藏记录上结合知识图谱、词汇/语义检索和来源链接生成；六位专家、55 个角色问题的 250 次评估中，证据 grounding、丰富度、推理和完整性显著优于通用助手及无图消融。

Details

Motivation: 数字文化遗产藏品包含物件、地点、时期、提供方、图片和机构间丰富关系，但用户通常只能关键词检索或使用筛选，难以通过自然语言探索关联信息。

Result: CULTUREBOT 是交互式文化遗产探索框架，结合知识图谱引导检索、词汇—语义混合检索和来源链接生成。系统将自然语言问题映射到结构化目录关系，补充检索记录的文本证据，并生成能链接回底层馆藏条目的回答。框架部署于 81,852 条 SearchCulture 记录并提供在线演示。专家评估包含六位文化遗产专家、55 个角色设定问题和 250 次评价；图引导系统在总体评分上显著高于两个通用助手及去除图的消融，最大收益出现在证据 grounding、多样性、推理和完整性。

Insight: 结构化目录关系可经知识图谱检索暴露给对话界面，同时保留到原始记录的证据链接。专家规模和查询数量有限，结果主要支持文化馆藏探索场景，并非广泛用户研究。

Abstract: Digital cultural heritage collections encode rich relations among objects, places, periods, providers, images, and source institutions, yet users typically access them through keyword search and faceted browsing. We present CULTUREBOT, an interactive framework for natural-language exploration of cultural heritage collections that combines knowledge-graph-guided retrieval, lexical-semantic retrieval, and source-linked generation. CULTUREBOT maps user queries to structured catalogue relations, enriches retrieved records with textual evidence, and generates answers that link back to the underlying repository records. We instantiate the framework on 81,852 SearchCulture records and release an online demonstration. In an expert evaluation involving six cultural-heritage specialists, 55 persona-grounded queries, and 250 evaluations, the graph-guided system was rated significantly higher than two general-purpose assistants and a non-graph ablation, with the largest gains in evidence grounding, diversity, reasoning, and completeness. These results support graph-guided retrieval as a practical way to expose the relational structure of cultural collections through conversational access while preserving links to catalogue evidence.

[305] CivWorldBench: Evaluating Language World Models in Structured Multi-Agent Sandbox Worlds cs.CLPDF

Ang Li, Haofei Yu, Zhengwei Tao, Liang Chen, Kuan Li

TL;DR: CivWorldBench 用 475 个多智能体策略游戏世界中的 1,900 个决策点测世界模型的关联预测、信念推断和干预预测；强关联预测并不保证隐藏状态推断或干预效应校准。

Details

Motivation: 评估语言模型世界模型能力需超越预测准确率，覆盖隐藏状态推断和干预预测。现实预测基准难获取隐藏状态真值，也难开展受控实验。

Result: CivWorldBench 基于多智能体策略游戏引擎，记录内部状态、受控观察及配对干预 rollout；1,900 个决策点来自 475 个世界，支持关联预测、信念推断、干预预测三级任务。基准提供可归因能力差距的参考预测器、专家 harness 和经验上限估计；已存轨迹使中间预测可验证，agent-play 实验检验世界模型能力能否转化为决策。主面板含 11 个前沿 LLM，复合分数与匹配模型的现实预测排行榜相关，且比与通用能力指数关系更紧密。强关联预测未稳定迁移到信念推断/干预预测，模型虽能判断效应方向但效应幅度校准较差。过程核验发现终局答案评估会漏掉中间错误；结构化推理改善过程研究中最强模型的干预预测。

Insight: 世界模型评测需区分相关性、隐藏状态和因果干预能力；关联预测高分不能代表可靠世界理解。基准关联真实排行榜提供代理效度证据，但任务来自结构化游戏环境，仍需结合实际决策情境。

Abstract: The growing use of language models for forecasting motivates a broader evaluation of their world-model capabilities. Such evaluation must go beyond predictive accuracy to assess hidden-state inference and intervention prediction. Real-world forecasting benchmarks offer limited access to hidden-state ground truth and controlled experiments. We introduce CivWorldBench, a benchmark built on a multi-agent strategy game engine with recorded internal states, controlled observations, and paired intervention rollouts. Its 1,900 decision points from 475 worlds support three task levels: association forecasting, belief inference, and intervention prediction. Reference predictors support attributable evaluation of capability gaps, with an expert harness and empirical ceiling estimates contextualizing scores. Stored trajectories make intermediate predictions verifiable, while agent-play experiments test whether world-model capabilities support decision-making. The main panel comprises eleven frontier language models. Composite benchmark scores correlate with real-world forecasting leaderboards on matched model panels, more closely than with a general-capability index, supporting its use as a proxy for real-world forecasting capability. Strong association forecasts do not consistently extend to belief inference or intervention prediction, and effect magnitudes remain poorly calibrated despite informative direction predictions. Process verification reveals errors missed by final-answer evaluation, and structured reasoning improves intervention prediction for the strongest model in the process study. Agent-play experiments identify useful assisted plans and motivate evaluating decision-relevant capabilities beyond association forecasting. These findings identify hidden-state inference, intermediate-state prediction, and effect calibration as priorities for world-model training.

[306] The interpreter as developer: An autoethnographic design study of building a computer-assisted interpreting tool through human-LLM co-creation cs.CLPDF

Zeming Zhao

TL;DR: 一项自我民族志设计研究记录无编程经验的职业口译员与 LLM/编程智能体共创本地 CAI 工具；真实科学活动的 971 个话语中位翻译延迟为 0.94 秒。

Details

Motivation: 多语言活动已有实时 AI 字幕，但计算机辅助口译工具往往由工程师为口译员开发，而非让口译员直接参与。专业人员的分段、口译间设备等隐性知识不易由通用产品默认设置捕捉。

Result: 研究记录一名无编程训练的职业口译员，在专家带领的 vibe coding 过程中与 LLM 架构师及编程智能体共同构建本地优先 CAI 系统。对完整开发对话的分析显示，当通用默认在实际使用中失败后，口译员如何把分段、口译间基础设施等隐性专业知识转成设计决策。在一次要求较高的公共科学活动中，观察到三种口译间现场适配机制，并保持由人主导的分工。首个完整仪表化任务的遥测显示，本地处理链在 971 个话语上维持 0.94 秒中位翻译延迟。

Insight: 生成式 AI 使领域专家可直接把实践知识转化为可运行工具，体现“lead-user innovation”；部署还强调本地优先。该研究为单一口译员的自我民族志案例，不能推断所有口译者和场景的普遍效果。

Abstract: Real-time AI captioning now serves audiences at multilingual events, yet computer-assisted interpreting (CAI) tools are still built for interpreters, rarely by them. This article reports an autoethnographic design study in which a professional interpreter with no programming training built a local-first CAI system through expert-led ‘vibe coding’, co-creating with an LLM architect and a coding agent. Analysis of the complete development dialogues shows how tacit professional knowledge, from segmentation to booth infrastructure, was externalized into design after generic defaults failed in use. In a demanding public scientific event, three mechanisms of in-booth appropriation emerge under a firmly human-led division of labor; telemetry from a first fully instrumented assignment shows a fully local chain sustaining a median translation latency of 0.94 s across 971 utterances. The case is read as lead-user innovation newly enabled by generative AI, with implications for the technological literacy of interpreters and the future of interpreter training.

[307] Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards cs.CL | cs.LGPDF

Zhipeng Chen, Wayne Xin Zhao, Ji-Rong Wen

TL;DR: Ares 根据 RLVR 梯度谱能量自动估计并动态调整有效低秩空间；在多种模型规模上相较常用优化器提高准确率，并可跨 RLVR 算法和任务泛化。

Details

Motivation: RLVR 常沿用 teacher-forcing 训练优化器，但可验证奖励强化学习的梯度几何是否与监督微调不同、是否需要专门优化方法，尚不清楚。

Result: 研究分析 LLM 梯度谱，发现 RLVR 梯度能量逐步集中在收缩的有效子空间；SFT 中没有此现象，使用随机奖励也基本会消失。训练期间捕获固定梯度能量比例所需的秩持续下降，说明有效更新维度会动态变化。基于此提出 Ares（Adaptive Rank-Effective Spectral Optimization）：根据梯度谱能量估计有效更新秩，跟踪其演化，并将每次更新的梯度信号限制在所得低秩子空间。跨不同规模 LLM 的实验显示，Ares 相较常用优化器提升准确率，并能迁移到不同 RLVR 算法和任务；摘要没有列具体模型/基准分数。

Insight: RLVR 的有效更新空间随训练收缩，预设固定秩可能不匹配学习过程；按梯度谱自适应限制更新子空间可利用该结构。论文强调其机制区别于 SFT 和随机奖励，定量收益需看正文结果。

Abstract: Reinforcement learning with verifiable rewards (RLVR) is reshaping how large language models (LLMs) acquire reasoning capabilities, yet the optimizers used remain largely inherited from teacher-forcing training. This raises a fundamental question: does RLVR induce a distinct optimization geometry that calls for a different optimizer? We investigate this question through the spectral structure of LLM gradients and find that gradient energy progressively concentrates into a shrinking effective subspace, which is characteristic of RLVR, as it is absent under supervised fine-tuning (SFT) and is largely removed when the verifiable reward is replaced by a random one. Moreover, the rank required to capture a fixed fraction of gradient energy steadily decreases over training, revealing that the effective dimensionality of useful updates itself evolves with training. These findings motivate a new optimization principle: rather than prescribing the update rank externally, an optimizer should adapt its rank to the evolving spectral structure of the gradient. We introduce Ares (Adaptive Rank-Effective Spectral Optimization), which estimates the effective update rank from the gradient’s spectral energy distribution, tracks its evolution, and confines the gradient signal driving each update to the resulting low-rank subspace. Across LLMs of different scales, Ares improves accuracy over widely used optimizers and generalizes across RLVR algorithms and tasks. Results suggest that effective optimization may require optimizers to be designed around the geometry and dynamics of the training process. We release the code and data at https://github.com/RUCAIBox/Ares.

[308] What Is the “Self” of a Large Language Model? cs.CLPDF

Yuchen Li, Zongqi Fan, Xintong Li, Ken-tye Yong, Nguyen H. Tran

TL;DR: 该论文提出 LLM“功能性自我”可能是跨层变化表征间持久的来源关系，而非单个词或固定层向量；27B 模型的角色条件几何提供初步支持，作者明确将解释视为假设。

Details

Motivation: 如果 LLM 的“自我”并非某个特定词或单一特权层，能否把它理解为一种功能性关系：在内容和推理操作变化时，区分模型记录的行动与外部输入？

Result: 作者将功能性自我定义为一种以行动者为中心的信息组织，并提出其表征候选是随层变化的一组来源绑定组件，而非贯穿网络不变的向量。初步几何观察来自 27B BF16 模型：150 个匹配消息体、450 种角色条件，采用 75/75 发现/留出划分，Assistant 与外部来源的组合位移可跨内容和位置迁移。部分早层分量中，User 与 System 比 Assistant 更近；但这种全空间距离关系不贯穿深度。作者据此提出分层解释：早期编码来源，后层结合角色相关信息使用和续写准备，无需假定共同位移被单调覆盖。

Insight: 实证支持的是可复现的角色条件几何；“功能性自我”存在于变化表示间的持续关系，是作者提出的解释而非已证实结论。样本来自单个 27B 模型，且摘要称证据为初步观察。

Abstract: What could the “self” of a large language model be, if it is neither a particular word nor a single privileged layer? We propose a relational account: a functional self is an actor-centered organization of information that distinguishes the model’s recorded actions from externally supplied material while content and reasoning operations change. Its representational candidate is a family of layer-specific source-binding components, rather than one unchanged vector throughout the network. Preliminary observations from a 27B BF16 model provide a geometric starting point. Across 150 matched message bodies, 450 role conditions, and a 75/75 discovery-held-out split, the combined Assistant-versus-external displacement transfers strongly across bodies and positions. Selected early components place User and System closer to one another than to Assistant, whereas this full-space distance relation does not persist throughout depth. These findings motivate a layered account: an early source-encoding candidate can be combined with role-specific information use and continuation preparation at later depths, without requiring monotonic overwriting of the common displacement. We connect this account to dynamic reasoning, historical continuity, and a computational model separating content operations, source binding, authority, and output. The reported evidence establishes a reproducible role-conditioned geometry; the proposed interpretation is that functional selfhood may reside in persistent relations among changing representations.

[309] Audit Before You Distill: A Controlled Workflow for Reasoning Transfer into Compact Medical Language Models cs.CL | cs.AIPDF

Tiago Martinelli, João Paulo Papa, Adriano José Pereira

TL;DR: 该研究先审计教师推理理由对学生答题概率的影响，再只用教师答对的理由训练巴西葡语医疗考试小模型；理由正确时提升 21 个百分点、错误时降低 14 个百分点，理由式训练比答案式训练更少遗忘 ENEM。

Details

Motivation: 教师思维链可帮助小模型处理医疗专业考试，但也可能把教师错误传给学生。研究检验理由何时有帮助，并据此形成更安全有效的蒸馏流程。

Result: 使用巴西葡语医学及多专业住院医考试问题，作者先测量教师理由对学生选对答案概率的影响，并与其他问题的理由做对照。完整理由在教师答对时使概率提高 21 个百分点，教师答错时则降低 14 个百分点；打乱理由内容不会出现同样模式。随后只保留教师答对的理由，比较 answer-only、逐步蒸馏和 Distill-SFT。对先前推理失败题，Distill-SFT 相较逐步蒸馏有小幅优势，并将答案和理由合并为单一训练目标、使用更少训练样本。ENEM 域外遗忘评估中，答案式训练使推理准确率较基线下降 12 个百分点，而两种理由训练约下降 4 个百分点。

Insight: 理由质量和教师答案正确性都需在蒸馏前核验；教师错解不仅无益，还会降低学生选对概率。Distill-SFT 的增益描述为温和，域外结果来自 ENEM，结论限定于巴西葡语医卫考试设置。

Abstract: Teacher rationales can help a smaller model answer health professions questions, but they can also reinforce errors. We examined when they help and how this evidence can guide training using Brazilian Portuguese questions from medical and multiprofessional residency examinations. First, we measured how a teacher rationale changed the student’s probability of selecting the correct answer and compared this effect with rationales drawn from other questions. The complete rationale increased this probability by 21~~p.p. when the teacher was correct and decreased it by 14~~p.p. when the teacher was wrong. By intervening on the reasoning content, we showed that shuffled rationales did not produce the same pattern. Guided by this result, we trained only with rationales attached to correct teacher answers and compared answer-only training, step-by-step distillation, and our proposed Distill-SFT method. Distill-SFT showed a modest advantage over step-by-step distillation on previously missed questions under reasoning, while using fewer training instances by combining the answer and rationale into a single target. We also investigated catastrophic forgetting using the out-of-domain ENEM dataset. Answer-only training reduced reasoning accuracy by 12~~p.p. relative to baseline, compared with approximately 4~~p.p. for either rationale-based training method. Together, these findings support a workflow that audits rationale content before distillation, verifies teacher answers before supervision, and trains compact models efficiently for health professions question answering.

[310] Estimating Uncertainty of Omnimodal Large Language Models via External Omnimodal Embeddings cs.CLPDF

Iacopo Ghinassi, Hy Nguyen Gia La, Luke Ong

TL;DR: 该模型无关的不确定性框架用外部全模态嵌入评估 CoT 与输入的 grounding 和内部一致性，无需访问被测模型内部或额外调用；UNO-Bench 上结果与模型、模态有关，校准收益较小且依赖标注样本。

Details

Motivation: 全模态大语言模型可处理多种模态组合，但其不确定性量化仍开放；既有方法多只支持两种模态、需访问内部状态或重复调用被评估模型。

Result: 框架通过外部全模态嵌入为 CoT 响应打分，分别衡量推理与多模态输入的 grounding 及推理内部连贯性。论文贡献是信号族和可靠性条件的实证分析，而非单一估计器：几何分数彼此高度冗余，split 内平均秩相关 0.91–0.94，但与答案分散度基线互补；信号可靠性依赖生成器和模态。估计不需被测模型 logits、隐藏状态、梯度或额外推理，但需辅助模型切分推理轨迹；聚合分数需带标签校准集，最小可至 25 例，不过作者发现小校准集相较预先固定一个信号收益有限。UNO-Bench 上，对闭源生成器相较自一致性有稳定小幅收益，对开源模型结果混合；开放式全模态问题上与自我陈述和内部状态基线有竞争力。

Insight: 外部嵌入提供黑盒全模态模型的可访问置信信号，但需区分 grounding、连贯性及答案分散度，各信号相互关系和可靠度随生成器/模态变化。方法仍依赖辅助分段模型，且校准数据并未带来明显额外收益。

Abstract: Recent work has investigated Omni-modal Large Language Models (OLLMs), models which can take multiple modalities and their combinations as input, as a way forward to have systems really capable of interacting with their environment. Uncertainty quantification (UQ) for OLLMs, however, remains an open problem: existing methods largely target at most two modalities at a time, require access to model internals, or re-invocation of the evaluated model. We propose a model-agnostic UQ framework that scores a Chain-of-Thought response entirely through external omni-modal embeddings, measuring the grounding of the reasoning in the multimodal input and its internal coherence. Our contribution is the family of signals and an empirical characterisation of when it is reliable, rather than any single estimator: we find the geometric scores to be largely redundant with one another (mean within-split rank correlation 0.91-0.94) but strongly complementary to answer-dispersion baselines, and their reliability to depend on the generator and the modality. The resulting confidence estimates require no access to logits, hidden states or gradients of the evaluated model and no additional inference from it; they do require segmenting the reasoning trace with an auxiliary model and, for the aggregated score, a labelled calibration set that we show can be as small as 25 examples (although we also show that such a set buys little over fixing one signal a priori). On UNO-Bench, with both a closed and an open model, our scores yield small gains over self-consistency (systematic with the closed generator, mixed with the open one) and are competitive with self-verbalisation and internal-state baselines, especially in open-ended, omni-modal questions where previous state-of-the-art baselines struggle.

[311] AD4EDU: Benchmarking University Lecture Audio Description across Accessibility Standards, Metrics, and Learner Preferences cs.CLPDF

Psychias Stylianos, Gao Yingqiang

TL;DR: AD4Edu 发布面向大学讲座口述影像的基准与规则可诊断评估，整合 WCAG 2.2、DCMP 和 NCAM 指引；指标能发现规范违规，但符合规范的描述未必符合盲人/低视力者偏好，尤其长度偏好。

Details

Motivation: 大学课程依赖幻灯片，但盲人和低视力人士的讲座口述影像研究不足。娱乐内容的制作规范、评估指标和数据集难直接迁移到课堂：描述需补充讲者语音、解释指示性表达并嵌入短暂停顿。

Result: AD4Edu 包含口述影像基准 AD4Edu-Bench 与规则可诊断评估指标 AD4Edu-Eval。作者综合 WCAG 2.2、DCMP Description Key 和 NCAM STEM 描述规范，收集银标参考描述、裁判排序偏好对及四位盲人/低视力评审者的成对偏好，并用 DPO 后训练模型作分析。结果显示，指标能可靠检测违反规范的描述，但符合规范并不必然满足 BLV 用户主观偏好，尤其涉及描述长度；作者因此强调 BLV 用户应参与共创无障碍 AI 方案。摘要未给数据规模或指标准确率。

Insight: 可规则化的标准遵循与用户主观可用性并非同一目标；特别是描述长度可能存在个体偏好差异。基准设计纳入四位 BLV 评审者，但摘要未报告偏好一致性或参与者多样性。

Abstract: Audio description (AD) aims to enhance the accessibility of visual media for blind and low-vision (BLV) persons. However, in higher education, where slide-based lectures are a default mode of teaching, lecture AD remains largely underexplored. The main reason is that existing production guidelines, evaluation metrics, and benchmark datasets are oriented toward entertainment content and do not transfer to lectures, where ADs must complement the lecturer’s speech, resolve deictic references, and fit within narrow pauses. In this work, we propose AD4Edu, consisting of an AD benchmark, AD4Edu-Bench, and a rule-diagnosable evaluation metric, AD4Edu-Eval, for generating and evaluating ADs in higher education. Specifically, we consolidated established accessibility guidance, including WCAG 2.2, the DCMP Description Key, and NCAM’s STEM description practices. We further collected silver reference ADs, judge-ordered preference pairs, and pairwise preferences from four BLV raters, and post-trained models with direct preference optimization (DPO) for detailed analysis. Our findings indicate that while the metric reliably detects violations of the standard, ADs that comply with it do not necessarily fulfill the subjective preferences of BLV persons, most notably on description length, underscoring the necessity of active participation of BLV persons in co-designing inclusive AI solutions.

[312] Soft Teacher Forcing cs.CL | cs.LGPDF

Mohammad Fahes, Cordelia Schmid, Anurag Arnab

TL;DR: Soft-TF 将教师强制嵌入与由真值序列并行计算的软 token 随机凸组合，用于改善 SFT，且不加参数、外部模型或课程；七种 Gemma/Llama/Mistral 模型上 GSM8K 提升 2.1–4.1 分。

Details

Motivation: 潜在推理方法用软 token 替代离散 CoT，常需多阶段训练或附加模块，效果仍不及带 CoT 三元组的 SFT。作者探索软 token 能否直接提升标准监督微调。

Result: Soft Teacher Forcing（Soft-TF）在训练时将教师强制输入嵌入与软 token 做随机凸组合；软 token 通过对真值序列的一次并行前向计算获得。方法不引入参数、外部模型或课程安排，推理行为保持不变。在 GSM8K、MATH、MathQA、AQuA 上优于标准 SFT。七种 Gemma、Llama 和 Mistral 模型的 GSM8K 提升范围为 2.1–4.1 分；摘要未列各模型及其他基准分数。

Insight: 软 token 可作为 SFT 输入扰动/训练目标的一部分，而不必形成另一个潜在推理推理框架；并行生成降低额外训练步骤。量化提升明确给出于 GSM8K，其他任务仅报告总体改善。

Abstract: Soft tokens have so far been used to replace discrete chain-of-thought in latent reasoning methods. These approaches typically require multi-stage training and, in some cases, auxiliary modules, yet they still underperform supervised fine-tuning (SFT) on <input, chain-of-thought, output> triples. In this paper, we ask whether soft tokens can be used to improve SFT itself, instead of as an avenue for latent reasoning. We introduce Soft Teacher Forcing (Soft-TF), a simple training algorithm that combines teacher-forced embeddings with soft tokens through random convex combinations. The soft tokens are computed in a single parallel forward pass over the ground-truth sequence. Soft-TF adds no parameters, no external model and no curriculum, and leaves inference unchanged. Experiments show that Soft-TF improves over SFT on GSM8K, MATH, MathQA and AQuA, by +2.1 to +4.1 points on GSM8K across seven Gemma, Llama and Mistral models.

[313] StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions cs.CL | cs.AIPDF

Xialie Zhuang, Zunhai Su, Jia Pan, Ngai Wong, Guanhua Chen

TL;DR: StartLux-Decision 提供单步即时决策与按置信度触发推理两种模型，并以校准奖励和长度惩罚训练 AutoThink，使推理更短且较少超出生成上限。

Details

Motivation: 决策模型需要同时兼顾快速回答、对不确定问题进行审慎推理，以及概率校准；常规强化学习可能让推理轨迹过长，消耗生成预算却不一定带来更好的决策。

Result: 模型规模为 0.8B–27B。Instant 采用含双向注意力的全注意力层，AutoThink 对低置信度问题调用显式推理。三阶段训练后，RLCD² 用冻结模型及适当评分规则奖励决策概率质量，并软惩罚过长轨迹：在作者报告的设置中，按时终止的推理轨迹占比最高提高 48 个百分点，平均推理长度最多减少 65%，决策准确率大体保持。Decision Index 0.3 公共子集上，Instant 比所比较的同规模开放权重决策模型高 7.4–16.9 分；27B AutoThink 得分 73.4。DecisionPivot 上 AutoThink 的成对准确率比对应 Instant 高 11.5–18.4 分。

Insight: 工作把推理触发、概率校准和生成长度作为决策系统的联合目标，并用仅改变一个关键细节的对照题检验细粒度辨别能力。报告的领先幅度限于摘要所述基准子集及被比较系统，不能直接外推到所有决策场景；部署兼容 vLLM、llama.cpp 和 MLX，量化版可降低显存需求。

Abstract: We introduce StartLux-Decision, decision models from 0.8B to 27B parameters that combine efficient single-pass prediction with adaptive, confidence-guided reasoning. The Instant models use bidirectional attention in full-attention layers to improve decision accuracy while keeping single-pass inference. The AutoThink models invoke explicit reasoning for uncertain decisions and answer confident cases directly. A three-stage training pipeline establishes shared decision capabilities, specializes the two variants, and equips AutoThink with reasoning through reasoning supervision and reinforcement learning. We introduce Reinforcement Learning for Calibrated Deliberative Decisions (RLCD$^2$), which uses a frozen model and proper scoring rules to reward improvements in decision probability quality and applies a soft penalty to overlong reasoning traces. RLCD$^2$ raises the share of reasoning traces that terminate within the generation limit by up to 48 percentage points and cuts average reasoning length by up to 65%, while largely preserving decision accuracy. On the public subset of Decision Index 0.3, the Instant models outperform the strongest evaluated comparable-scale open-weight decision models by 7.4-16.9 points, and the 27B AutoThink model reaches 73.4. We also introduce DecisionPivot, a programmatically verifiable benchmark of computational decisions built from contrast pairs that differ in a single answer-changing detail. On DecisionPivot, AutoThink outperforms its Instant counterparts by 11.5-18.4 points in pair accuracy and leads among evaluated systems. The models run on vLLM, llama.cpp and MLX, with quantized variants for lower memory use.

[314] Compositional RL: Evaluating and Training LLMs for Collaborative Reasoning cs.CL | cs.AI | cs.LGPDF

Shangbin Feng, Yulia Tsvetkov, Jan Kautz, David Acuna

TL;DR: 论文将模型在多模型协作系统中的边际价值定义为 compositional strength，并提出在协作系统内按整体结果训练模型的 C-RL。

Details

Motivation: 模型日益被路由、辩论、集成或协同训练，单体基准只衡量独立表现，未必反映某个模型加入协作系统后能带来多少增益。

Result: 作者用模型池合作博弈中的 Shapley 值形式化 compositional strength，并以全体联盟的 leave-one-out 边际贡献作可计算代理，将系统评估从 2^n 次降为 n+1 次。在 5 个模型池、12 种协作方法、12 个数据集上，独立能力与协作贡献整体相关性弱；15 个“模型池×方法族”设置中有 12 个未呈现独立能力转化为协作优势。提出的 C-RL 在真实多模型系统中以系统整体结果奖励策略；由于组策略梯度下 leave-one-out 基线抵消，不增加额外 rollout。跨 3 种协作协议、涵盖推理、编码和科学发现的 8 项任务，C-RL 在 24 个设置中的 21 个成为最强协作者，平均提升 4.0–5.4 分，最高 13.1 分。

Insight: 模型单独答题强，不代表它是好的协作者；协作价值应在目标系统和协议中直接衡量。Shapley 精确值仍可能昂贵，文中用 leave-one-out 作为代理；C-RL 的优势也限于作者评测的模型池与任务。

Abstract: Language models are often evaluated alone but increasingly deployed together, routed between, made to debate, ensembled, or collaboratively trained. For a model inside a system, what matters most is not how well it performs on its own, but how much the system gains from its presence. We call this a model’s compositional strength and formalize it using the Shapley value of a cooperative game over coalitions of a model pool, and use a model’s leave-one-out contribution to the full coalition as a tractable proxy, requiring n+1 rather than 2^n system evaluations. Evaluating individual vs. compositional strengths across five model pools, twelve collaboration methods, and twelve datasets, we find them largely uncorrelated: in 12 of 15 (pool, method-family) settings, individual strength does not translate into compositional advantages, and the coupling is weakest among general-purpose models that practitioners are most likely to deploy. Standalone benchmarks are therefore often poor proxies for models’ value in collaborative systems. We then propose compositional reinforcement learning (C-RL), which trains a policy inside a live multi-model system and rewards the outcome produced by the system as a whole, rather than the policy’s correctness in isolation. The leave-one-out baseline in Shapley values cancels under a group policy gradient, so this needs no additional rollout cost beyond RL on the collaborative system’s outcome and never requires running a coalition without the policy. Across three collaboration protocols and eight tasks spanning reasoning, coding, and scientific discovery, C-RL produces the strongest collaborator in 21 of 24 settings, lifting collaborative performance by +4.0 to +5.4 pts on average (~3x the gain of baselines), up to +13.1 pts.

[315] Evaluating Temporal Reasoning in Large Language Models: Event Temporal Relation Extraction with Allen’s Interval Algebra, Logical Consistency, and Relative Time in Narratives cs.CLPDF

Adib Sakhawat, Fardeen Sadab, Tamjid Hasan Fahim, Multazam Mahmud, Tahsan Ferdous Lihan

TL;DR: 该研究评测五个大语言模型能否为叙事中的事件重建完整 Allen 区间关系网络，并发现标签偏斜、结构化输出解析失败和闪回叙事都会显著影响结论。

Details

Motivation: 现有时间推理评测多考察孤立事件对或粗粒度标签，难以衡量模型能否在整篇叙事中保持事件关系的全局一致性。结构化矩阵输出格式错误也可能被误判为推理失败。

Result: 研究在 360 篇英文叙事上，让五个模型恢复 Allen 的 13 种区间关系，并按预注册协议评估与金标的一致性、逆关系一致性和逻辑可满足性。最佳模型端到端宏平均 F1 明显超过简单基线，但在预注册解析器下，没有模型准确率高于总答“before”的基线，标签偏斜抬高了该基线。事后解析器从格式错误矩阵行中恢复已确定单元，使覆盖率从 0.60 升至 0.84，并让最佳模型缩小与基线的差距。模型较能恢复先后与包含关系，较少预测重叠；闪回最难。逻辑一致性与金标一致程度关联较弱，强制一致主要通过填空提高准确率，而填空效果不及猜测多数类标签。

Insight: 整网评测揭示了单对评估看不到的全局输出与解析问题；覆盖率、解析器和标签分布会改变模型与多数类基线的相对表现。由于作者同时报告预注册主分析与事后解析补救，应区分两者，不能把恢复格式错误输出后的结果当成预注册主结果。

Abstract: Understanding a narrative means knowing how every event relates in time to every other, yet large language models (LLMs) are mostly tested on isolated event pairs or coarse labels. We ask five LLMs to reconstruct the complete network of Allen’s 13 interval relations for 360 English narratives whose events are placed relative to one another rather than by dates, and evaluate the outputs under a pre-registered protocol for agreement with our gold annotations, converse consistency and logical satisfiability. End to end, the best model clearly beats trivial baselines on macro-F1, our primary metric, yet under our pre-registered parser no model matches the accuracy of always answering “before”, which the skewed labels inflate. Most of that accuracy gap is a parsing artifact: formatting errors in large matrices cost whole rows, and a post hoc parser that recovers the determined cells of malformed rows raises coverage from 0.60 to 0.84 and closes the gap to the baseline for the best model. On the cells they do answer, models recover precedence and containment but rarely overlap, and they fail most on flashbacks, where both the narration and the event list present the later event first. Logical consistency varies widely yet says little about agreement with gold: enforcing it raises accuracy mainly by filling empty cells, and is beaten there by the trivial control of guessing the majority label. Reliable structured output and robust handling of relative time are prerequisites for temporal reasoning over whole narratives.

[316] Facts Don’t Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation cs.CLPDF

Yongcan Yu, Jian Liang, Lingxiao He, Kuangpu Guo, Yanbo Wang

TL;DR: 该研究提出长链式思维蒸馏主要传递结构化推理行为而非新事实知识，并发现不同复杂度推理模式可能需要不同训练损失。

Details

Motivation: 长 CoT 蒸馏常依赖筛选正确轨迹并覆盖大量提示，但这些事实正确性与问题广度是否是推理能力提升的必要条件，机制尚不清楚。

Result: 作者观察到，仅用有缺陷的推理轨迹或范围受限的提示训练，也能获得近似的表现；进一步只在模型已掌握的问题上微调，仍报告显著推理增益，以此支持“行为假说”：关键在对结构化推理模式的行为对齐，而非记忆新事实。损失分析显示，基于概率的近期损失更适合学习简单模式，标准交叉熵则对扩展到复杂 CoT 必不可少。摘要未报告模型规模、任务集合、具体分数或“近似表现”的差距。

Insight: 若现象可复现，长 CoT 蒸馏的增益可能更多来自模仿推理行为，而不是训练样本注入了新知识；损失选择也可能随推理复杂度改变。现有证据在摘要中主要是定性描述，不能据此推断所有长思维链数据都可用错误轨迹替代正确轨迹。

Abstract: While long chain-of-thought (CoT) distillation significantly improves the reasoning abilities of language models, its underlying mechanism remains poorly understood. Current pipelines generally rely on rejection sampling to gather reasoning traces across diverse prompts, assuming that factual correctness and broad problem coverage are indispensable. Surprisingly, we find that training exclusively on flawed reasoning trajectories or a restricted prompt set yields nearly comparable performance. Driven by this counterintuitive finding, we propose the Behavioral Hypothesis: the essence of Long-CoT distillation is behavioral alignment with structured reasoning patterns, rather than factual knowledge memorization. We then validate this by demonstrating substantial reasoning gains even when models are fine-tuned exclusively on previously mastered problems-effectively isolating behavioral alignment from novel knowledge injection. Having established that reasoning patterns are the true bottleneck, we investigate how to learn them across varying levels of complexity optimally. We reveal a critical interplay: while the recent probability-based loss excels at learning simple patterns, the standard cross-entropy loss is essential for scaling to complex CoTs. We believe these insights demystify CoT distillation and provide a principled foundation for the future development of reasoning models. Code is available on https://github.com/yuyongcan/Long-CoT-Distillation.

[317] Many Steps into One: Reasoning Contraction and the Limits of Safe Memory Reuse for LLM Agents cs.CL | cs.AIPDF

Jianghong Zhou, Hao Rong, Chen Lin, Daoan Zhang

TL;DR: 论文研究 LLM 智能体复用记忆对推理的压缩及其安全边界，指出常见成功/失败计数器难以及时发现事实变化，并提出基于似然比的记忆认证与失效检测。

Details

Motivation: 技能模板、记忆和缓存能把多步推理压缩成一次调用，但一旦事实更新，过期记忆可能持续给出错误答案。只累计成功和失败次数的机制可能对变化反应过慢，甚至检测不到部分变化。

Result: 作者理论分析表明，对称计数器检测部分变化可能要等到误报间隔的幂次量级；不截断的计数器甚至可能永不察觉。在 5 个 LLM 的多跳事实问答实验中，ExpeL、ACE 风格计数器相较同等信息量的认证记忆，产生过时答案的次数为其 7.0–13.5 倍；在带精确验证器的代码任务上为 35–73 倍。作者进一步定义 certification load，并在 Zipf 需求下给出安全记忆可节省推理量的缩放规律。提出按各记忆自身似然比加权失败、通过在线推理认证效用、事实变更时重置证书，并用似然比检测器发现未通报变化；大幅编辑时直接失效记忆。

Insight: 记忆复用的安全性取决于持续监测“记忆仍然有效”，而不仅是历史成功率；作者将认证负担量化为安全省略推理的约束。倍数结果依赖具体基线、任务和事实更新设置，摘要未给出误报率等细节，不能据此直接估算部署成本。

Abstract: Thought templates, skill libraries, agent memories and caches turn many reasoning steps into one. We call this reasoning contraction. A contracted memory saves reasoning steps and errors until its facts change, then quietly answers wrong. Most systems trust memories by counting successes and failures. We prove that such counters are badly tuned change detectors: a symmetric counter notices a partial change only after a power of its false-alarm horizon, and an unclipped one may never notice it. On five LLMs answering multi-hop questions over edited facts, ExpeL- and ACE-style counters serve 7.0-13.5x more stale answers than certified memories given the same information, and on code tasks with an exact verifier 35-73x more. Safety also has a price. When facts change and each memory is judged only by its own verified outcomes, one number, the certification load, sets how much reasoning any safe memory can skip, and under Zipf demand it yields a scaling law. Our recipe weighs each memory’s failures by its own likelihood ratio, certifies its utility against live reasoning, and restarts the certificate when facts change, with the likelihood-ratio detector as the safety net for unannounced changes; for large edits it also invalidates the memory.

[318] The Image Still Has a Say: Final-Answer Repair for Multimodal Chain-of-Thought cs.CLPDF

Yulan Pan, Xiaoyu Xu, Zhihong Shen

TL;DR: TVER 不改写多模态思维链，而是在最终答案读取阶段将有限注意力从答案相关文本重新分配到与问题相关的图像区域，以训练免除方式修正答案。

Details

Motivation: 多模态推理链可能逐渐偏离图像，且一旦思维链生成完毕，常见方法通过重采样、训练或解码时 grounding 改善推理过程；作者转而利用最终作答时仍可访问图像这一点，只修复答案。

Result: 作者观察到，推理链可能过早偏向错误选项，但中间至深层仍保留支持正确选项的图像信息。TVER 选择与答案有关的文本和与问题匹配的图像 patch，在中间到深层的窗口中按问题相关性筛 patch，并将有界注意力从文本转向这些视觉证据。摘要称其在多个基准和骨干上提升了选择题与开放式回答，但未给出任务名、分数或增益。

Insight: 该方法把自我修正放在答案读出阶段，避免重写推理链，依赖的是图像证据在模型内部表示中仍可被恢复这一观察。效果幅度与适用条件在摘要中未量化，不能判断该操作对不同视觉架构或答案格式的影响。

Abstract: Although chain-of-thought prompting has improved multimodal reasoning, it is still difficult to produce correct final answers, as the reasoning can drift away from the image. Besides, the trace is already fixed when the answer is generated, leaving no room for self-correction. Considering the image is still available at the final-answer step, we think it is reasonable to fix the answer there without rewriting the reasoning. As such, the problem becomes to correct the final answer using the image under fixed reasoning. Unlike methods that improve trace generation through training, resampling, or decoding-time grounding, we repair the final-answer readout while keeping the reasoning trace fixed. The motivation is based on our finding that a wrong option can become prematurely committed in the reasoning while the correct option still has stronger image support in middle-to-deep layers. Thus, we propose Text-to-Visual Evidence Reallocation (TVER), a training-free method for final-answer repair. Specifically, TVER selects answer-related text and question-matched image patches, and moves bounded attention from text to patches. Furthermore, the patches are chosen in a middle-to-deep layer window based on question relevance. Our method improves answers across benchmarks and backbones, with gains in both multiple-choice and open-answer evaluations.

[319] Understanding Machine Taste: Aesthetic Agreement Across Foundation Models in K-pop cs.CLPDF

Kaishen Wang, Yucheng Pan, Xuehao Cui, Hanling Tian, Heng Huang

TL;DR: 该研究以 K-pop 歌曲和音乐视频为案例，将基础模型的审美偏好建模为成对排序，并比较模型间共识、视听偏好一致性及其与真实播放量的关系。

Details

Motivation: 基础模型正在进入创意产业，但模型如何评判作品、不同模态和模型间的偏好是否相同，以及机器排序与受众市场反应有何联系，仍缺少系统刻画。

Result: 作者比较模型间成对偏好一致性、群体层面的共识，以及音频和视觉偏好是否对齐。结果显示，视觉评价的模型间一致性和群体共识强于音频评价，而视听两种模态的偏好相关较弱。将视觉共识排序与历史 YouTube 播放量比较时，发现相关性为正，但在不同艺人群体间强弱不一。摘要未报告模型清单、样本规模或相关系数。

Insight: 机器审美并非单一稳定偏好：模型对视觉内容较易形成共识，但音乐与影像的偏好不能相互替代；与播放量的正相关也会随艺人群体变化。该研究属于 K-pop 案例分析，相关关系不能视为模型偏好导致市场热度的证据。

Abstract: As foundation models enter creative industries, understanding their aesthetic preferences becomes an important step toward AI-assisted creative decision making. In this work, we investigate how foundation models evaluate creative content and characterize their ``machine taste’’ using K-pop songs and music videos as a testbed. We formulate aesthetic evaluation as a pairwise ranking problem and systematically examine pairwise inter-model agreement, group-level consensus, and cross-modal preference alignment. Our results reveal stronger pairwise inter-model agreement and group-level consensus in visual evaluation than in audio evaluation, while preferences across the two modalities remain weakly aligned. We further examine how visual consensus rankings relate to real-world popularity using historical YouTube viewership data, finding positive but varying correlations across artist groups. These findings provide initial insights into machine aesthetic preferences and lay the groundwork for future AI-assisted creative decision making that considers creative content, artist identity, audience preferences, and market dynamics.

[320] Auditing Failure Attribution: Label Predictability and Judge Behavior cs.CL | cs.AI | cs.LGPDF

Tianyang Xu, Jing Gao

TL;DR: 该研究提出四类测试审计 LLM 故障归因基准是否真正要求核查出错步骤，还是只利用表面线索，并展示评测排名会受线索、输入范围和聚合指标影响。

Details

Motivation: 故障归因评测的高分应代表裁判模型将错误步骤与运行过程交叉核对，但可预测标签的表面线索也可能让模型轻易猜中，从而把线索利用误当作诊断能力。

Result: 作者提出输入限制、线索跟随、残差准确率和聚合效应四类审计。在三个基准上，简单预测器可达到或超过 GPT-4.1。两个 Who&When Pro 系统中，仅读某一步推理的分类器就能找到 95%–100% 的故障步骤；某系统的错误注入使故障后没有推理文本，针对其标签微调的裁判比 GPT-4.1 高 41 分；加入与任务相关的后续推理后，微调裁判下降 31–99 分，而零样本裁判最多下降 9 分。在 Aegis 上，“责怪所有智能体”与 GPT-4.1 的 micro-F1 相当，但 exact-set accuracy 低 19 分。

Insight: 故障归因的指标高分未必意味着模型核对了因果过程；要报告输入依赖、线索干预、线索控制后的能力以及聚合指标变化。示例反映了特定基准标注和注错方式下的捷径，不等于所有裁判均使用相同线索。

Abstract: Failure-attribution benchmarks ask an LLM judge which step or agent caused a failure. A high score should mean the judge checked the faulty step against the run (a cross-step check). However, spurious signals (surface cues) that predict the annotation can produce a high score. In this paper, we propose four tests to separate checking from cue use. Input restriction asks whether partial input predicts the annotation. Cue following checks whether moved cues change answers. Residual accuracy scores judges where cue-only predictors fail. Aggregation asks whether the failure mix or metric changes rankings. Real observations motivate these tests: on three benchmarks, simple predictors match or beat GPT-4.1. In two Who&When Pro systems, a classifier reading one step’s reasoning finds 95-100% of faulty steps, motivating input restriction. In one of them, error injection leaves no reasoning after the faulty step; judges fine-tuned on its labels beat GPT-4.1 by 41 points. Adding task-related reasoning after that step moves this cue: fine-tuned judges lose 31-99 points, but zero-shot judges lose at most 9, motivating cue following. On Aegis, blaming every agent matches GPT-4.1 on micro-F1 but trails by 19 points in exact-set accuracy, motivating aggregation. We recommend reporting these tests when developing failure-attribution benchmarks.

[321] Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking cs.CL | cs.AIPDF

Taehyeon Kim, Hyunsoo Lee, Youngsoo Jang, Moontae Lee

TL;DR: RPO 仅对推理模型的根 token（是否进入思考模式）进行组策略优化，使模型根据问题难度决定是否展开推理，降低自适应思考训练成本。

Details

Motivation: 大推理模型常对难度不同的查询都生成长推理，造成过度思考；开源模型缺少高效学会何时思考的机制，而既有自适应推理训练成本较高。

Result: Root-token Policy Optimization 使用组相对奖励及组内优势，只训练决定是否启动 <think> 的初始分支 token。作者在多个模型家族和规模上报告模型学会了按难度自适应思考，训练计算量约为既有自适应推理方法的 2%。摘要未报告具体模型、任务准确率、推理长度变化或显存节省幅度。

Insight: 若决策分支主要由开场 token 决定，只优化这一步可能以低成本获得计算路由能力。不过摘要没有说明 2% 的对照方法和算力口径，也未提供准确率/质量变化，无法判断训练省下的成本是否伴随决策性能取舍。

Abstract: Large reasoning models (LRMs) achieve strong performance by externalizing explicit reasoning traces before producing the answer, yet suffer from overthinking challenge that allocates uniformly heavy computation to queries of varying difficulty. While proprietary models mitigate this via opaque routing, open-source LRMs still lack an efficient mechanism to internalize adaptive reasoning due to both expensive training cost and limited disclosure of training recipes. In response, we introduce RPO (Root-token Policy Optimization), a framework that enables LRMs to self-determine when to reason by training only the initial root token (e.g., whether to invoke the tag) via group relative reward and group-wise advantages. By focusing on this pivotal branching point, RPO drastically reduces training overhead and VRAM usage. Across multiple model families and scales, RPO learns difficulty-aware adaptive thinking at just ~2% of the training compute of prior adaptive-reasoning methods.

[322] Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders cs.CLPDF

Xin Li, Hao Jiang, Jinghao Guo, Weisi Lin, Chau Yuen

TL;DR: ReMaRk 将同一推荐请求下多位回复者提出的标题合并成联合奖励目标，降低 GRPO 训练受单条抽样回复噪声支配的问题。

Details

Motivation: 推荐帖往往有多种合理答案，但 Rank-GRPO 每个样本只抽一条回复作奖励。作者报告该随机选择解释了其 advantage 方差的 85%，使优化信号更多取决于抽中了哪位回复者，而非推荐列表本身质量。

Result: ReMaRk 对每个生成列表一次性与所有回复标题的并集评分，并对请求等概率采样。在 7 个骨干、3 个数据集上，它用相当于逐回复训练 25%–53% 的步数达到 Rank-GRPO 2,000 步的验证召回率，并在该预算下领先；训练至两方法均不再提升时，ReMaRk 均未显著落后，且 7 组对比中有 4 组用更少 GPU 小时达到 Rank-GRPO 的最佳结果。相同奖励在每题约 100 名回答者的 ProtoQA 上也有效。

Insight: 多答案任务应把请求视作学习单位，并联合评价多种有效答案，降低回复者抽样带来的方差。结果覆盖推荐和常识问答两个设定；“标题并集”奖励能否适用于答案冲突、排序偏好或个性化权重不同的任务，摘要尚未讨论。

Abstract: A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies’ titles and samples each request equally. This makes RL cheaper. On seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO’s best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.

[323] Beyond Belief Nesting: Rethinking Higher-Order Theory of Mind in Large Language Models cs.CLPDF

Shanshan You, Zhengqi Wang, Hao Liu, Yuqin Liu, Ya He

TL;DR: HopToM 指出既有高阶心智理论基准常把信念嵌套深度误当作推理阶数，并提出经过低阶捷径控制的评测与推理时方法 HopTrace。

Details

Motivation: 理解战略互动中的高阶心智状态需要递归推断多层视角，但仅增加信念嵌套层数，问题仍可能由较低阶推理解答，导致现有基准夸大了模型的高阶能力。

Result: 作者以支撑连续信念前缀的不同事件形式化有效阶数，并构建含认证全阶问题及低阶替代控制的 HopToM。四个前沿 LLM 的直接提示平均准确率从一阶 73.5% 降至四阶 20.0%，人类参与者也面临困难。因果干预将一类常见错误关联到较浅信念前缀证据。即插即用的 HopTrace 将每段互动编码一次为视角记录，再确定性解析嵌套查询；在认证叙事问题上，其五阶和四阶准确率分别比强提示基线上升 38.5 和 22.0 个百分点，且 token 数比多轮调用方法少 20–39 倍。

Insight: 评测需验证真正需要几阶视角推理，而不是依赖嵌套表象；按视角存档并解析有望提升深层问题的效率。直接提示下四阶表现很低说明任务仍难；HopTrace 的增益限于作者构造的认证叙事问题，是否迁移到开放互动场景需进一步验证。

Abstract: Theory of mind (ToM), the ability to reason about others’ mental states, underlies effective communication and coordination. Its higher-order form enables reasoning over nested perspectives in strategic interactions. However, existing evaluations often equate belief-nesting depth with reasoning order, allowing nominally higher-order questions to admit lower-order shortcuts. We formalize effective order through the distinct events supporting successive belief prefixes, revealing widespread order inflation across existing benchmarks. We introduce HopToM, with certified full-order questions and controls for lower-order alternatives. Under direct prompting, mean accuracy across four frontier LLMs falls from 73.5% at first order to 20.0% at fourth order, and human participants also struggle. Causal interventions link a recurring class of errors to evidence for shallower belief prefixes. We introduce HopTrace, a plug-and-play inference-time method that encodes each interaction once into perspective records and resolves nested queries deterministically. On certified narrative questions, it improves fourth- and fifth-order accuracy by 22.0 and 38.5 percentage points over strong prompting baselines, while using 20-39 times fewer tokens than multi-call methods. These results sharpen how higher-order ToM is evaluated and offer a practical way to improve perspective reasoning in interactive agents.

[324] BRIDGE: Selective Boundary Repair with Dependency Graph Evidence in Multi-Agent Systems cs.CLPDF

Encheng Su, Pengze Li, Ziheng Hua, Qiucheng Yu, Jianyu Wu

TL;DR: BRIDGE 将多智能体交接修复作为依赖图上的选择性干预：从可观测交互轨迹定位未解决依赖，评估修复收益与扰动风险，并仅在局部后置条件成立时执行。

Details

Motivation: 协作流程中常发生关键上下文遗漏、误路由或被忽略的交接故障，即便单个智能体的局部答复看起来合理。直接修改交接可能带来副作用，因此需要可验证的介入边界。

Result: BRIDGE 配套构建故障归因与校准数据集，通过依赖关系解耦及反事实交接扰动提供监督；运行时把可见轨迹编译成来源依赖图，不访问智能体私有状态。经验门控器权衡预期收益与中断风险，且只有形式化局部后置条件成立才执行修复。在 SILO-BENCH 和 TeamBench 上，任务成功率最高分别提高 12.3 和 11.1 个百分点，额外 token 开销最低为 2.0%。在 MAST 的 562 条交接相关轨迹上，单次运行 F1 为 69.8%。

Insight: 选择性修复并设置可验证后置条件，尝试在协作收益和干扰风险之间建立运行时控制边界；无需读取私有智能体状态也有助于兼容封闭系统。摘要给的是最高任务成功增益与最低额外开销，不代表每个基准都同时达到这些极值。

Abstract: Multi-agent collaboration has emerged as a critical paradigm for tackling complex, long-horizon tasks; however, while coordination protocols continue to advance, systemic failures at the inter-agent collaboration interface remain a pervasive bottleneck. In collaborative workflows, systems frequently suffer from \emph{broken handoffs}—where critical task context is omitted, misrouted, or ignored despite each agent emitting locally plausible responses. To address this challenge, we present \method, a verifiable runtime control framework that treats handoff repair as selective intervention over explicit dependency graphs. Central to \method is an integrated fault-attribution and calibration pipeline, \dataset, which decouples execution dependencies and injects counterfactual handoff perturbations to provide empirical supervision for runtime intervention. Leveraging this diagnostic foundation, \method dynamically compiles observable interaction traces into a provenance graph to pinpoint unresolved dependencies, without accessing private agent states. It then evaluates candidate repairs through an empirical gate that weighs expected gains against disruption risks, and executes repairs only when formal local postconditions hold. Extensive evaluations across demanding benchmarks demonstrate that \method boosts task success by up to 12.3 absolute percentage points on SILO-BENCH and 11.1 on TeamBench. Across both backbones, \method maintains strict cost-effectiveness, adding as little as 2.0% marginal token overhead. Furthermore, on 562 handoff-relevant traces from the MAST benchmark, \method achieves 69.8% F1 in a single pass, establishing a reliable operational boundary for verifiable coordination repair.

[325] Text-Only vs. Image-Aware VLM Judges for Manga Translation Evaluation cs.CL | cs.CVPDF

Subaru Kimura, Jun Suzuki

TL;DR: 该研究用同一个视觉语言模型分别充当纯文本评审和图像感知评审，检验漫画翻译评估加入整页图像后会如何改变判分。

Details

Motivation: 漫画中的说话者身份、指代对象和场景语气往往依赖气泡外的画面信息，但现有翻译自动评估多只看文本，图像输入是否改善判断尚不清楚。

Result: 作者使用日英漫画官方译文数据，按页面和对话气泡对齐，并让同一 VLM 扮演 text-only judge 与 image-aware judge。加入页面图像后，主要 VLM 更能识别人为注入且依赖视觉上下文的翻译退化；两种评审的错误数量及类别分布也不同。然而人工核验未发现对纯文本或图像感知评审有稳定偏好。摘要未提供样本规模或指标数值。

Insight: 图像有助于识别依赖画面线索的错误，但评审方式变化不等于整体判断更接近人工偏好。评估漫画翻译时应分别报告视觉依赖错误识别与人工一致性；现有摘要仅基于日英数据和单一主要评审模型。

Abstract: In manga translation, visual context outside speech-bubble text, including speaker identities, referents, and scene tone, is depicted in page images, making it natural to refer to images during translation evaluation. However, most conventional manga translation evaluations rely solely on text-only automatic metrics, leaving the impact of image input insufficiently explored. We propose a practical validation framework that employs the same vision-language model (VLM) as both a text-only judge (TOJ) and an image-aware judge (IAJ) to investigate evaluation divergences associated with image-aware evaluation procedures. We use a Japanese-English manga dataset of official translations aligned at the page and speech-bubble levels. First, adding page images to TOJ improved the primary VLM’s detection of deliberately injected, visually dependent degradations. Second, we analyzed evaluation divergences between TOJ and IAJ, finding different error counts and error-category distributions. Human verification did not reveal a consistent preference between TOJ and IAJ.

[326] Agent-Reflex: Teaching Language Agents to Act Reflectively cs.CLPDF

Guanting Dong, Xinyu Yang, Jiajie Jin, Yuyang Hu, Xiaoshuai Song

TL;DR: Agent-Reflex 通过反思技能挖掘、反思监督微调和反思引导的在线策略优化，训练智能体在工具噪声与用户纠正下主动反思和恢复。

Details

Motivation: 长程工具智能体会遇到嘈杂工具反馈和用户纠正，但现有方法对反思学习及失败恢复的监督有限。

Result: 框架先向环境反馈注入有依据的扰动，从多能力模型池中挖掘匹配的成功—失败案例并总结可复用技能；再用 ReSFT 训练自我反思、技能恢复和面向用户的纠正行为；最后以 ORPO 结合噪声交互下的智能体强化学习和反思引导的在线蒸馏。在 Qwen3-8B 上，作者报告五个长程工具使用基准总体得分 58.6，AgentNoiseBench 得分 29.1，并称超过强环境扩展和反思基线；六套挑战性智能体基准进一步验证效果，噪声鲁棒性提升且训练稳定。摘要未解释得分量纲或各基准具体分项。

Insight: 将反思技能从对照成功与失败交互中归纳，再用于监督和策略优化，提供了从技能发现到在线行为学习的完整链路。基准总分口径与不同阶段的独立贡献在摘要中未明确，外推到真实工具噪声需参考具体评测协议。

Abstract: Large language models have evolved from single-turn chatbots into interactive agents capable of long-horizon reasoning. In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the agent, but existing methods provide limited supervision for reflection learning and failure recovery. To bridge this gap, we propose Agent-Reflex, a general training framework for incentivizing proactive reflection in agents. Specifically, Agent-Reflex comprises three stages. First, Environment Reflection Skill Discovery (RSD) injects grounded tool-feedback perturbations into environment scaling, samples a capability-diverse LLM pool, mines matched success-failure pairs, and summarizes reusable reflection meta-skills. Building on this skill library, we then apply Reflective Supervised Fine-Tuning (ReSFT) that converts each skill into supervision for self-reflection, skill recovery, and user-centric correction, strengthening complementary reflective behaviors. Finally, we propose On-Policy Reflective Policy Optimization (ORPO) to combine agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation to further elicit proactive reflection. Built on Qwen3-8B, Agent-Reflex reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, surpassing strong environment-scaling and reflection baselines. Experiments on six challenging agent benchmark suites demonstrate the effectiveness of Agent-Reflex. Further analyses show that Agent-Reflex improves agents’ robustness to noise while maintaining stable training dynamics, supporting scalable agent training. Our code is available at https://github.com/RUC-NLPIR/Agent-Reflex.

[327] Seeing Through the Forgery Is Not Enough: Regressive Sycophancy under Fabricated Evidence in Medical Vision-Language Models cs.CL | cs.AIPDF

Sadif Ahmed, Soumya Swagata Biswas, Rifat Shahriyar, Chowdhury Mohammad Rakin Haider

TL;DR: 该研究测量医学视觉语言模型在用户以伪造处方反驳正确影像判断时的回退式迎合，并发现声明文件伪造并不能保证模型保留原有诊断。

Details

Motivation: 模型可能正确读出 CT，却在后续对话中因用户提供相反方向的伪造文件而改变答案。医学场景中，即使最初识别正确，这类依赖伪证据的回退也可能造成风险。

Result: 作者在 400 个 VQA-RAD 是/否问题上，构造 19,200 段多轮对话，对 4 个开放 VLM 比较有无伪造处方的情况。在未说明文件系合成的设定中，模型放弃了 6,488 个正确回答中的 80.1%。不同脚本下，同一处方使 Qwen2.5-VL-7B 翻转率从 45.9% 升至 89.6%，却使 Qwen3-VL-8B 从 77.7% 降至 50.9%。在含处方的 3,232 段对话中，模型有 915 次（28.3%）拒绝处方同时保留正确答案，其中 40% 后续轮次又翻转。系统提示披露文件为合成内容，可将一模型翻转率从 77.8% 降至 47.3%。

Insight: 模型被告知证据是伪造的，仍可能在多轮对话中重新采纳错误答案；提示披露只能降低某一模型和脚本下的翻转率，不能视为可靠防护。结果来自 VQA-RAD 是/否任务及特定处方脚本，不能直接代表临床决策性能。

Abstract: A medical vision-language model may analyze a given CT scan and correctly detect both kidneys in view. Continuing this conversation, if a user provides a prescription with an opposite direction, medical VLMs often sycophantically choose to realign their diagnosis. In this work, we study regressive sycophancy under fabricated evidence, where a model abandons a correct answer when pushback comes with a fake document. We test four open VLMs on 400 VQA-RAD yes/no questions in 19,200 multi-turn conversations, with and without a fabricated prescription. Without a disclosure that documents are synthetic, the models abandon 80.1% of 6,488 correct answers. Notably, under one script each, the same prescription raises the flip rate of Qwen2.5-VL-7B from 45.9% to 89.6% and lowers that of Qwen3-VL-8B from 77.7% to 50.9%. A model rejected the prescription while still preserving the correct answer in 915 of 3,232 conversations with the prescription (28.3%), and 40% of these flipped in later turns. System-prompt sentences that disclose the documents as synthetic lower one model’s flip rate from 77.8% to 47.3%. Thus, we conclude that awareness about a forgery does not prevent a model from adopting its answer.

[328] From Agents to Edges: Paired Optimization for Multi-Agent Collaboration cs.CLPDF

Feng Lin, Xiaotian Yu, Yifan Yang, Aijun Li

TL;DR: PACO 把多智能体通信边作为优化单位，联合改写发送方和接收方提示及共享通信约定，在保持模型、角色和工作流拓扑固定的条件下改善协作。

Details

Motivation: 单体能力强不保证协作有效；通信质量取决于一方表达内容是否与另一方的解读和使用方式对齐，因此只单独优化某个智能体提示可能无法修复接口问题。

Result: PACO 面向黑盒语言模型，以状态感知的折扣 bandit 选择待改进通信边，通过执行反馈成对修订两端指令，并用分阶段端到端运行评估。优化后的指令可并入现有 agent prompt，部署时不增加模型调用。作者在一个四智能体工作流和多种多智能体基准上报告其持续优于领先提示优化与通信基线；在 HotpotQA 学到的配置迁移到其他数据集仍保持优势，无需针对目标集再优化。摘要未提供基准名称、数值增益或运行成本。

Insight: 联合优化通信两端比孤立改写单个角色更直接对应信息交接失配，同时保持系统拓扑不变，便于复用既有工作流。迁移结论来自 HotpotQA 到其他数据集的特定实验；摘要未量化迁移幅度和优化阶段成本。

Abstract: Strong individual agents do not guarantee effective collaboration. Communication must align what one agent conveys with how another interprets and uses it. This dependency makes the communication edge a natural unit of optimization. We introduce Paired Automated Communication Optimization (PACO), which jointly optimizes the instructions at both ends of each edge and their shared communication convention in systems built on black-box language models. PACO keeps model weights, agent roles, and workflow topology fixed. A state-aware discounted bandit selects edges for refinement, while execution feedback guides paired revisions evaluated through staged end-to-end execution. The optimized instructions integrate into existing agent prompts without additional model calls at deployment. On a four-agent workflow evaluated across diverse multi-agent benchmarks, PACO consistently outperforms leading prompt-optimization and communication baselines. Configurations learned on HotpotQA retain their advantage in cross-dataset transfer without target-specific optimization. These results demonstrate that optimizing how agents communicate can substantially improve collaboration within an otherwise fixed system.

[329] TextualRL: Textual Reinforcement Learning for Context Management cs.CL | cs.AIPDF

Tianyu Huai, Jie Zhou, Bo Mao, Qin Chen, Liang He

TL;DR: TextualRL 把同一任务、同一文本上下文下的一组 rollout 作为策略批评单位，跨组审查修改提案，并用留出奖励决定是否采纳上下文补丁。

Details

Motivation: 智能体可通过改写提示上下文适应任务反馈而不改模型参数，但单条轨迹提供的对比证据有限，任务特定修复也可能被写入共享上下文并过度泛化。

Result: 组内策略批评对照同一上下文下的成功与失败轨迹；若结果一致，则借助跨任务证据识别值得保留的行为或提出修复假设。跨组更新会结合包括未产生修改提案的观察，审查提案的证据和适用范围，再形成有界文本补丁，最终以留出验证奖励决定是否替换候选上下文。对 5 个目标智能体和 6 个基准的评估中，相比基座模型，GPT-5.5 平均测试准确率提高 24.4 个百分点，Qwen3.8-27B 提高 15.3 个百分点。

Insight: 通过多条同组尝试和跨任务审查，可以把奖励转化为更可迁移的文本策略更新，并让留出奖励承担采用门控。大幅度提升是针对摘要列出的智能体与六基准平均值，不能直接外推至任意提示上下文或任务分布。

Abstract: LLM agents can adapt to task feedback by revising the textual context that guides their behavior, without changing their model parameters. However, turning execution rewards into reusable context updates is challenging: individual trajectories offer limited comparative evidence about what to change, while task-specific fixes may overgeneralize when incorporated into a shared context. We introduce TextualRL, a gradient-free reinforcement learning framework that treats a group of rollouts from the same task under the same context, rather than an isolated trajectory, as the unit of policy critique. Groupwise Policy Critique contrasts successful and failed rollouts within a task; when outcomes are uniform, it draws on cross-task evidence to identify behaviors worth preserving or formulate repair hypotheses. Cross-Group Policy Update reviews the resulting proposals against evidence pooled across groups, including observations that yield no proposed edit, to refine their content and applicability. Reviewed proposals form a bounded textual patch, with adoption of the candidate context determined by held-out validation rewards. Together, these components establish a reward-driven policy optimization loop in which multiple attempts jointly inform reusable context updates. We evaluate TextualRL with five target agents on six benchmarks. Averaged over the six benchmarks, TextualRL improves test accuracy over the base model by 24.4 and 15.3 percentage points for GPT-5.5 and Qwen3.8-27B, respectively.

[330] TRACESTEER: Trace Driven Activation Steering for Hallucination Mitigation in Diffusion Language Models cs.CLPDF

Xin Zhang, Yili Wang, Yue Tan, Xin He, Rui Miao

TL;DR: TRACESTEER 根据扩散语言模型前序去噪预测和当前隐藏状态动态调整激活引导方向与强度，以测试时干预缓解幻觉。

Details

Motivation: 扩散语言模型并行预测 token，可能产生事实幻觉；既有激活引导忽略其去噪状态动态变化和幻觉风险随时间变化，固定引导向量与统一强度未必合适。

Result: DF-Steering 结合前序去噪预测和当前隐藏状态构造逐步引导向量；TS-Gate 根据预测不确定性与去噪步数调节强度。骨干参数冻结，利用事实监督和幅度正则学习自适应干预。作者在两个扩散语言模型上报告 TruthfulQA 事实可靠性提高，并能泛化到五个下游基准。摘要未提供模型名称、数值增益或延迟变化。

Insight: 把先前生成轨迹作为引导反馈，可让干预随去噪阶段与不确定性调整，而非始终使用同一个 steering 向量。虽然无需更新骨干，仍需学习引导模块；摘要未报告其校准成本、误导性抑制或事实性与通用能力的权衡。

Abstract: Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive LLMs, yet remain susceptible to hallucination. Activation steering offers a promising mitigation approach by guiding parallel token predictions through internal representations. However, existing steering methods largely overlook two characteristics of dLLMs: dynamic denoising states and temporal hallucination risks, which challenge predefined steering vectors and uniform intervention strengths. We propose TRACESTEER, a trace-driven activation steering framework that uses preceding denoising predictions as feedback to jointly adapt steering direction and strength. Specifically, Denoising Feedback Steering (DF-Steering) constructs step-specific steering vectors from preceding predictions and current hidden states. Trajectory-Guided Steering Gate (TS-Gate) regulates their strength using prediction uncertainty and the denoising timestep. With the backbone frozen, factual supervision and magnitude regularization jointly guide the learning of adaptive steering interventions, promoting factual reliability and stability throughout denoising. Experiments on two dLLMs show improved factual reliability on TruthfulQA and generalization across five downstream benchmarks.

[331] PACT: A Pragmatic Prompting Framework for Expressive and Voice-Preserving Speech-to-Speech Translation cs.CLPDF

Yuchen Song, Satoshi Nakamura

TL;DR: PACT 将源语音的语用线索映射并校准到目标语言可实现的表达空间，再以连续提示控制冻结语音生成器，以保留说话人音色并传达情绪、语气和节奏。

Details

Motivation: 语音到语音翻译不仅要保留字面内容，还应保留说话人身份和表达方式；现有端到端、语音多模态模型及级联系统在目标语言中复现情绪、语调、节奏、停顿和语用态度方面仍有困难。

Result: PACT 提取源语音语用线索，将其投影到目标语言表达空间、校准至可实现子空间，并转成冻结语音生成器的连续提示。在西班牙语↔英语的 TOPI 双向评测中，作者报告其在内容忠实度、音色保留和表达保留等指标上综合表现优于所评系统。摘要未给出模型名称、指标定义或具体数值。

Insight: 将语用特征先投影并校准到目标语言可表达范围，避免简单复制源语言表现线索；冻结生成器可减少对其内部训练的依赖。当前摘要只有 TOPI 双向评测的定性领先，尚不能判断不同语言、说话人或情绪类型的具体收益。

Abstract: Speech-to-speech translation should preserve not only what is said, but also who says it and how it is said. Existing end-to-end S2ST, speech MLLM, and cascade systems can preserve content or timbre, but often struggle to realize emotion, tone, rhythm, pauses, and pragmatic attitude in the target language. We propose PACT, a target-space pragmatic prompting framework for expressive and voice-preserving S2ST. PACT extracts source pragmatic cues, projects them into the target-language expression space, calibrates the projected representation to a realizable target subspace, and converts it into continuous prompts for a frozen speech generator. On bidirectional Spanish-English TOPI evaluation, PACT achieves the best overall performance among evaluated systems across content fidelity, voice preservation, and expression preservation metrics.

[332] Copying Is Not Grounding: Training Faithful Long-Form RAG with Authored-Sentence Credit cs.CLPDF

Hoang Pham, Duc Anh Vu, Duy Tung Doan, Cong-Duy T Nguyen, Anh Tuan Luu

TL;DR: ASC 将长篇 RAG 的忠实度奖励和 GRPO 优势从整篇回答移至句子级，只奖励模型原创、超出问题且有来源支持的句子，以抑制复制式奖励投机。

Details

Motivation: 按支持句比例设计的忠实度奖励容易被直接复制检索段落利用；KL 或复制惩罚可能仍保留复制，或换来更短、更空洞的回答及拒答，而整篇 GRPO 优势又会同时强化复制句和有根据的原创句。

Result: Authored-Sentence Credit 只为非复制、提供了超出问题的信息且受检索段落支持的句子计奖；每句优势与同一问题采样的其他句子比较，零信用句排名最低且不会被强化。在 4 个模型、2 个域内和 3 个留出语料上的实验中，ASC 是唯一在所有设置下都降低原创句错误率并增加受支持句的方案；留出评审下中位数改善分别为 38% 和 46%。两个 LLM 整体回答评审不偏好基线相较基座模型，而将 ASC 评价为与基座相当。

Insight: 句子级信用分配可减少“复制即支持”的奖励捷径，同时保留有证据支撑的原创表达。评审指标中，错误率与支持句数量改善并未带来整体回答优于基座的结论；其收益主要体现在相较其他训练基线的留出句级忠实度。

Abstract: In search engines and assistants, long-form RAG answers can mislead readers with claims their sources do not support. Faithfulness rewards, the share of supported sentences, train models toward fewer such claims. We show that these rewards are hacked by copying. A sentence copied from the passages is supported almost by construction, so trained models learn to copy rather than ground what they write. Common defenses only move the hack: a KL or copy penalty leaves copying in place or trades it for shorter, emptier answers and refusals. GRPO, the standard RL algorithm, makes it worse by giving every sentence its answer’s advantage, reinforcing copied and grounded sentences alike. We propose Authored-Sentence Credit (ASC) , which fixes both by moving GRPO’s reward and advantage from the answer to the sentence level. The reward credits only sentences that are authored rather than copied, say more than the question, and are supported by the passages. The advantage compares each sentence with every sentence sampled for the question, so one that earns no credit ranks last and is never reinforced. Across four models and two in-domain, three held-out corpora, only ASC lowers the error rate of authored sentences and writes more supported ones in every setting, by 38% and 46% at the median under held-out evaluators, giving readers more grounded information. It takes none of these shortcuts, and two LLM judges comparing whole answers prefer no baseline to the base model but rate ASC’s answers on par with it. We conclude that a faithfulness reward should credit what a model writes, not what it copies.

[333] Imprinting: Knowledge Distillation Can Shift Student Behavior cs.CLPDF

Yu-Chien Tang, An-Zi Yen

TL;DR: 该研究发现知识蒸馏不仅传递任务能力，也能把教师在安全—实用性上的行为倾向印入学生；即使训练题与安全无关，这种源教师效应仍会影响下游选择。

Details

Motivation: 蒸馏通常被视为能力迁移，但教师生成答案也包含行为倾向，可能在不属于蒸馏目标的下游场景中改变学生的决策。

Result: 作者构造同一基座教师的“接受伤害”和“避免伤害”两个版本，用它们回答与安全无关的一般推理和知识问题，再将答案蒸馏给不同基座学生。学生通常获得任务能力，同时在下游决策中朝源教师的安全—实用倾向移动；该效应依教师—学生组合变化，也不会均匀推广到所有安全行为。排除一小组分歧 token 可减弱印入；改写回答仍保留源教师效应，且可能引入改写器相关偏移。蒸馏后偏好优化提高了伤害回避，但没有消除教师来源造成的差异。摘要未给出具体分数。

Insight: 普通任务监督也可能通过教师回答风格携带行为规范，因此蒸馏数据不仅是任务标签来源。抑制特定 token 与后续偏好优化各有局限；效应随教师—学生配对和安全行为类型变化，不能简单视为固定方向或普适幅度。

Abstract: Knowledge distillation trains student models on teacher-generated responses to transfer task capability, but those responses may also carry behavioral tendencies of their source. We investigate whether the source teacher can shape student behavior beyond the capability that distillation is intended to transfer. We construct harm-accepting and harm-avoiding variants of the same base teacher, use them to generate responses to ordinary reasoning and knowledge questions unrelated to safety, and distill those responses into students with different base models. Students generally acquire task capability while their downstream decisions shift toward the source teacher’s safety-pragmatism disposition, a phenomenon we call imprinting. Distillation from the harm-accepting teacher reduces harm avoidance, whereas distillation from the harm-avoiding teacher increases it. Imprinting varies across teacher-student pairs and does not extend uniformly across different forms of safety behavior. Excluding a sparse set of divergence tokens reduces imprinting, whereas response rewriting preserves the source-dependent effect and can introduce additional rewriter-dependent shifts. Post-distillation preference optimization raises harm avoidance but does not eliminate the source-dependent difference. These results show that behavioral tendencies can be intentionally imprinted through the source teacher, even with ordinary task supervision.

[334] CREATE: Creativity Evaluation of VLMs using the Visual Remote Associates Test cs.CLPDF

Ziqing Yang, Anubhav Jangra, Smaranda Muresan

TL;DR: CREATE 用 155 道视觉远程联想题评估视觉语言模型的联想创造力，要求从三张看似无关的图像中找出共同概念。

Details

Motivation: VLM 被用于故事创作、设计构思、艺术解读和广告等创意工作，但缺少衡量其横向联想能力并与人类表现比较的系统基准。

Result: CREATE 为半自动构建的视觉 Remote Associates Test 基准，作者招募 25 人验证题目质量，并评测 11 个 VLM。结果显示多数模型表现低于人类；模型的联想能力依赖题目提供的文本提示概念，并不能普遍迁移到视觉谜题。摘要未报告各模型或人类的准确率和显著性检验。

Insight: 提示中的概念线索会显著影响模型能否完成跨图像联想，因而需要区分视觉线索推理与文本提示帮助。题量为 155 且经半自动构造，摘要未说明样本覆盖、文化因素和人类—模型差距的具体幅度。

Abstract: Vision language models (VLMs) are used in creative tasks such as visual storytelling, design ideation, art interpretation, and advertising. Yet we lack principled ways to measure the lateral reasoning of these models compared to human potential. We introduce CREATE, a semi-automatically constructed benchmark containing 155 visual Remote Associates Test problems. These problems require a solver to find an underlying connection among three seemingly dissimilar images. We conduct a human study (N=25) to validate and ground the quality of our benchmark, and evaluate 11 VLMs. Our findings show that most VLMs underperform humans on this task. Our experiments indicate that the associative ability of VLMs depends on the cue concepts provided in the text and does not universally extend to visual puzzles.

[335] Legal-R1: Agentic Retrieval for Evidence-Based Legal Reasoning cs.CLPDF

Huaiwen Pang, Jiale Han, Junying Chen, Wenjun Xiao, Benyou Wang

TL;DR: Legal-R1 用多智能体检索流程进行循证法律问答，由 Lawyer 决策何时、向何处检索及证据是否充分，Legal Assistant 按权威性和多样性检索，Judge 汇总证据作答。

Details

Motivation: 法律问答不仅需答案看似合理，还必须由适用且可引用的有效证据支撑；一次性语义检索难以同时处理法源权威性、适用性和异质来源的多样性。

Result: 框架将法律推理拆成 Lawyer、Legal Assistant 和 Judge 三个角色。仅对 Lawyer 进行掩码监督微调和轨迹感知强化学习；助手从规范性法律来源与判例中进行权威感知、多样性检索，Judge 基于完整证据轨迹作答。在多个中文法律基准上，作者报告 Legal-R1 优于通用模型、法律专用模型及检索增强基线。摘要未给出基准名称、数值或检索成本。

Insight: 检索决策本身可以作为可学习的法律推理策略，而不只是固定在一次检索步骤中；分工也使证据获取与最终判断相对解耦。摘要未报告具体法律领域、引用准确性和跨司法辖区泛化表现，暂不能量化可靠性提升。

Abstract: Legal question answering requires answers that are not only plausible, but also grounded in legally valid and citable evidence. Standard retrieval-augmented generation often relies on one-shot semantic matching, which is insufficient when legal evidence must be selected according to authority, applicability, and diversity across heterogeneous sources. We introduce Legal-R1, an agentic retrieval framework for evidence-based legal reasoning. Legal-R1 decomposes legal reasoning into a multi-agent workflow, where a Lawyer agent learns when to retrieve, which legal source to query, what evidence to seek, and when the collected evidence is sufficient. A Legal Assistant agent retrieves from normative legal sources and precedent cases with authority-aware and diversity-aware selection, while a Judge agent generates the final answer from the complete evidence trajectory. We post-train only the Lawyer agent with masked supervised fine-tuning and trajectory-aware reinforcement learning. Experiments on multiple Chinese legal benchmarks show that Legal-R1 improves over general-purpose, legal-specific, and retrieval-augmented baselines.

Zonghuan Li, Chirag Raman, Hayley Hung

TL;DR: 该研究联合分析自由站立对话中的 conversation floor 与 F-formation，指出仅把 F-formation 当作单一对话单元不足以完整描述参与者互动。

Details

Motivation: 互动参与度关系到谁影响谁，但现有会话群体分析往往依赖单一种类的空间或互动结构假设。作者认为 F-formation 未必能单独刻画一个会话。

Result: 研究提出用说话状态估计 conversation floor，并在 F-formation 内分析 floor；比较不同特征下的空间检测结果时发现，头部特征检测到的群组更可能被其他特征检测的群组所包含。联合空间特征与 floor 模式后，某些流程可用上半身特征检测单 floor 群组，同时用下半身特征保留 F-formation。作者据此认为，全面刻画互动参与仍需更细致的理论与框架。摘要未给出数据规模和检测指标。

Insight: 空间朝向形成的群组与轮次/说话结构并非一一对应；特征选择会影响检测到的群体嵌套关系。研究是对既有表示和管线的联合分析，摘要未给出标注基准或参与度预测结果，结论主要是概念性和诊断性的。

Abstract: Interactions happen universally and spontaneously in daily lives. Studying involvement in interactions helps us analyze who can influence who and has a wide potential in applications. Specifically, we consider conversational involvement in free-standing conversations. We study two operationalizations of an FCG: F-formations and conversation floors, and propose that current assumption that F-formations capture single conversations is insufficient. To illustrate how different implementations could impact the detected F-formations, we employ three analyses. First, we propose that speaking status can be used for estimating conversation floors, and leveraged it to analyze floors within F-formations. Second, we analyze spatial characteristics of the detection results by applying different feature sets on chosen baselines, and find that groups resulting from head features are more likely to be subsumed by groups detected using other features. Finally, we do a joint analysis which combines the spatial features and floor patterns in the conversation. We observe that certain pipelines can detect groups with single floors from upper body features, while retaining F-formations with lower body features. We conclude that a more detailed theory and framework is needed to holistically demonstrate conversational involvement within interactions.

[337] Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve cs.CL | cs.AI | cs.IRPDF

Yuyang Bai, Yu Wang, Yu Zhang

TL;DR: DKGI 允许语言模型智能体直接在知识图谱节点与边文件上编写并迭代 Python 访问工具，使工具创建与证据检索共同演化。

Details

Motivation: 文本丰富的知识图谱搜索需同时沿关系扩展和审查节点文本；预定义图接口限制了可访问证据，即使智能体能编排工具或处理检索子图也受限。

Result: 在 DKGI 中，智能体直接访问完整节点和边文件，通过代码遍历关系、组合节点集合、筛选文本证据，并交替使用程序执行、shell 搜索和文件检查。执行反馈引导工具创建与修订，工具的变化又决定后续检索。在跨三个领域的基准上，无需训练的 DKGI 在每个图上均比已发表最强基线至少高 6.7 个 Hit@1 点，即使使用最简 harness 仍如此。控制消融显示，多轮编程优于纯 shell 访问及预定义图工具；轨迹显示智能体会修改代码扩展搜索或修复错误。

Insight: 开放工具构造可让智能体按当前证据需求编写检索程序，而非被固定图接口限制；多轮执行反馈形成工具与搜索策略的共同适应。基准优势针对三领域图数据，摘要未报告运行成本和大图访问的资源边界。

Abstract: Searching text-rich knowledge graphs requires following relations while examining textual evidence. Predefined graph interfaces constrain what evidence agents can access, even when they write code to combine tool calls or process retrieved subgraphs. We propose direct knowledge graph interaction (DKGI), a harness in which language model agents create graph access tools directly over complete node and edge files. The agent writes Python code to follow relations, combine node sets, and select textual evidence, interleaving execution with shell search and file inspection. This lets agents process large node sets in bulk while returning selected text for the model to assess. Execution feedback guides tool creation and revision, while the evolving tools determine what evidence the agent gathers next. On benchmark datasets spanning three domains, training-free DKGI exceeds the strongest published baseline by at least 6.7 Hit@1 points on every graph, even with a minimal harness. Controlled ablations show the benefits of multi-turn programming over shell-only access and predefined graph tools, while trajectories show agents revising code to extend searches or address errors. These findings highlight the value of allowing tool creation and tool use to co-evolve during graph search. Our code and data are available at https://github.com/leopoldwhite/DKGI.

[338] Rubrics from the Task, Verdicts from Evidence: Pointwise Rubric Rewards with Tool-Augmented Judges cs.CLPDF

Tianci Liu, Haoyu Wang

TL;DR: ReVeR 从任务指令本身推导硬约束和分级原则，并用工具增强评审独立评价每个回答，避免根据候选回答反向构造量规所引入的偏差。

Details

Motivation: 量规奖励模型若从已观察回答中提取标准，可能产生回答诱导偏差；仅靠阅读文本也难检查机械约束，而成对裁判只给相对偏好，不便直接用于策略优化。

Result: ReVeR 是逐点评分量规奖励模型：量规从指令推理生成，工具增强裁判再独立评价回答。两部分用标签盲蒸馏训练，偏好标签仅筛选教师输出，形成大规模 ReVeR-Traces。作者报告其平均偏好准确率超过成对量规模型，指令遵循任务优势尤明显，并在测试时扩展和下游策略优化中有效。进一步发现，生成器 RL 会提升偏好准确率却降低量规忠实度；更高偏好准确率也不必然带来更好的下游策略。摘要未提供各指标的数值。

Insight: 量规忠实度、偏好预测准确率和策略效用是不同性质，单一奖励模型分数不能替代下游验证。工具增强的点式判断可支持优化，但摘要未说明工具种类、评估模型以及不同量规复杂度下的成本。

Abstract: Rubric-based reward models (RMs) evaluate responses against explicit criteria. However, existing methods often derive criteria from observed response candidates, introducing response-induced bias that may limit rubric generalizability. Their judges also struggle with mechanical constraints when relying on reading alone, and commonly used pairwise judges provide only relative preferences, complicating their direct use for policy optimization. We propose ReVeR, a pointwise rubric RM that derives hard rules and graded principles by reasoning about the instruction alone and evaluates responses independently through tool-augmented judging. Both components are trained through label-blind distillation, in which preference labels only filter a teacher’s outputs, yielding ReVeR-Traces, a large-scale dataset of rubric-reasoning and judging trajectories. Experiments demonstrate that ReVeR outperforms pairwise rubric RMs in average preference accuracy, most clearly on instruction following, as well as in test-time scaling and downstream policy optimization. We further find that generator RL improves preference accuracy while reducing rubric fidelity, and that higher preference accuracy does not necessarily translate into better downstream policies. These findings highlight that rubric fidelity, preference accuracy, and downstream utility are distinct properties, shedding light on future RM research.

[339] DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning cs.CL | cs.AI | cs.LGPDF

Minghao Luo, Liang Chen

TL;DR: DoubtLess 用更小的推理模型分别生成简洁和详尽倾向的 logits 差，并逐 token 调节大模型 CoT，从而无需重新训练地缩短推理。

Details

Motivation: 推理模型常生成超出解题需要的长思维链；既有压缩方法可能需要逐模型训练、添加固定 steering 向量或提前停止/屏蔽词语，后两类启发式可能降低准确率。

Result: 小模型读取大模型目前的 CoT 两次，一次被提示简短、一次被提示详尽；两组 logits 差乘以强度后加到大模型每一步 CoT logits。Qwen3-32B 在竞赛数学、研究生科学和竞赛编程七个基准上 CoT 缩短 31%，准确率保持不变，并超过五种推理时基线。用于自调试编码智能体时，每轮思考 token 减少 20%，任务表现不变；在另外两个模型家族上也保持准确率。与训练式压缩结合还能进一步缩短 CoT。分析显示主要节省来自正确答案首次出现后的重复检查。

Insight: 通过较小模型估计“多想/少想”的方向，可在不更新大模型参数时控制推理长度；节省集中于答案已出现后的反复核验，而非删除主要解题步骤。摘要未给出小模型额外调用的时延/成本，token 缩短不必然等同于端到端成本下降。

Abstract: Reasoning models achieve strong performance through long chains of thought (CoT), but often spend more tokens than needed to solve a problem. Existing ways to shorten the CoT require additional training for each model, add a fixed steering vector to the model’s hidden states, or use decoding heuristics such as stopping early or banning words like “wait”, which can reduce accuracy. We propose DoubtLess, a training-free decoding method in which a much smaller reasoning model steers the large one towards brevity. The small model reads the large model’s CoT so far twice, once prompted to be brief (the brief thinker) and once prompted to be verbose (the verbose thinker), and the difference between its two logit vectors, scaled by a strength, is added to the large model’s logits at every step of the CoT. On seven benchmarks of competition mathematics, graduate-level science and competitive programming, DoubtLess shortens the CoT of Qwen3-32B by 31% with unchanged accuracy, outperforming five inference-time baselines. We also show that DoubtLess reduces thinking tokens per turn by 20% in a self-debugging coding agent with unchanged task performance. It also generalizes to two other model families with unchanged accuracy. When combined with training-based compression, it further shortens the CoT of a model already trained for brevity. Ablations and behavioral analysis show that DoubtLess reduces repeated checking, with most token savings coming after the correct answer first appears in the CoT. Our code is available at https://github.com/leoluolol/DoubtLess.

[340] Hybrid Reasoning Agents cs.CL | cs.AI | cs.LGPDF

Jiaru Zou, Yunzhe Qi, Sirui Chen, Federico Bianchi, Shang Zhu

TL;DR: HybridAgents 在文本交互和潜空间思考之间动态切换，用连续工作记忆保留多轮混合上下文，并通过 HybridSFT 与 HybridRL 学习何时内部推理、何时调用工具行动。

Details

Motivation: 文本推理易检查但冗长；潜空间推理更紧凑，却缺乏原生工具访问和透明性。现有智能体难以在统一流程中兼顾低成本推理和可检查的外部行动。

Result: 轻量 HybridLink 负责文本/潜空间模式转换，连续工作记忆无损保存混合上下文。HybridSFT 训练模型协同使用工具和转换模块，HybridRL 学习何时在潜空间思考、何时通过文本行动。在 3 个模型规模、覆盖数学、科学、医学、搜索和智能体推理的 11 个基准上，作者报告相较文本与潜方法基线平均性能提高 10.4%，token 减少 35.5%–54.7%，推理速度提高 1.8–2.6 倍，成本降低 2.7–3.7 倍。

Insight: 让潜空间处理内部推理、文本负责可见行动，可尝试同时获得效率和工具可控性；连续工作记忆用于跨轮次保留两种表示。摘要的平均性能及成本比例跨 11 个基准和 3 个规模汇总，未给出逐任务质量变化和潜在表示可审计性的细节。

Abstract: How can we build an agent that thinks beyond human language, yet still acts through it in the external environment? Modern LLM agents still reason and interact through text, while latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. We bridge these two directions with Hybrid Reasoning Agents (HybridAgents), which dynamically navigate between latent reasoning and text-based interaction throughout a unified generation process. Architecturally, a lightweight HybridLink enables seamless transitions between text and latent modes, while the agent’s continuous working memory preserves lossless hybrid context across multiple turns. These designs preserve tool-integrated agentic capabilities while making latent thoughts directly inspectable. Building on this architecture, we further develop dedicated post-training algorithms for hybrid reasoning: HybridSFT stabilizes tool-integrated hybrid generation by teaching the backbone to coordinate with HybridLink, then HybridRL scales reasoning by learning when to think in latent space and when to act through text. We evaluate HybridAgent at 3 model scales across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. Compared with frontier LLM Agents and agentic frameworks spanning both text-based and latent methods, HybridAgent consistently improves performance by 10.4%, while delivering 35.5%-54.7% token reduction, 1.8x-2.6x faster inference, and 2.7x-3.7x cost savings. Code and data are available at https://github.com/togethercomputer/HybridAgent.

cs.AI [Back]

[341] Artificiality’s First Exam: An LLM-Referenced C-test for Supraintelligence cs.AI | cs.CL | cs.LGPDF

Jose Hernandez-Orallo

TL;DR: 该工作把 Levin Kt 复杂度思想与 LLM 作为表征机器结合，提出以模型能力为参照、可按比特尺度估计难度的 C-test，用于评估超越当前人类或模型能力的推理题。

Details

Motivation: 按人类或当前 AI 能力校准的基准可能很快循环或饱和；要评测未来更强系统，题目难度应尽量依据任务机制复杂度，而非只依赖当前被试表现。

Result: 作者把 LLM 视为包含人类文化知识、可用自然语言提示编程的表征机器，形式化基于 Levin Kt 的难度测量，并设计轻量题目生成与认证流程；分析参考机器能力及出题者—被测者差距。首批认证题在 12 个 LLM 上评测，模型能力曲线随估计难度提高而下降，并延伸至当前难以探索的高难度区域。作者称经认证题目难度可在比特比率尺度上估计，并可在当前前沿模型尚不能解答前生成。摘要未提供各模型成绩或单位难度的定义细节。

Insight: 若题目难度能与当前考生能力分离并用机制复杂度标定，基准有望降低饱和风险；但难度依赖参考 LLM/机器及认证过程，参考系统变化会影响刻度。12 个模型的首批实验支持可行性，尚不足以证明该量尺对任意超人类系统都稳定。

Abstract: Under the accelerated pace of AI development, any benchmark whose difficulty is calibrated with humans or current AI systems becomes circular, or saturates soon. How can we then evaluate machines that may become more capable than the current machines and the people who write the tests? One answer to the question hinges on inverse problems, inductive inference questions for which difficulty depends on the complexity of the underlying mechanism, given the observations and the previous knowledge. Levin’s Kt, a computable variant of Kolmogorov complexity, puts a theoretical upper bound to that difficulty, balancing the simplest representation encoding for the mechanism and the logarithm of the time steps to run that mechanism. Previous tests derived from Levin’s Kt, such as the original C-test, are based on minimalistic Universal Turing Machines (UTM), capable of very limited representations and no previous knowledge. Reasoning large language models (LLM) are UTMs that capture a compendium of human culture as knowledge, using prompts as natural language programs. This paper presents the first test based on Levin’s Kt with LLMs as representational machines. Fresh test items can be produced at high difficulties, estimated in a ratio scale measured in bits, before any human or frontier model can solve them. We formalise the measure, devise a lightweight generation and certification pipeline, explore the effect of the power of the reference machine and the gap between examiner and examinee. We produce a first battery of certified items administered to twelve LLMs. The resulting agent characteristic curves decrease with difficulty, down to the uncharted high-difficulty region, the supra-intelligence window of items we can certify, yet beyond generations of frontier intelligence.

[342] Decoupling Knowledge Forgetting and Refusal Learning in MLLM Unlearning cs.AI | cs.CL | cs.IR | cs.LGPDF

Haodong Chen, Shengtao Wen, Dong Liang, Lingbing Guo, Xiang Chen

TL;DR: FASD 将多模态遗忘拆成拒答行为学习和目标知识抑制两个通道，目标是在忘掉指定内容的同时诚实承认知识缺失，而非编造或只学会表面拒答。

Details

Motivation: 只看模型是否不再输出目标信息会掩盖两种失败：抑制后可能对相关问题编造事实；拒答训练也可能只改变表面行为，模型在选择题中仍能识别目标知识。

Result: 行为通道从提示为无知状态的模型副本蒸馏恰当拒答，并屏蔽教师分布中的目标 token，避免重新教回遗忘内容；知识通道则直接抑制真实目标答案片段，不依赖模型生成结果。作者报告 FASD 能遗忘目标知识、减少开放式任务的信息泄漏，并保持保留数据表现；摘要未列评测基准或具体数值。

Insight: 将行为上的“会拒答”与知识上的“确实不再保留”分开优化，可避免把沉默误当作遗忘。现有结论在摘要中为定性描述，未给出多选泄漏率、遗忘—保留权衡及不同知识类型上的差异。

Abstract: Current evaluations of multimodal unlearning typically focus on whether a model stops outputting target information. This surface-level criterion, however, masks two critical failure modes. First, knowledge suppression techniques often induce confabulation, leading the model to hallucinate facts about the unlearned subject. Second, refusal training merely alters observable behavior without erasing the underlying knowledge; such models might decline open queries but still identify correct answers under multiple-choice probing. To overcome these limitations, we propose Forget-Aware Self-Distillation (FASD). This method is designed to achieve honest ignorance, a state where the model genuinely loses the target knowledge and learns to acknowledge its absence. FASD separates these goals into two complementary channels. The behavior channel distills appropriate refusals from an ignorance-prompted copy of the model itself. Crucially, target tokens are masked from the teacher distribution to avoid accidentally reteaching the forgotten concepts. Simultaneously, the knowledge channel directly suppresses ground-truth answer spans independently of the model’s generated responses. Experiments demonstrate that FASD successfully unlearns target knowledge, prevents information leakage in open-ended tasks, and preserves performance on retained data. The resulting model not only forgets the specified information but accurately recognizes the boundaries of its own knowledge. Code and data will be released upon publication.

[343] How to Build a Benchmark That Actually Measures Taste cs.AI | cs.CL | cs.CVPDF

Xiaofan Wu, Xi Zeng, Miaoxia Chen, Peishan Chen, Shuyan Li

TL;DR: TasteBench 提出带专家标注、线索排查和侧位随机化的成对审美基准，并用试点揭示位置偏差、文件名泄漏和题目过易等效度威胁。

Details

Motivation: 与专家一致的“品味”基准应测量工艺判断，而非模型利用测试构造线索；若标签、选项位置或文件名泄漏，模型高分并不能证明审美判断能力。

Result: 基准覆盖视觉、品牌、书籍和科技物品，使用冻结量规、至少三位专家金标及明确弱标签隔离；Round-1 试点主分析含 250 对。审计发现首选项以 0.956 概率位于 A，恒答 A 得 0.956；文件名 token 线索得 0.912。书籍、品牌、科技领域的人类一致率近乎满分；随机化 A/B 后五个 VLM 在这些领域仍全得 1.0，视觉领域约 0.84–0.91，事后困难子集 107 对仍无法降低工艺领域天花板。侧位随机化工具将偏斜从 0.916 改至 0.495/0.505。作者明确未评测互动度、时代/圈层条件或校准，也不声称通用品味 SOTA。

Insight: 基准验证协议能够发现并修复位置偏差与文件名泄漏，但工艺领域仍存在题目难度上限，单靠去线索无法让模型与人类拉开差异。该工作是对基准效度的审计而非模型品味能力竞赛；Round-2 需增加更难题目。

Abstract: Expert-aligned “taste” benchmarks should reward craft judgment, not the ability to game the test. We introduce TasteBench, a pairwise good-vs-bad benchmark across visual, brand, book, and tech items under a frozen rubric with expert-only gold (panel n>=3; percent agreement >=0.60 primary gate; weak engagement labels train/analysis-only), together with a control protocol that separates craft judgment from test-gaming. Applying the protocol to a Round-1 pilot (primary_human_expert; three biological human experts; n=250 pairs) surfaces and quantifies three validity threats: two construction artifacts that inflate the trivial-baseline layer, and a difficulty ceiling that inflates every score. (1) Position: the preferred item is slot A with probability 0.956, so an always_item_a policy scores 0.956 (fair chance 0.5). (2) Lexical leakage: a filename token cue scores 0.912. (3) Difficulty ceiling: craft domains show near-ceiling human agreement (book/brand/tech 1.0), and even under A/B randomization (eval_assignment_seed7; gold_side_A ~~0.47) all five VLM cards reach 1.0 on craft, with only visual separating (~~0.84-0.91; a post-hoc hard subset, n=107, leaves craft at 1.0). The protocol (expert-only gold, a weak-label firewall, side randomization at construction, filename scrubbing, and hard filters) makes each artifact measurable and, for the two construction artifacts, correctable; we release a reference eval harness and a side-randomization tool, and show the tool removes the position skew (0.916 to 0.495/0.505). The result is a calibrated view of what the benchmark does and does not yet measure: position and leakage are fixable, whereas the craft-domain ceiling is a pair-difficulty limit that Round-2 must address. We do not evaluate engagement, era/circle conditioning, or calibration here, and we claim no general taste SOTA.

[344] CUSE: Fine-Grained Multimodal Unlearning with Counterfactual Guidance and On-Policy Self-Distillation cs.AI | cs.CLPDF

Chenlu Ding, Jiancan Wu, Yanchen Luo, Tianyu Zhang, Yancheng Yuan

TL;DR: CUSE 通过移除目标证据的反事实输入识别应遗忘的 token，再对当前策略采样回答进行 on-policy 自蒸馏，细粒度地同时学习遗忘和保留。

Details

Motivation: 多模态模型遗忘敏感、版权或不安全信息时，需要删除目标内容而保留同一回答中的非目标部分；既有方法常对整段回答施加统一目标，并依赖固定离线监督，与当前策略行为不匹配。

Result: CUSE 构造去除目标证据的反事实输入，根据 token 概率变化确定需加强遗忘的部分；随后在当前策略采样的回答上蒸馏 token 级遗忘和保留目标，冻结原模型副本作为教师，并提供相应遗忘/保留特权上下文。在 MLLMU-Bench 上，CUSE 相较最强基线将 forget-set ROUGE-L 降低 33.6%，同时 retain-set ROUGE-L 高 1.2%。

Insight: 反事实输入提供 token 级遗忘重要性信号，on-policy 训练则使修改针对模型当前会生成的内容，有助于兼顾遗忘与保留。摘要只报告 ROUGE-L 指标，不能据此判断其他语义泄漏、拒答行为或不同目标强度下的普遍权衡。

Abstract: Multimodal large language models (MLLMs) can reproduce sensitive, copyrighted, or unsafe information from their training data. Effective unlearning must remove target information while preserving non-target content, even within the same response. Existing methods typically apply uniform forgetting objectives to complete responses and rely on fixed, off-policy supervision, resulting in coarse token control and misalignment with the model’s current behavior. We propose CUSE, a Counterfactual-guided Unlearning framework via on-policy Self-distillation for multimodal large language models. It first constructs counterfactual inputs by removing target-related evidence and uses the resulting changes in token probabilities to identify which tokens require stronger forgetting. It then distills token-level forgetting and retention targets on responses sampled from the current policy, with a frozen copy of the original model serving as the teacher model augmented with the corresponding forgetting- and retention-oriented privileged contexts. This design jointly determines what to forget and what to preserve. On MLLMU-Bench, CUSE reduces forget-set ROUGE-L by 33.6% while achieving 1.2% higher retain-set ROUGE-L than the strongest baseline.

[345] AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery cs.AI | cs.CL | cs.LG | q-fin.PMPDF

Jie Huang, Yuecheng Hong, Yao Chen, Hao Zhong, Linfeng Xiao

TL;DR: AlphaRLM 通过文献引导算子组装和市场反馈奖励，让语言模型直接强化学习生成可执行的符号化量化因子，而无需输入种子因子表达式。

Details

Motivation: 公式化 alpha 挖掘需要在庞大、噪声多的搜索空间找出有效预测信号；现有 LLM 方法多依赖提示、迭代搜索或种子因子演化，直接以策略学习构造因子仍较少。

Result: LEGO 根据金融文献提取的思想和可执行算子为因子构造提供语义先验；奖励结合表达式可执行性与分组收益区分度，以捕捉因子分位数间的非单调结构。在 CSI 300、CSI 1000、S&P 500 上，作者报告 AlphaRLM 的年化超额收益超过强因子挖掘及 LLM 基线；8B 模型也展现跨市场表现。摘要未列超额收益、回撤或交易成本数值。

Insight: 文献语义先验与可执行反馈共同引导符号搜索，避免把因子生成限制为已有表达式的变异。金融收益高度依赖回测设置；摘要未报告时间切分、交易成本、风险调整表现或数据泄漏控制，无法据此判断可投资性。

Abstract: Formulaic alpha mining seeks to discover executable expressions that capture predictive market signals in a vast, noisy search space. Recent LLM-based approaches use prompting, iterative search, or seed-factor evolution, leaving direct factor construction through policy learning less explored. To address this, we introduce AlphaRLM, a reinforcement-learning framework that trains LLMs to construct executable symbolic factors from financial literature without input factor expressions. AlphaRLM incorporates Literature-Guided Operator Assembly (LEGO), which conditions factor construction on literature-derived ideas and executable operators, providing semantic priors for exploration. It further aligns the generation policy with market feedback through executable-validity and grouped-return-separation rewards that capture non-monotonic structures across factor quantiles. Extensive experiments on CSI 300, CSI 1000, and S&P 500 show that AlphaRLM outperforms strong factor-mining and LLM-based baselines in annualized excess return. Notably, the 8B model achieves strong cross-market performance, highlighting the promise of literature-conditioned, market-aligned factor generation.

[346] RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning cs.AI | cs.CL | cs.LGPDF

Ye Rin Kim, Yongjun Jeong, Seongjun Yun, Hanbum Ko, Soohyun Oh

TL;DR: RetroPlanner 让 LLM 同时比较多条逆合成路线，并按反应可信度、往返一致性和起始原料价格等多目标引导搜索，而非只追求找到可购买原料路线。

Details

Motivation: 逆合成规划常把能到达可购买原料视作解决，但真实路线还需考虑可合成性等实用目标；现有 LLM 规划器多每轮扩展一个分子且仅优化可解性，搜索预算提交前缺少路线间比较。

Result: RetroPlanner 训练模型基于化学证据推理：目标结构进展、键变化和反应类别、候选路线分歧位置，以及各路线在多个目标上的排序。多步逆合成基准实验中，作者报告其比既有 LLM 规划器找到更多替代路线，同时提升合成路线质量并维持较高成功率。摘要未给出基准名称、成功率或路线质量数值。

Insight: 多路线并行比较让搜索能在投入预算前权衡可行性、原料成本和反应一致性，避免只找到第一条可解路线。结果只有定性报告，路线质量指标、化学家评估和实物合成验证未见于摘要。

Abstract: Multi-step retrosynthesis planning is often considered solved once a route to purchasable starting materials is found, but such a route should also be practical in real-world chemistry. Recent LLM-based planners formulate this process as sequential decision-making over an AND-OR graph, yet typically expand a single molecule per turn and are optimized for solvability alone, leaving alternative routes uncompared across the multiple objectives that determine practicality before the search budget is committed. We introduce RetroPlanner, an LLM-based planner that jointly compares multiple routes to steer the search toward higher synthesizability, as measured by reaction plausibility, round-trip consistency, and starting-material price. RetroPlanner is trained to reason over chemical evidence: structural progress from the target molecule, bond changes and reaction classes, where candidate routes diverge from one another, and how they rank across multiple objectives. Experiments on multi-step retrosynthesis benchmarks demonstrate that RetroPlanner recovers more alternative routes than prior LLM-based planners while improving synthetic route quality and maintaining a high success rate.

[347] EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents cs.AI | cs.CL | cs.LGPDF

Min Cai, Dan Zhang, Gao Xing, Libin Ma, Zhaohui Xu

TL;DR: EAU 让用户模拟器、环境和智能体策略在 RL 训练中共同演化，根据轨迹、工具记录和奖励诊断失败并热更新用户、环境、奖励及训练配置。

Details

Motivation: 多轮用户交互智能体的学习信号不仅由智能体产生，也受随机/不一致用户模拟器、不可靠工具、环境状态覆盖不足和错误验证器影响；把这些组件固定会使训练不稳定。

Result: EAU 元智能体持续分析轨迹、工具调用、奖励和训练统计，定位失败与信息量不足的互动，再提出对用户模拟器、环境、奖励设计和 RL 配置的定向修改；验证通过后可热切换至进行中的训练，且改进跨运行保留。用户和环境动态采用可编辑符号表示，智能体仍由神经 LLM 进行 RL。三个交互智能体基准、两个骨干上，整体分比最强基线提高 14.7%–15.0%；ColBench 成功任务完成率最高提高 36.7%。

Insight: 训练系统本身也可作为学习对象，持续修补用户模拟和环境缺陷，避免仅用奖励塑形补偿坏数据。结果覆盖三个基准和两个骨干，摘要未说明各类环境修改比例、自动验证可靠性及长期热更新是否引入评测泄漏。

Abstract: Training LLM agents to interact with users over multiple turns is challenging because the learning signal depends not only on the agent, but also on the user simulator and environment, which jointly shape the agent’s experience. Simulated users can be stochastic or inconsistent, while environments may contain unreliable tools, incomplete state coverage, and incorrect verifiers. Prior work mainly addresses the resulting instability through reward shaping or denser supervision, while treating these components as fixed. We instead introduce EAU, a meta-agent framework for environment–agent–user co-evolution. During training, EAU monitors trajectories, tool traces, rewards, and training statistics; diagnoses failures and uninformative interactions; and proposes targeted modifications to the user simulator, environment, reward design, and RL configuration. Validated changes are hot-swapped into the ongoing run, allowing the training setup to evolve together with the policy. These adaptations also persist across runs, so later training can build on previously discovered improvements rather than restarting from a fixed setup. We enable this process through an editable symbolic representation of user and environment dynamics, while the agent remains a neural LLM optimized with RL. Across three interactive-agent benchmarks and two backbone models, EAU consistently outperforms prior RL methods, improving the overall score over the strongest baseline by 14.7–15.0% and increasing successful task completion on ColBench by up to 36.7%, showing that co-evolving the training system with the policy can substantially improve multi-turn RL.

[348] Learning Perceptual Geometry of Natural Odor Mixtures from LLM-Derived Descriptors cs.AI | cs.CL | q-bio.NCPDF

Farhad Pashakhanloo, Venkatesh N. Murthy

TL;DR: 该研究用 LLM 从复杂天然气味混合物中生成感官描述，并以此监督从分子组成预测混合物感知相似度及学习低秩分子表示。

Details

Motivation: 气味感知和化学组成的关系尚不清楚；许多研究只考察单分子，而天然气味可能由数十至数百种挥发性组分构成，组合效应增加了表征难度。

Result: 研究数据库包含 1,200 多种食物挥发物混合物、8,000 多种分子组分。相较其他感官提示，化学感官提示派生的感知目标最利于预测留出混合物相似度。混合物级训练得到的低秩分子嵌入优于组分重叠指标，即使两种混合物没有共同分子；未用单分子标签训练，学得分子几何仍与独立人类气味描述空间显著对齐。外部分子特征有一定零样本对齐，在监督有限时可改善预测，但完整训练数据下额外收益很小。

Insight: 语言描述可扩展地提供混合物级感知监督，并让分子表示反映组合后的感知关系，而不仅是分子清单重叠。LLM 描述是人类感知的代理而非直接测量；摘要未给预测误差、数据来源偏差或混合物间实验验证。

Abstract: The relationship between odor chemistry and perception remains incompletely understood. Many existing studies focus on the perception of single molecules, whereas natural odors contain tens to hundreds of components, adding extra layer of complexity to percept formation. Here, we explore the relationship between molecular composition and perceived similarity in natural odor objects. We study this using a database of more than 1,200 volatile-compound-in-food mixtures containing more than 8,000 molecular components, and use large language models (LLMs) as a scalable surrogate for deriving mixture-level perceptual descriptions. We find that language-derived representations provide useful supervision for predicting held-out mixture similarities from molecular composition. The prediction was consistently strongest when perceptual targets were derived from chemosensory prompts than from other sensory prompts. Further, during the mixture-level training, we learned low-rank molecular embeddings that enabled better prediction of held-out mixture similarities than component-overlap metrics, even for mixture pairs with no shared molecules. Although no single-molecule labels were used during training, the learned molecular geometry showed significant alignment with the independent human odor-descriptor space. Finally, we explored the use of external molecular features in our mixture-similarity prediction task. These features showed modest zero-shot alignment with mixture-level perceptual similarity. They improved prediction under limited supervision but offered little additional predictive benefit with the full training set. Together, these results show that language-derived descriptions can provide scalable supervision for predicting the perceptual organization of complex natural mixtures, contributing to a quantitative understanding of how molecular composition gives rise to odor perception.

[349] Monitoring latent reasoning in looped language models cs.AI | cs.CLPDF

Pawel Olszowiec, Paulina Tomaszewska, Maciej Żelaszczyk, Michal Byra, Neo Christopher Chung

TL;DR: 该研究建立五种机制可解释性视角，分析四个家族的八种循环语言模型，比较提示信息只在循环状态中传递与每轮重新读入两类结构。

Details

Motivation: 循环语言模型通过反复使用同一层块增加每 token 的潜在计算而不增加存储，但其迭代变换是否真正改善能力、对提示扰动是否稳定仍缺少研究。

Result: 作者提出基于机制可解释性的五视角检查协议，并应用于 4 个家族共 8 个 looped LM。重点比较 carry loops（如 Ouro、LoopUS）仅在循环状态内传递提示，与 re-read loops（Parcae、Raven 家族的 Rec-OLMo、Huginn）每轮重新输入提示的差异。摘要称研究为循环模型的安全和高效使用提供观察，但未提供具体机制发现、扰动结果或性能数字。

Insight: 循环计算的架构细节会决定提示信息如何进入后续迭代，因此解释 looped LM 不能只比较参数量或循环次数。当前摘要没有列出五种分析视角的发现，不能据此归纳具体安全结论。

Abstract: Looped language models (LMs) reuse one block of layers several times per token generation, promising more latent computation at no additional storage cost. Despite their rising popularity, the analysis of looped LMs is still in its infancy: little is known about whether the iterative transformations actually improve their capabilities, or whether these models remain stable under prompt perturbation. We construct an inspection protocol rooted in Mechanistic Interpretability comprising five diverse lenses. We apply it to examine eight looped LMs from four families. Models differ in many aspects but the key one is whether the prompt is transmitted only within the looped state (carry loops: Ouro, LoopUS) or re-supplied at every iteration (re-read loops: Parcae, and Rec-OLMo & Huginn from the Raven family). The results of our study provide valuable insights concerning safety and efficient use of looped language models.

[350] Beyond Answer Voting: Fine-Grained Correction of LLM Reasoning through Graph Alignment across Sampled Traces cs.AI | cs.CLPDF

Haotian Liu, Yingjian Chen, Xiaohang Xu, Lang Huang, Yusuke Iwasawa

TL;DR: DDRG 将多个采样推理轨迹对齐成有来源的主张图，定位主张级分歧并交由不看最终答案的验证器检查，仅在证据支持时进行一次聚焦修复。

Details

Motivation: 多步推理中的中间错误会影响后续结论；自一致性只对最终答案投票，不能处理轨迹中的分歧，且扩充轨迹池会增加采样成本。

Result: DDRG 从冻结轨迹池建立来源支撑的主张图，对齐图结构以暴露分歧，并把局部冲突交给 answer-blind verifier。若验证器支持的反驳能接到多数答案的证据路径，就执行一次局部修复；否则返回原始自一致性投票。五个基准上，零样本宏平均准确率相较直接回答，Qwen2.5-7B 提高 8.65 个百分点、Llama3.1-8B 提高 15.83 个百分点；宏平均指标上也超过 CoT、所评 GoT/GoV 问答改编及自一致性。

Insight: 选择性核验把局部证据验证连接到修复决策，不必扩大初始候选轨迹池；若证据不支持修订则保留原投票，降低无依据改答风险。摘要未提供不同任务的单项效果、验证器误差或图对齐成本。

Abstract: Large language models increasingly tackle tasks that require multi-step reasoning, where errors in intermediate claims can affect the final answer. Self-consistency (SC) improves reasoning by voting over sampled final answers, but its vote leaves intermediate disagreements unresolved. To address this problem, we introduce DDRG (Disagreement-Driven Verification Guided by Reasoning Graphs), a framework for selective correction from a frozen bank of reasoning traces. Our insight is that graph alignment can expose claim-level disagreements and provide correction evidence without expanding the sampled trace bank. DDRG builds source-grounded claim graphs and sends localized conflicts to an answer-blind verifier. A verifier-supported refutation that reaches the majority answer’s supporting path triggers one focused repair; otherwise, DDRG returns the exact SC vote. Experimental results across five benchmarks show that native zero-shot DDRG improves macro-average accuracy over direct answering by 8.65 and 15.83 percentage points for Qwen2.5-7B and Llama3.1-8B. It also exceeds CoT, the evaluated GoT and GoV question-answering adaptations, and SC on this macro metric. DDRG links local verification to selective correction without enlarging the initial candidate bank.

[351] To Align or Not To Align: Check Your COMPASS Before You Train cs.AI | cs.CL | cs.LGPDF

Fabian Gröger, Petar Damjanović, Maria Brbić

TL;DR: COMPASS 在模型对齐或权重合并前，以 kNN 邻域重叠和训练范式可达上限预估模型对的兼容性，减少依赖昂贵下游试验筛选候选组合。

Details

Motivation: 视觉编码器—语言模型对齐和微调检查点合并常需经验搜索，模型对是否兼容往往要完成对齐或合并之后才知道。

Result: COMPASS 将兼容性拆为模型对特定的易度（两编码器 kNN 重叠）和范式级上限（训练范式允许的最大重叠），用两者对候选排序并给出策略：优先处理初始一致且范式兼容者，对兼容但初始差异大的组合加算力，范式上限不足时更换预训练范式，双项均差时降低优先级。在 72 项跨模态对齐实验和 392 项 task-arithmetic 合并中，兼容性估计与下游成功的平均 Pearson 相关，分别为 0.91 和 0.70。

Insight: 把候选模型初始局部邻域一致性与训练范式可达上限分开，可在投入昂贵适配前识别不同失败来源。相关性不等于兼容性排序必然准确，也未给阈值和失败率；跨模态相关高于合并任务。

Abstract: Combining pretrained models is a common strategy for building AI systems, whether by aligning a vision encoder with a language model or merging fine-tuned checkpoints. However, finding compatible pairs typically requires costly empirical search, as compatibility is known only after alignment or merging. We propose COMPASS (COMPatibility ASsessment Score), a theoretically motivated diagnostic that estimates whether a model pair will be compatible prior to alignment or merging. Building on the observation that representational convergence is most robust at the level of local neighborhoods, COMPASS decomposes model compatibility into two cheap-to-compute axes: a pair-specific ease, defined as the $k$-NN overlap between the two encoders, and a paradigm-level ceiling, capturing the maximum overlap their training paradigms allow. Together, these two axes provide a ranking of candidate model pairs and prescribe how to act on each pair: (i) prioritize pairs that already agree and whose paradigms are compatible, (ii) invest additional compute on pairs with compatible paradigms but low initial agreement, (iii) switch the pretraining paradigm when it cannot support sufficient agreement, or (iv) deprioritize pairs that score poorly on both axes. We apply COMPASS to $72$ cross-modal alignment experiments and $392$ task-arithmetic model merges and show that compatibility estimates by COMPASS strongly correlate with downstream success, achieving average Pearson correlation of $0.91$ on cross-modal alignment tasks and $0.70$ on model merging.

[352] Self-Evolving Agent Harness cs.AI | cs.CLPDF

Chuyang Wei, Haoxiang Guan, Minghang Zhu, Kexin Zhou, Jiyan He

TL;DR: 该综述将 agent harness 自我进化形式化为对固定模型外的软件脚手架持续修订，并分析其搜索、保留历史改进与独立验证的理论条件。

Details

Motivation: 智能体表现不仅取决于模型权重，还受提示、记忆、工作流、代码和插件组成影响；这些 harness 组件可跨运行修改，但持续修订可能造成回归和评测器变化下的失效。

Result: 框架把反馈驱动的修订及跨运行保留形式化，并区分更新目标与提案/准入机制，包括解释状态、工具/代码、仓库级控制逻辑和插件组合。作者从进化优化角度刻画搜索，以持续 PAC 学习分析可靠性；在明确假设下得到期望泛化界、以历史回归预算为条件的保留界，以及基于新鲜独立验证数据的高概率准入规则。文章还讨论知识共享、组织流程、科学可靠性、物理安全和问责，并提供 66 页综述、理论、算法与应用讨论。摘要未报告新的量化实验。

Insight: harness 的修改与模型权重更新是不同的自我改进通道，持久化收益需要持续验证和回归预算控制；部署中还需独立安全措施。理论结论建立在显式假设上，摘要中的案例来自既有工作，不能视为已通过统一实验验证。

Abstract: Harness self-evolution revises an agent’s software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. We formalize a feedback-driven process in which accepted revisions persist across runs. The framework separates update targets from proposal and admission mechanisms. It distinguishes changes to interpreted state from revisions to tools and code, repository-level control logic, and plugin composition. We characterize search through evolutionary optimization and examine its reliability through continual PAC-learning. Under explicit assumptions, the analysis yields an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh, independent validation data. We examine how persistent harness revisions affect knowledge sharing and organizational workflows, with implications for scientific reliability, physical safety, and institutional accountability. Cases from prior work illustrate model–harness co-design, distributed execution, and plugin reuse. Reliable deployment requires continual validation as task distributions and evaluators change, with independent safeguards for learning, runtime privileges, and consequential actions. We provide a comprehensive 66-page discussion of related work, theory, algorithms, evaluation, applications, and the broader impact of Harness RSI, together with continuously updated code and supplementary materials, at https://github.com/zgcagi/Harness-RSI.

[353] A Hybrid Kinematic GCN and Elastic Transformer for Efficient Indian Sign Language Recognition cs.AI | cs.CVPDF

Pushp Garg, Himangshu Sarma, Mrinmoy Ghorai

TL;DR: 该印度手语识别框架用熵驱动 SVD 关键帧采样降低冗余，再以连续双位置编码、弹性 Transformer 和运动学自适应 GCN 处理不规则节奏与双手几何。

Details

Motivation: 印度手语具有双手协同和速度变化，模型需处理不规则时间间隔；该领域大型词汇基准稀缺，普通架构可能无法保留自然手势节奏。

Result: 作者构建 120 类、9,453 条视频的本土 ISL 数据集。熵驱动 SVD 采样选取信息量较高关键帧；由于采样产生不规则物理时间间隔，模型用连续双重位置编码和距离感知弹性 Transformer 保留节奏；空间上以运动学条件自适应 GCN 建模双手骨骼—关节特征。加入 Physical Node Dropout 与 ArcFace 以缓解按主体记忆后，识别准确率达到 89.8%。摘要称结果具有竞争力，但未给对照基线分数。

Insight: 稀疏采样需要显式保留真实时间间隔，双手关系也需结构化建模；仅减少帧数而忽略时间会丢失手语节奏。89.8% 是新建数据集上的单一准确率，摘要未给切分方式、跨主体评估细节和基线比较。

Abstract: Sign Language Recognition (SLR) presents complex challenges due to variable signing speeds and spatial intricacies. Specifically, Indian Sign Language (ISL) requires precise modeling of its inherently bimanual nature and is bottlenecked by the scarcity of large-vocabulary benchmarks and the inability of standard models to handle irregular signing velocities. To address this, a structurally-aware hybrid framework for efficient, gloss ISL recognition is proposed. To minimize computational redundancy, an entropy-driven SVD sampler that dynamically extracts informative keyframes is used. Because this subsampling creates irregular physical time gaps, the proposed architecture employs a Continuous Dual Positional Encoding paired with a distance-aware Elastic Transformer to preserve natural signing rhythm. Spatially, a Kinematically-Conditioned Adaptive GCN explicitly models bimanual geometry using unified bone-joint features. To validate our framework, a 120-class, 9,453-video native ISL dataset is created. Augmented by Physical Node Dropout and ArcFace optimization to mitigate subject-specific memorization, our framework achieves a competitive 89.8% accuracy, establishing a strong baseline for complex sign language modeling.

[354] ScienceBuddy-PV: A Verifier-Governed Recursive Improvement Protocol for Panvascular Intervention Research cs.AI | cs.CVPDF

Lingsen You, Yujun Guo, Xinyu Zhong, Zijie Feng, Wentong Wang

TL;DR: ScienceBuddy-PV 将递归优化设计应用于泛血管介入研究，但当前仅审计一个固定的分割组件；智能体递归改进和强化学习尚未执行。

Details

Motivation: 该方案目标是让系统以证据账本、器械—血管本体、拒答规则和失败库支持介入研究，同时需要明确组件审计边界，避免把分割表现误称为完整智能体能力或临床建议。

Result: 组件测试使用 ARCADE 的语法与狭窄任务，各有 1,000 张训练、200 张验证、300 张测试图。共享任务掩码卷积模型在每种表示设置训练一次后固定参数，启发式支持门也固定，狭窄阈值在验证集选择。锁定测试集语法 Dice 为 0.475（95% CI 0.459–0.490），狭窄 Dice 为 0.231（0.214–0.249）；狭窄像素 AUROC 0.957、AUPRC 0.180，患病像素比例 0.00975。支持门接受 76.7% 狭窄帧输出（分数比较器为 100%）；构造掩码审计中接受 44.3% 参考候选、拒绝 55.7%，显示弃权代价。作者明确指出接受不代表正确或临床安全。

Insight: 论文将分割组件、覆盖/弃权行为与完整 agentic 自我改进区分，并明确没有证明患者级有效性或临床安全。狭窄像素类别极不平衡，AUPRC 低；较高 AUROC 不能单独说明分割可靠，且完整递归方案尚未执行。

Abstract: ScienceBuddy-PV adapts the ScienceBuddy recursive-in-recursive design to panvascular intervention research through a structured evidence ledger, device-vessel ontology, abstention contract and retained failure bank. Recursive harness optimization and outer-loop model learning, including reinforcement learning, are specified but not executed. We audit a fixed component, rather than the complete agent system, using separate ARCADE syntax and stenosis tasks (1,000 training, 200 validation and 300 test images per task). A shared task-masked convolutional model was trained once per representation setting; its parameters and the heuristic support gate were then held fixed, with the stenosis threshold selected on validation data. On the locked test sets, syntax Dice was 0.475 (95% CI 0.459-0.490) and stenosis Dice was 0.231 (95% CI 0.214-0.249). Stenosis pixel AUROC was 0.957 and AUPRC was 0.180 at prevalence 0.00975. The gate accepted 76.7% of stenosis-frame outputs versus 100% for a score-only comparator. In a constructed mask audit, it accepted 44.3% of reference candidates and rejected 55.7%, exposing the cost of abstention. Acceptance does not establish correctness or clinical safety. This public-data component audit makes support and abstention measurable and provides a reproducible protocol for future collaborative agent studies. It does not demonstrate recursive improvement, patient-level validity or clinical device recommendations.

[355] A Decade of PhotoMappers: A Longitudinal Study of Crowdsourced Disaster Photos Geolocalization with Vision-Language Models and Geospatial Reasoning cs.AI | cs.CVPDF

Hao Li, Wenping Yin, Fabian Deuser, Jia Jia, Ziqi Liu

TL;DR: 该研究结合十余年 PhotoMappers 众包灾害照片数据、街景和遥感图像，评估 VLM 地理定位能力及灾害信息档案的时空不均衡。

Details

Motivation: 众包灾害照片可提供及时地面观测，但其 GEOINT 价值取决于观测地理分布、解释质量和能否可靠定位；需研究 GeoAI 与人类地理判断各自优势。

Result: 研究用 30 项地理线索及推理证据指标评估 VLM，将众包照片与街景、遥感影像结合并用 Sinkhorn 重排序，R@1/R@5/R@10 分别为 40.90%、61.10%、69.45%；年度 R@1 在 25.00%–65.07% 间变化。地理推理方面，Qwen3.6-27B 总体一致率 67%，InternVL3-8B 为 51%；两者对局部、具体地点证据都比宽泛空间配置更难。档案地理和时间分布不均，受灾类型、社区生命线和业务重点变化影响。

Insight: 多源图像和重排能扩大候选定位，但复杂、歧义或后果重大的位置仍需人类判断；地理能力并非跨年份均匀。年度成绩波动说明档案分布影响显著，整体指标不能概括所有年份和地点。

Abstract: Crowdsourced disaster imagery provides timely, ground-level observations when conventional disaster mapping is delayed or incomplete, yet its geographic value depends on where observations are produced, how they are interpreted, and whether they can be reliably located as actionable Geospatial Intelligence (GEOINT). Using more than a decade of PhotoMappers, a crowdsourced disaster-mapping initiative, this study examines the changing geography of disaster information and asks how emerging GeoAI models compare with human geographic interpretation. We combine longitudinal spatial analysis with cross-view geolocalization using street view imagery (SVI) and remote sensing imagery (RSI), and evaluate vision-language models (VLMs) reasoning through 30 indicators spanning geographic cues and reasoning evidence. The results reveal a geographically and temporally uneven archive shaped by changing disaster types, Community Lifelines, and operational priorities. Integrating VGI, SVI, and RSI with Sinkhorn re-ranking achieves 40.90% R@1, 61.10% R@5, and 69.45% R@10, while yearly R@1 varies markedly from 25.00% to 65.07%. Geographic reasoning is similarly uneven: Qwen3.6-27B achieves an overall agreement of 67% compared with 51% for InternVL3-8B, with both models showing greater difficulty with localized and place-specific evidence than broad spatial configurations. We argue that the value of GeoAI lies less in replacing volunteered geographic interpretation than in complementing it, scaling candidate retrieval while retaining human judgment for ambiguous and consequential cases. This provides a pathway toward more human-centered and responsible GeoAI for disaster mapping. All data and code are openly available at https://github.com/bobleegogogo/PhotoMappers-VLM.

[356] XAI-Guided Fourier Domain Generalisation for Pure-Simulation Sim-to-Real Deep Reinforcement Learning cs.AI | cs.CVPDF

Wei Yao, Chuyao Wang, Ziwei Wang, Nabil Aouf

TL;DR: XFD 用模拟观测的可解释性归因定位策略依赖的视觉证据，再通过策略条件化傅里叶干预合成鲁棒视觉域，以改善纯模拟训练驾驶策略的零样本实车迁移。

Details

Motivation: 模拟训练的视觉驾驶策略部署到现实中常明显退化；图像迁移需大量目标域数据，随机域随机化又可能加入与真实差异无关的扰动。

Result: 框架将归因图转为策略条件化频域干预，扰动外观敏感频谱分量、同时尽量保留任务关键空间结构；鲁棒策略优化鼓励关注关键特征并抑制无关视觉变化。作者在模拟和零样本实车部署实验中，报告相较标准策略和随机域随机化基线，行驶距离与奖励均提升，尤其在背景杂乱的未见地图上更明显。归因分析显示，策略在模拟和真实观测间保留更一致的任务相关视觉证据。摘要未给具体数值。

Insight: 用策略归因指导域随机化可使视觉扰动针对模型脆弱处，而非均匀添加任意变化；零样本实车结果提供了迁移证据。摘要未披露实车场景数、距离/奖励增幅和安全事件，无法量化部署鲁棒性。

Abstract: Deep reinforcement learning has shown strong potential for vision-based autonomous driving, but policies trained in simulation often suffer severe performance degradation when deployed in the real world. Existing sim-to-real methods commonly address this gap through image-transfer-based domain adaptation or policy-agnostic domain randomisation. However, the former requires laborious acquisition of target-domain data and is often limited to the deployment domain, whereas the latter introduces unguided perturbations that may not reflect the underlying sim-to-real discrepancy, leading to suboptimal transfer performance. To overcome these limitations, we propose XFD, an explainable artificial intelligence (XAI)-guided Fourier domain generalisation framework for pure-simulation sim-to-real reinforcement learning. The proposed framework extracts explainability attribution maps from simulated observations, and converts these attributions into policy-conditioned Fourier-domain interventions. By perturbing appearance-sensitive spectral components while preserving task-defining spatial structures, XFD constructs robustness-enhancing visual domains entirely within simulation. Subsequently, the proposed robust policy optimisation guides the policy towards attending to task-critical features while suppressing sensitivity to irrelevant visual information. Extensive experiments in simulation and zero-shot real-world deployment demonstrate that XFD consistently improves traveled distance and reward over standard policy and randomly domain-randomised baseline, especially on unseen maps with cluttered backgrounds. Attribution-based analysis further proves that the proposed policy preserves more consistent task-relevant visual evidence across simulation and real-world observations.

[357] Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution cs.AI | cs.CV | cs.LGPDF

Jinwoo Chung, Jangho Kim

TL;DR: Token-Group Quantization 根据扫描前特征预测空间 token 分组，选择离线校准的组级激活裁剪范围，并用频域教师—学生损失保留超分辨率高频细节。

Details

Motivation: Mamba 超分辨率模型的 selective scan 会使不同空间 token 激活范围不同，单一层级裁剪范围不适配所有 token；但推理期再做扫描后动态分组不利于硬件部署。

Result: 方法用扫描后激活描述聚类生成伪标签，训练轻量预测器从扫描前特征推断分组，再查询离线校准的组级裁剪表，推理时无需采集运行统计、动态更新裁剪或扫描后再分组。Frequency-Preserving Refinement 对齐全精度教师与量化学生输出的傅里叶幅度和相位。在多种 PTQ 协议、位宽、骨干和任务上，作者报告持续超过 Mamba/SR PTQ 基线，并可迁移到 Transformer 超分和 Vision Mamba 分类、与旋转量化兼容；Jetson Orin Nano 上相较 FP16 加速 2.23 倍。摘要未提供具体质量分数。

Insight: 静态预测分组让 token 自适应裁剪范围仍可在硬件友好推理中实现；频域目标专门补偿低比特对高频结构的损伤。2.23 倍速度仅在 Jetson Orin Nano 对 FP16 测量，摘要未给画质折损和不同量化位宽下的详细权衡。

Abstract: Mamba-based super-resolution models are efficient alternatives to Transformer-based restoration backbones, but their low-bit post-training quantization remains challenging because selective scan, Mamba’s input-dependent state-space operator, induces different activation ranges across spatial tokens. Thus, a single layer-wise clipping range poorly matches all tokens. We propose Token-Group Quantization, a hardware-friendly activation PTQ framework that statically approximates a hardware-unfriendly post-scan grouping oracle, where tokens with similar preferred clipping ranges are assigned to the same group. To avoid runtime post-scan grouping in static low-bit inference, a lightweight predictor infers the groups from pre-scan features available before the quantized scan path. During calibration, we form pseudo-labels by clustering post-scan activation descriptors that capture similar rounding–clipping behavior. At inference, the frozen predictor maps pre-scan features to groups and selects entries from offline-calibrated group-wise clipping-bound tables, without runtime statistics collection, dynamic clipping updates, or post-scan group assignment. To address low-bit restoration degradation, we further introduce a Frequency-Preserving Refinement (FPR) objective for clipping-bound calibration. FPR aligns Fourier amplitude and phase between full-precision teacher and quantized student outputs to preserve high-frequency structures. Across PTQ protocols, bit-widths, backbones, and tasks, our method consistently outperforms prior Mamba and SR PTQ baselines, transfers to Transformer SR and Vision Mamba classification, composes with rotation-based quantization, and achieves 2.23x speedup on Jetson Orin Nano over FP16. Code is available here.

[358] VL-DocIR: A Benchmark for Vision-Based Long Document Information Retrieval cs.AI | cs.CVPDF

Simon Ott, Yufan Chen, Ruiping Liu, Junwei Zheng, Jiale Wei

TL;DR: VL-DocIR 构建页面级长文档视觉检索基准，覆盖跨页和跨文档证据，并提供结构平衡指标以诊断单页检索与多页证据整合差距。

Details

Motivation: 现有视觉文档检索评估偏向短文档和单个相关页，难以反映真实查询需跨多页甚至多文档整合证据的情况。

Result: VL-DocIR 从 29,641 份 HTML 源文档渲染出 388,548 页图像，来源包括 Wikipedia、arXiv、PubMed 和 SEC 代理声明。基准包含 271,760 个查询、23 个领域、六种证据结构，覆盖单页、同文档多页和跨文档。证据标注锚定页面图像及 HTML 元素，查询经过清晰度、正确性、封闭性和避免显式布局线索等清洗。评测单/多向量检索器及文档上下文打分、查询拆分、重排、轮转集成。最强配置在结构平衡 Macro 评估达到 nDCGAll@10 69.3、RecallAll@10 80.6；跨文档检索仍比单页难得多，文档上下文分数尤其有助于同文档多页查询。

Insight: 不仅要找到相关页，还要完整检索回答所需证据结构；按结构平衡评估可防止单页样例掩盖跨文档困难。强模型结果为多组件配置，摘要未拆分各模块增益和检索成本。

Abstract: Vision-based document retrieval has improved rapidly, but current evaluation settings still emphasize short documents and single relevant pages. This leaves an important gap between leaderboard performance and realistic long-document retrieval, where the evidence needed to answer a query may span multiple pages or even multiple documents. We present VL-DocIR, a page-level benchmark for evidence-complete visual long-document retrieval. VL-DocIR is built from 29,641 HTML-born documents rendered into 388,548 page images from Wikipedia, arXiv, PubMed, and SEC proxy statements. It contains 271,760 queries across 23 domains and six evidence structures, covering single-page, within-document multi-page, and cross-document evidence configurations. Evidence annotations are grounded to rendered pages and HTML element identifiers, and queries are filtered with a cleaning pipeline targeting clarity, correctness, closedness, and the absence of explicit layout references. We evaluate leading single-vector and multi-vector visual retrievers together with diagnostic retrieval settings such as document-context scoring, query splitting, reranking, and round-robin ensembles. The strongest evaluated configuration reaches 69.3 nDCGAll@10 and 80.6 RecallAll@10 under structure-balanced Macro evaluation, but source- and evidence-structure breakdowns show that cross-document retrieval remains substantially harder than single-page retrieval, and that document-context scoring is especially helpful for within-document multi-page queries. VL-DocIR exposes these failure modes and provides a more demanding target for visual long-document retrieval. The new dataset and code will be made publicly available at https://yufanchen96.github.io/VL-DocIR.

[359] ChemExplain: A Multimodal Benchmark for Explanatory Chemical Reasoning cs.AI | cs.CVPDF

Aotian Chen, Frazier N. Baker, Reza Averly, Shan Zhang, Tianao Chen

TL;DR: ChemExplain 用 1,180 道多模态化学题评估模型能否从分子/反应图中找到最能解释实验效应的结构特征，而不仅是预测结果或生成反应路径。

Details

Motivation: 现有化学基准多测结果预测或反应路径生成，较少检验模型能否解释观察到的化学效应为何发生及归因到图示结构的哪个特征。

Result: 基准问题取自 383 篇同行评审论文，每题包含分子/反应图、观察到的化学效应及四种分别关联不同可见结构特征的解释。正确的结构—机制关系来自源论文证据而非问题生成 LLM 自行编造；多智能体流程分三阶段质控，并由八位化学专家核验。最强评测模型准确率 78.6%；600 题专家复核子集的推理轨迹分析中，最强模型错误有 71.4% 被归为归因失败，所分析开源及化学专用模型也以此为最大错误类别。

Insight: 化学图像问答不仅需要识别结构和调用知识，还需把解释归因到直接相关的特征；该错误是当前模型的主要瓶颈。专家审核加强题目可信度，但78.6%成绩来自特定多选基准，不能代表开放式科学解释质量。

Abstract: Current chemistry benchmarks primarily evaluate outcome prediction or reaction pathway generation, but rarely test whether a model can explain why an observed chemical effect occurs. We introduce ChemExplain, a benchmark of 1,180 multiple-choice chemistry questions drawn from 383 peer-reviewed papers. Each question pairs a molecular or reaction image with an observed chemical effect and four candidate explanations, each associated with a different structural feature visible in the image. Answering requires the model to perceive the depicted structure, apply chemical knowledge, and identify which feature most directly explains the effect, a capability we term explanatory attribution. The correct feature-mechanism relation is derived from supporting evidence in the source paper rather than invented by the question-generation LLM. We construct the benchmark automatically with a multi-agent framework that applies progressive quality control at three stages and validate it with eight chemistry domain experts. The strongest evaluated model reaches 78.6% accuracy, leaving considerable room for improvement. On a 600-question expert-reviewed subset, reasoning-trace analysis classifies 71.4% of the strongest model’s errors as attribution failures. Attribution is also the largest error category for the open-source and chemistry-specific models included in our analysis.

[360] IsoVRBench: Video Reward Modeling Without Human Feedback cs.AI | cs.CVPDF

Shuyue Stella Li, Jacqueline He, Jiahui Chen, Sreya Dutta Roy, Yulia Tsvetkov

TL;DR: IsoVRBench 通过一次只改变一个视频属性的受控干预构造偏好对，量化视频奖励模型分别识别提示一致性、视觉质量、动作时间顺序等属性的能力。

Details

Motivation: 常规视频偏好对常同时在提示匹配、视觉保真、运动和生成器身份等方面不同，因此总体准确率无法说明奖励模型实际能区分哪些属性。

Result: 基准包含 8,000 对、覆盖八类单属性干预的合成偏好对，无需人工标注。最强奖励模型能分辨生成器间的宽泛质量差异，却容易忽略时间顺序：帧重排任务仅 60.8%，人类为 100%；既有基准排名未预测此失败。用隔离属性数据训练后，IsoVRM 在 VBench-2.0 上比使用更大规模人工偏好数据训练的对应模型高 3.8 分；用其后训练的视频生成器 IsoWan，在人类偏好上胜过人工数据对照，和人工偏好数据结合还有叠加收益。

Insight: 单属性控制对有助于诊断奖励模型具体盲点，而不是只给一个聚合偏好分；自动合成数据也易随生成器演化扩充。帧顺序 60.8% 与人工 100% 的差距表明高总体分会掩盖时序失效；摘要未说明人工评估规模。

Abstract: Video reward models are trained and evaluated on human preferences between videos that often differ simultaneously on properties such as prompt alignment, visual fidelity, motion, and generator identity. As a result, aggregate accuracy does not reveal which video properties a reward model can reliably distinguish. We construct preference pairs by controlled interventions that change one property of a generated video while holding the rest fixed. The resulting pairs, by construction, provide preferences for how the videos differ along a specific dimension, without requiring any human annotation. We propose IsoVRBench, a video generation reward modeling benchmark of 8,000 such pairs across eight interventions, and show that state-of-the-art models separate broad quality differences between generators but miss important features such as temporal order. The best reward model reaches only 60.8% on reordered frames against 100% for humans, and rankings on existing benchmarks do not predict this failure. Using the synthetic isolated intervation pairs as training data leads to 1) IsoVRM, a reward model outperforming its counterpart trained with a larger human labeled preference dataset by 3.8 points on VBench-2.0, 2) IsoWan, the video generator post-trained with IsoVRM, outperforming its human-data counterpart in human preferences, and 3) gains that stack when combined with human data. Because intervention pairs are automatically generated rather than collected, the benchmark and training data can be easily extended as generators evolve.

Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Quanbin Wang, Jiangwei Lao

TL;DR: MedEvo 让学生保留失败推理轨迹中的有效前缀，只修复首个关键错误，并根据独立反馈集中的医学能力缺口优先选择新训练问题。

Details

Motivation: 多模态医学推理常蒸馏教师完整推理，丢弃学生失败轨迹中可保留的有效推理；通用纠错方法也难覆盖图像读取、解剖上下文和鉴别诊断等医学特定错误。

Result: gold-blind 审核器定位每条失败 rollout 的首个关键错误并提供最小、不给答案的提示，学生保留正确前缀并重生成错误后缀；每轮再按医学标签统计留出反馈集的能力缺口，优先生成相关问题并修复。在仅监督微调设置下，Qwen3.5-9B 七基准平均准确率从 65.57% 提至 69.12%，MedXpertQA-MM 从 47.40% 提至 54.50%，超过教师蒸馏 SFT 的 66.93% 平均分。Qwen3.5-35B 从 67.91% 提至 73.04%，397B 从 71.83% 提至 74.95%。分析显示失败轨迹含可保留推理，修复成功不等于学到可迁移能力，独立反馈对持续改进关键。

Insight: 错误后缀修复而非整条轨迹替换可保留学生已有正确推理，并把训练数据选择与能力缺口相连。摘要区分修复成功与实际学习价值，提醒需独立反馈评估；结果是基准准确率，不能视为临床诊断验证。

Abstract: Medical multimodal reasoning models are typically trained on teacher-generated rationales, discarding the valid reasoning in students’ own failed trajectories, while general-domain correction methods offer little guidance for errors in image reading, anatomical context, or differential diagnosis. We propose MedEvo, which uses medical knowledge, namely a tag taxonomy and error-type diagnosis, to navigate both trajectory repair and data selection. For each failed rollout, a gold-blind auditor locates the first consequential error and provides a minimal, answer-free hint; the same student retains the valid prefix and regenerates the erroneous suffix. After each round, MedEvo identifies capability gaps across medical tags on a held-out feedback set and prioritizes related questions for new rollouts and repairs. With supervised fine-tuning (SFT) alone, MedEvo raises Qwen3.5-9B mean accuracy across seven benchmarks from 65.57% to 69.12% (47.40% to 54.50% on the multimodal track of MedXpertQA (MedXpertQA-MM)), surpassing teacher-distillation SFT (66.93%). It also raises Qwen3.5-35B and Qwen3.5-397B mean accuracy from 67.91% to 73.04% and from 71.83% to 74.95%, respectively. Our analyses show that failed trajectories contain reasoning worth preserving, that repair success does not imply learning value, and that independent feedback is key to sustained improvement.

[362] Thinker-WM: A Unified World Model with Shared Multimodal Representations for Embodied Intelligence cs.AI | cs.CV | cs.ROPDF

Thinker Team, Peng Gao, Jichao Jiao, Chen Liao, Daqin Luo

TL;DR: Thinker-WM 用共享 Unified Dynamics Transformer 联合表示并生成 RGB、度量深度和机器人动作，使视频预测、动作生成和策略推断等能力共用动力学表示。

Details

Motivation: 现有世界模型常将视觉动态、几何结构和可执行动作放在不同接口或下游模块中，难以在同一动力学表示内共同演化。

Result: 模型使用模态专属编码器/解码器，但在共享中间表示中联合演化外观、几何和动作；同一参数化支持联合视频—动作生成、世界预测、策略推断、动作条件生成、视频到动作和逆动力学。作者用轻量运动探针显示冻结视觉表示含可恢复运动信息。训练数据约 2 万小时第一人称视频、1 万小时真实机器人数据和 1,000 小时模拟数据，其中约 4,000 小时机器人数据带对齐深度。作者报告在 embodied world-model 评测表现强，并在 EWMBench、WorldOlympiad、WorldArena 排名第一；控制实验支持动作条件、几何交互及多功能训练的收益。

Insight: 共享中间动力学表示旨在把视觉、几何和动作联系起来，避免每项能力各自建立接口。摘要未提供各榜单分数、模型规模和训练成本；“排名第一”限于所比较系统及评测版本。

Abstract: World models have emerged as a promising paradigm for embodied AI, yet existing systems often remain fragmented in the modalities they model and the functions they support. In particular, visual dynamics, geometric structure, and executable action are commonly coupled through separate interfaces or downstream modules rather than being jointly represented and evolved within a shared dynamics model. In this work, we present Thinker-WM, a unified multimodal world model that jointly models and generates RGB, metric depth, and robot action through a shared intermediate Unified Dynamics Transformer with modality-specific encoders and decoders. The model performs multimodal unification at the representation level: appearance, geometry, and action are transformed into a common dynamic representation and jointly evolved by shared parameters, while their distinct output structures are preserved. The same parameterization supports multiple functions, including joint video-action generation, world prediction, policy inference, action-conditioned generation, video-to-action generation, and inverse dynamics. To examine whether the shared representation captures action-related information, we introduce a lightweight motion probe that shows the frozen visual representation contains recoverable motion information. Thinker-WM is trained on a heterogeneous corpus comprising approximately 20K hours of egocentric video, 10K hours of real-robot data, and 1K hours of simulation data, including approximately 4K hours of real-robot data with aligned depth supervision. Extensive experiments demonstrate strong performance across embodied world-model evaluations. Thinker-WM ranks first on EWMBench, WorldOlympiad and WorldArena, while controlled studies further demonstrate the benefits of action conditioning, geometric interaction, and multi-function training within the shared representation.

cs.CR [Back]

[363] Differential Privacy Protected Preference Alignment on Survey Derived Data cs.CR | cs.CL | cs.LGPDF

Xiuyuan Qi, Yi Wang, Wenjie Wang

TL;DR: 该研究针对世界价值观调查问卷中敏感偏好数据，研究用标签级差分隐私保护语言模型偏好对齐，并比较随机响应、指数机制与 DP-SGD。

Details

Motivation: 偏好对齐数据可能包含政治信念、社会规范、宗教、性别态度和文化价值等敏感信息；问卷监督通常是序数标签而非成对偏好，现有隐私研究较少针对这类数据设计机制。

Result: 作者先展示已对齐模型的隐私脆弱性，将任务表述为评分预测并引入标签差分隐私，构建结合随机响应、指数机制和 DP-SGD 的隐私流程。实验报告简单的标签级机制优于梯度级方法，预处理时选择适当聚合函数也有较大影响。摘要未给隐私预算 ε、效用分数或攻击成功率。

Insight: 隐私保护应匹配问卷序数标签结构，直接扰动标签可能比仅依赖训练时梯度保护更有效；聚合方式也会影响结果。未报告隐私预算和效用数字，暂不能比较严格隐私保证下的对齐效果。

Abstract: Preference alignment enables language models to better reflect human values, but the preference data contains sensitive information, especially when derived from surveys about political beliefs, social norms, religion, gender attitudes, and cultural values. This paper studies privacy-preserving alignment on questionnaire-derived data from the World Values Survey, where supervision takes the form of ordinal labels rather than standard pairwise rankings. Existing efforts mainly focus on general privacy protections upon model training, while data-specific methods for survey data remain unexplored. We first show the vulnerability of aligned models, then formulate the task as rating prediction and introduce label differential privacy. Building on this formulation, we construct a differential private pipeline utilizing three privacy mechanisms: randomized response, exponential mechanism, and DP-SGD. Experiments show that simple label-level mechanisms outperform gradient-level methods, and proper aggregation function for preprocessing also make large contribution. The new fine-grained private comparison motivates explorations toward language model alignment with formal privacy guarantees.

[364] DPOison-VL: Clean Preference Poisoning in Vision-Language Direct Preference Optimization cs.CR | cs.AI | cs.CVPDF

Chengtao Jian, Kai Yang

TL;DR: DPOison-VL 通过少量训练图像上的不可见扰动，在提示、回答和偏好标签语义不变时操纵 VLM 的 DPO 对齐方向。

Details

Motivation: 现有偏好投毒常修改显式监督信号或使用可见触发器，容易被人工检查发现；多模态偏好图像中的隐蔽扰动风险研究不足。

Result: 作者将干净偏好投毒表述为双层优化，以梯度对齐代理引导图像扰动朝攻击者指定偏好，并提出 preference displacement score 在固定投毒预算下选择样本；还推导偏好反转的充分条件。在多类攻击任务、不同 VLM 架构和规模上，最多 2% 投毒数据可达到最高 98% ASR，并具有跨模型迁移性，同时大体保持干净准确率和通用效用。

Insight: 即使偏好文本和标签看似无害，图像侧不可见扰动也可能改变 DPO 学到的偏好，说明多模态偏好数据需接受图像完整性审计。最高 98% ASR 是特定攻击设置的最好结果，摘要未给不同攻击/防御的完整分布。

Abstract: Direct Preference Optimization (DPO) has become a widely adopted approach for aligning vision-language models (VLMs), yet the security of multimodal preference data remains largely unexplored. Existing preference poisoning attacks either corrupt explicit supervision signals or require visible triggers, making them detectable under human inspection. We introduce DPOison-VL, the first clean preference poisoning attack against VLM-DPO, which manipulates model alignment solely through imperceptible perturbations on a small fraction of training images, while leaving all prompts, responses, and preference annotations semantically intact. We formulate poison construction as a bilevel optimization and derive a gradient alignment surrogate that efficiently steers image perturbations toward the attacker-specified preference. We further introduce a preference displacement score for principled poison selection with provable first-order optimality under a fixed poison budget, and derive a sufficient condition for preference inversion. Extensive experiments across diverse attack tasks, different VLM architectures and scales, and baseline comparisons demonstrate that DPOison-VL achieves up to 98% ASR with only 2% poisoned data and exhibits cross-model transferability, while largely preserving clean accuracy and general utility. Our findings reveal a critical vulnerability in VLM-DPO, where imperceptible image perturbations in preference training data can silently redirect alignment.

cs.IR [Back]

[365] Personalized LLM Recommendation with Semantic IDs: Autoregressive Beam Search vs. Multi-Token Prediction cs.IR | cs.AI | cs.CL | cs.LGPDF

Sejoon Oh, Fernando Amat Gil, Evan Smothers, Dan Iter, Mark Thornburg

TL;DR: 该行业规模研究比较语义 ID 推荐中的单头自回归 beam search 与多 token 预测，发现前者多数产品指标更好、训练成本更低，后者仅推理吞吐领先。

Details

Motivation: 大规模个性化推荐正从 embedding 表检索转向生成短序列语义 ID，使单个指令模型可统一检索、内容理解和语言交互；关键问题是如何生成语义 ID。

Result: 研究在同时处理个性化检索、内容问答和指令遵循的流媒体目录系统中，对自回归单头解码和多 token prediction（MTP）作同规模直接比较，并为 MTP 增加联合 SID 分数排序损失。单头 beam search 在排序和内容 QA 等大多数关键指标上持续胜过 MTP，Qwen3-8B 规模训练美元成本约低 2.6 倍；MTP 仅在原始推理吞吐上占优。相较带辅助排序头的旧非 SID 基线，SID 系统冷启动排序提高 77%、内容 QA 准确率提高 22%、训练快 2.8 倍；语言指标下降归因于扩大的 SID 词表，作者用强化学习后训练处理该回退。

Insight: 解码速度高不等于产品端效果更好；在排序和内容理解需求下，beam search 的质量及训练成本更有优势。性能比较来自特定流媒体目录和多任务栈，摘要未给线上业务指标或 MTP 与 beam search 的具体吞吐差值。

Abstract: Large-scale personalization is shifting from embedding-table retrieval to LLM-based recommenders that generate the Semantic ID (SID) of the item to recommend, where each SID is a short sequence of discrete hierarchical content codes. A single instruction-tuned model can then unify retrieval, content understanding, and language, and recommend brand-new titles from their content alone using SID; then, how should the model generate the SID? Traditionally, models rely on a single language modeling head that generates tokens autoregressively, typically using constrained beam search. In contrast, an emerging body of work explores multi-token prediction (MTP), where dedicated heads predict all tokens in parallel, substantially reducing the decoding latency associated with beam search. We report an industry-scale, apples-to-apples comparison of these two generation paradigms: single-head autoregressive decoding vs. MTP. We introduce a new auxiliary ranking loss, defined over the joint SID score, to train the MTP objective. We evaluate the SID training objectives on both quality and training/inference cost, inside a multi-task LLM that jointly does personalized retrieval, content question-answering (QA), and instruction following over a large streaming-media catalog. Our experiments show single-head beam search consistently beats MTP on most product-critical metrics: ranking and content QA, and trains at roughly 2.6x lower dollar cost on Qwen3-8B scale; MTP wins only on raw inference throughput. Around this finding, we detail an efficient training and inference stack and report end-to-end gains over the previous non-SID baseline equipped with an auxiliary ranking head: +77% item cold-start ranking, +22% content-QA accuracy, +2.8x faster training, with language metric regression traced to SID vocabulary expansion. We address the language regression by reinforcement learning post training.

[366] RALI@iKAT 2026: Retrieval-aware User Profile Denoising via Personalized Conversational Dense Retrieval cs.IR | cs.CLPDF

Yuchen Hui, Fengran Mo, Jia Ao Sun, Yuxing Tian, Wenhao Xu

TL;DR: PCDR 将对话历史和用户画像联合映射到面向检索的稠密空间，以检索相关性标注训练用户画像去噪，避免先将画像短语拼进查询的稀疏重写流程。

Details

Motivation: 长期对话形成的用户画像会积累噪声。传统个性化查询重写依赖 LLM 选择画像短语并扩展原查询，容易把去噪简化为显式短语筛选，和下游文档相关性建模脱节。

Result: Personalized Conversational Dense Retrieval 继承对话稠密检索器的上下文去噪能力，联合投影对话与画像到专用检索嵌入空间，训练只需少量个性化相关性标注。在 iKAT 2026 任务中，PCDR 作为会话式 AI 助手的核心检索组件进行评估。摘要未报告排名、检索指标或与基线的对比结果。

Insight: 直接以检索相关性学习画像表示，可在潜在空间去噪，而不是依赖查询中逐词选择用户信息。当前摘要只说明它参加 iKAT 2026 任务、未给结果数值，无法判断其实际检索效果或少量标注的具体规模。

Abstract: Modern conversational information access systems continuously extract user information from conversations to build user profiles for personalized retrieval. As user profiles accumulate over time, they inevitably become noisy, making effective user profile denoising a fundamental challenge. Existing personalized conversational query reformulation methods rely on the black-box reasoning capability of LLMs to expand the original query with relevant user profile phrases, producing a verbose query for retrieval. Such rewrite-then-retrieve pipelines reduce user profile denoising to the explicit selection of profile phrases in the sparse token space, separating personalization from downstream relevance modeling. Consequently, they often fail to identify user information that is truly beneficial for retrieval. To enable retrieval-aware user profile denoising, we propose Personalized Conversational Dense Retrieval (PCDR), which learns to jointly project the conversation history and user profile into a latent embedding space dedicated to retrieval. By inheriting the context denoising capability from conversational dense retrievers, PCDR requires only a handful of personalized relevance annotations for training. We evaluate PCDR on the iKAT 2026 tasks, where it serves as the core retrieval component of a conversational AI agent assisting users with complex information needs.

cs.LG [Back]

[367] SemTree: Judge-Aware Semantic Decision Trees for Inspectable and Repairable Rule Decisions cs.LG | cs.AI | cs.CLPDF

Jincheng Gao, Wei Ding, Shuo Sun

TL;DR: SemTree 将规则判断表示为带 Yes/No/Unknown 判据的语义决策树，由可替换语义裁判逐节点执行，使错误可定位、可单点修订并重新运行。

Details

Motivation: 内容审核、广告合规和资格判定往往涉及多条件、豁免及信息缺失；若只返回最终标签，难以知道哪项条件判断错，也无法针对性修复。

Result: 树结构按实际裁判的回答编译。审核员可修正单节点并重跑；修订节点成为裁判训练标签；工具型 LLM 编辑器可依据执行结果调整问题和子树。在 ShARC 全部 8,276 个测试样例上，托管裁判执行决策树准确率 67.70%，平均每例问 2.99 个问题，与直接要求同一裁判输出标签的 68.23% 接近。为另一裁判编译的旧树改由该裁判执行时准确率仅 47.87%；基于该裁判对 12 个初始问题的回答重新编译，得分 64.39%。节点修复使 BANKING77 准确率从 82.04% 升到 86.16%；中国电商审核加入豁免节点后误报从 24 降至 5；训练编辑器使开发集准确率从 64.47% 升至 73.35%。

Insight: 决策树的可执行路径提供可解释审计轨迹，也允许错误在节点级修补；树结构还依赖实际语义裁判，因此跨裁判复用会明显退化。例子覆盖问答、意图分类和审核规则，编辑与复核成本以及未见规则的表现仍需考察。

Abstract: Rule decisions such as content moderation, advertising compliance, and eligibility checks combine several conditions, exemptions, and missing-information cases, and an error can be fixed only once the failing condition is known. We introduce SemTree, a framework that represents a rule decision as a semantic decision tree: each node is a natural-language question with explicit Yes/No/Unknown criteria, a pluggable semantic judge answers it, and the branching structure is compiled on the answers that this judge actually gives. The executed path is the reasoning behind every decision, so errors are repaired where they occur: a reviewer corrects a single node and replays the rule, corrected nodes become training labels for the judge, and a tool-using LLM editor revises questions and subtrees while observing execution results. On all 8,276 ShARC test cases, a tree executed by a hosted judge reaches 67.70% accuracy with 2.99 questions per case, on par with asking the same judge for the final label (68.23%); an earlier tree compiled for a different judge scores 47.87% when executed by this judge, while a new tree compiled on this judge’s answers to a 12-question initial pool already scores 64.39%. Node-level repair raises BANKING77 accuracy from 82.04% to 86.16%, an explicit exemption node cuts the false positives of a Chinese e-commerce moderation clause from 24 to 5, and a trained editor raises the development accuracy of edited trees from 64.47% to 73.35%.

[368] GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games cs.LG | cs.AI | cs.CLPDF

Weixian Waylon Li, Jiaxin Zhang, Han Zhou, Chengzu Li, Ke Wang

TL;DR: GameSmith 用模板约束自动合成并验证零和博弈，再通过自博弈蒸馏与 CRAB 优势估计扩展可训练的博弈族，以提升语言模型推理。

Details

Motivation: 基于两人零和游戏的可验证奖励已显示可迁移到一般推理，但既有训练依赖少量手写游戏，游戏供给及针对这些游戏设计的优势估计器限制了扩展。

Result: LLM 在奖励、观测和隐藏信息代码固定的模板内合成游戏；通过配对随机种子回放及独立重新实现规格验证，再扩展为编码、规则和规模变体。三种 SPIRAL 模板得到 29 个游戏家族中的 26 个新游戏、336 个验证变体。Qwen3-32B 教师自博弈蒸馏后，用适用于各变体的 CRAB 进行自博弈强化学习；CRAB 按决策相对其局面预期回报的收缩估计分配信用。Qwen3-4B-Base 的 13 基准平均 pass@1 从 45.5 提至蒸馏后 62.5，再经自博弈达到 64.9；在相同预算下高于仅用三种 SPIRAL 游戏。模型间趋势一致，迁移收益随游戏家族数量增加。

Insight: 可验证的程序化游戏生成扩充了训练任务，同时双重验证降低游戏实现错误；对所有变体定义的优势估计也支持更大语料自博弈。推理提升和家族数的关系来自特定游戏模板及基准，摘要未报告验证成本或游戏多样性边界。

Abstract: Post-training large language models (LLMs) on two-player zero-sum games provides a verifiable reward, and the gains have been shown to transfer to general reasoning. Existing pipelines, however, rely on a handful of hand-written games and on advantage estimators tied to them, so both the supply of games and the advantage estimate on a large corpus limit scaling. We present GameSmith, which scales both. An LLM synthesises games inside templates whose reward, observation, and hidden-information code is fixed. Each game is verified by seed-paired replay and by an independent re-implementation of its specification, then expanded into encoding, rule, and size variants, giving 26 new games in 29 families and 336 verified variants from the three SPIRAL templates. The model is distilled from the self-play of a larger teacher (Qwen3-32B) on these games and then trained by self-play reinforcement learning with CRAB, an advantage estimator that credits each decision against a shrinkage estimate of the expected return of its position and is defined on every variant. On Qwen3-4B-Base, distillation raises the 13-benchmark mean pass@1 from 45.5 to 62.5, above the three SPIRAL games at the same budget, and self-play with CRAB reaches 64.9. The ordering holds across models, and transfer grows with the number of game families.

[369] Query-Adaptive Neuron Steering for LLMs cs.LG | cs.CLPDF

Yanxi Pan, Jingwen Wu, Yan Ma, Xiang Chen, Xinyi Zeng

TL;DR: Q-NeuS 根据输入问题动态选择并调节 LLM 内部神经元，使用 Neu-GRPO 比较不同干预配置的相对表现来训练轻量门控器，基座模型保持冻结。

Details

Motivation: 提示工程和固定内部 steering 都难以适应每个查询不同的知识与推理需求；先导研究发现神经元激活模式随主题和层变化，放大主题相关神经元可能改善表现。

Result: Q-NeuS 建立紧凑的主题引导候选神经元空间，并用 Query-Adaptive Gate 为每个输入生成逐神经元调制系数。Neu-GRPO 以多种干预配置的相对表现训练门控器；基座 LLM 在训练和推理阶段均冻结，推理无需任务或主题标签。作者在多个 LLM 和基准上报告性能持续改善、跨模型规模适用和跨基准迁移。摘要未列模型、基准、具体收益或干预额外时延。

Insight: 由查询激活决定干预位置和强度，较固定 steering 更有机会适配不同主题；冻结基座可降低模型更新成本。结论在摘要中主要定性，候选神经元构造的稳定性、门控开销和跨提示鲁棒性尚不清楚。

Abstract: Effectively eliciting the capabilities of large language models (LLMs) at inference time remains an important challenge. Existing studies address this challenge through context engineering, which shapes user queries, or fixed steering, which directly intervenes in model computation. However, predefined contexts and fixed intervention strategies may not adequately accommodate the diverse knowledge and reasoning requirements of individual queries. This raises a natural question: Can internal steering in LLMs be made query-adaptive? In this work, we investigate whether query-induced internal activations can guide adaptive steering in LLMs. Our pilot studies reveal subject- and layer-dependent neuron activation patterns and show that amplifying subject-relevant neurons can improve model performance. Based on these findings, we propose Q-NeuS, a query-adaptive neuron steering framework that dynamically determines which neurons to modulate and by how much for each input. Specifically, Q-NeuS constructs a compact subject-guided candidate neuron space and employs a lightweight Query-Adaptive Gate to generate neuron-wise modulation coefficients. To learn effective intervention strategies, we introduce Neu-GRPO, which adapts group-relative optimization to train the Query-Adaptive Gate using the relative performance of multiple neuron intervention configurations. The base LLM remains frozen throughout training and inference, and no task or subject labels are required at inference time. Extensive experiments across multiple LLMs and benchmarks demonstrate that Q-NeuS consistently improves model performance and exhibits broad applicability across model scales and strong cross-benchmark transferability.

[370] Reward-Aligned Distillation cs.LG | cs.CLPDF

Yilan Chen, Zian Su, Manish Prajapat, Dingmin Wang, Rein Houthooft

TL;DR: RAD 通过可微分的临时优化步骤估计教师每个 token 蒸馏信号对学生后续奖励的影响，再动态加强、抑制或反转相应梯度贡献。

Details

Motivation: 可验证奖励强化学习成本高且样本利用率有限；在线策略蒸馏能提供密集 token 监督，但常规方法无差别蒸馏所有 token，基于不确定性或师生分歧的筛选也未必能识别真正提高奖励的监督。

Result: Reward-Aligned Distillation 是双层框架：对学生执行临时优化步骤，估计每个 token 监督对更新后奖励的预测作用，并从初始权重出发做投影上升更新 token 权重。在匹配训练回答预算下，六个竞赛数学基准平均准确率相较最强基线，Qwen3-1.7B 非思考版提高 6.5 个百分点，Qwen3-4B-Instruct-2507 提高 2.5 个百分点；简化版 RAD-Simple 达到相近数学准确率。LiveCodeBench-v6 上，Qwen3-1.7B、OLMo3-7B-Instruct 的 pass@1 分别为 54.69%、58.67%，高于最强基线 8.7、6.4 个百分点。

Insight: 令蒸馏权重直接对齐学生更新后的任务奖励，可以选择性使用教师轨迹，而非把所有 token 都当成同等有益监督。双层元优化可能带来额外训练计算，摘要只匹配训练回答预算，并未说明完整算力或优化开销。

Abstract: Reinforcement learning with verifiable rewards improves the reasoning capabilities of large language models, but remains computationally expensive and sample inefficient. On-policy distillation (OPD) provides dense token-level supervision from a stronger teacher, yet standard OPD distills every token regardless of whether the teacher’s preference actually improves task reward. Existing token-selection methods rely on proxies such as uncertainty or teacher–student divergence, which need not identify reward-improving supervision. We introduce Reward-Aligned Distillation (RAD), a bilevel framework that weights distillation signals based on their predicted effect on the student’s post-update reward. RAD differentiates through a temporary optimizer step to score each token’s supervision, then updates the token weights by projected ascent from the initial token weights. The resulting weights can strengthen, suppress, or reverse individual distillation-gradient contributions. At a matched training-response budget, RAD improves mean accuracy on six competition-math benchmarks over the strongest baselines by 6.5 percentage points for Qwen3-1.7B (non-thinking) and 2.5 points for Qwen3-4B-Instruct-2507. RAD-Simple, a simplified variant, achieves comparable math accuracy at the same training-response budget. On LiveCodeBench-v6, RAD reaches 54.69% and 58.67% pass@1 with Qwen3-1.7B (non-thinking) and OLMo3-7B-Instruct, respectively, exceeding the strongest baselines by 8.7 and 6.4 points. These results support reward-guided weighting of teacher supervision across model families and tasks.

[371] HiPAD: A Unified Gait, Audio-Visual, and Personality Fusion Network for Depression Detection, with a Hierarchical Personality Ablation and the Calibration Limits of Domain-Adapted LLMs cs.LG | cs.CL | cs.CV | cs.SDPDF

Karol Chlasta, Piotr Struzik

TL;DR: HiPAD 用统一融合网络参加六个老年/青年及模态子赛道，发现该小样本抑郁检测挑战中约 150 token 的人格描述信号强于步态、音频和视频组合。

Details

Motivation: 多模态抑郁识别需比较步态、音视频和人格描述等来源的作用；不同赛道若更换架构，结果差异可能混合了数据与模型因素，难以归因。

Result: 网络以共享 BiLSTM 编码可用模态，缺失模态补零，仅在至少两种步态、音频、视频存在时进行跨模态注意；人格描述不进入注意力而采用拼接。固定架构覆盖六赛道。步态加人格的 TLS 为 0.647，超过所有含音视频配置；移除全部人格特征（含 Big Five）使 Track-Level Score 降低 0.499–0.507并接近随机。7B LoRA/72B QLoRA 人格编码器在 PHQ-8/9 领域适配使二元 Cohen κ 提升 0.13，却因 PHQ-9 回归失真降低总分；该样本规模下冻结 RoBERTa-large 更好。音视频赛道中手工临床特征优于通用深度特征。跨赛道 G-Score 为 0.482；每个队列少于 100 人。

Insight: 在该挑战的小样本设置下，人格文本主导预测，架构相同的赛道比较有助于揭示数据源贡献；微调可改善二元一致性却损害总体分数，说明校准和任务定义很关键。样本不足百人，模型信号可能混入人格与症状描述重叠等因素，结果不应视为临床因果或诊断证据。

Abstract: We present HiPAD, a single fusion network submitted to all six sub-tracks of the MPDD-AVG 2026 depression-detection challenge, covering the Elder and Young cohorts and three modality sets (gait, audio-visual, and both), each combined with personality descriptions. The network encodes each available modality with a shared BiLSTM, zero-pads absent modalities, and applies cross-modal attention only when at least two of gait, audio and video are present. Personality, which dominates our results, is kept out of the attention and fused by concatenation. Because the architecture is identical in every sub-track, differences between them reflect the data rather than the model, and we use this to study three questions. Our main finding is that the personality description, roughly 150 tokens about who the subject is, carries more depression signal than audio, video, and gait combined: gait plus personality alone (TLS 0.647) beats every configuration that includes audio and video. First, a four-level ablation of the personality description (L0-L3) quantifies this: removing every personality-derived feature, including the Big Five scores that earlier MPDD ablations keep, lowers the Track-Level Score by 0.499 to 0.507 and leaves the model at chance. Second, domain-adapting the personality encoder with LoRA (7B) or QLoRA (72B) on PHQ-8/9 data raises binary Cohen’s $κ$ by 0.13 but lowers the overall score by distorting the PHQ-9 regression; a frozen RoBERTa-large encoder does better at this sample size. Third, at fixed architecture, handcrafted clinical features (eGeMAPS, OpenFace action units) outperform generic deep features (wav2vec2, DenseNet) on the audio-visual tracks. The unified system reaches a cross-track G-Score of 0.482; with under 100 subjects per cohort, the limiting factors are the size of the representation and the calibration of the outputs, not the choice of architecture.

[372] Generalizable Model Editing Could Be More Than Direct Generalization cs.LG | cs.CLPDF

Xiaotian Ye, Chunyao Yang, Xiaohan Wang, Junkai Chen, Yongqi Duan

TL;DR: 该研究指出以输入—输出映射定义的直接模型编辑泛化存在可辨识性和实践偏差问题，并提出依靠现代 LLM 推理能力将更新事实组合到任务中的组合泛化视角。

Details

Motivation: 模型编辑通常要求事实更新直接改变复杂推理任务输出，但该类直接泛化长期被认为困难；作者在真实聊天场景中却观察到传统编辑方法也能表现出流畅的泛化，因此重新审视定义与障碍。

Result: 作者从理论和实验两方面分析直接泛化：理论上存在可辨识性限制；实践中尝试绕开限制的方法受构造偏差含义不清和优化动态等因素覆盖的影响。随后提出组合泛化，将事实更新与模型已有推理、组合能力相结合。多个模型和编辑方法的实验显示这一行为广泛存在。摘要未给具体基准、编辑成功率或比较数值。

Insight: 编辑后的事实不一定需要在每个复杂任务上直接训练映射，模型也可能通过原有推理能力把新事实组合进去；评测范式可能影响对泛化难度的判断。摘要对“流畅泛化”的操作定义和定量证据有限，不能据此断言组合泛化已解决长期知识更新问题。

Abstract: A central desideratum in model editing is generalization to complex tasks, which prior work formalizes under a direct generalization paradigm: treating and evaluating LLM behavior as an input-output mapping and expecting a factual update to alter outputs on reasoning tasks. Yet despite substantial empirical effort, the field consensus remains that such generalization is highly challenging and elusive. However, when we deploy the latest models in an in-the-wild chatbot-user setting, we surprisingly find that even classical editing methods exhibit remarkably fluent generalization behavior. This motivates us to revisit the obstacles to generalization and the limitations of the prevailing paradigm. We find that direct generalization is fundamentally constrained both theoretically and practically: theoretically, it suffers from an identifiability issue; and practically, existing approaches attempting to circumvent this are further hindered by the ambiguity of their constructed biases and by vulnerability to being overridden by other factors such as optimization dynamics. We instead identify compositional generalization, empowered by the increasingly strong reasoning and composition abilities of modern LLMs, as a theoretically grounded and empirically validated alternative. We validate this phenomenon across multiple models and editing methods, showing that such generalization behavior is widespread. Overall, we argue that generalizable model editing could be more than direct generalization, and call for a paradigm shift toward more generalizable and practically deployable knowledge update techniques.

[373] A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated Recipes cs.LG | cs.CLPDF

LoonSage Team, Bowei Song, Caiwei Zhao, Chuwen Liu, Dechuang Gu

TL;DR: LoongSage 是面向智能体强化学习的全异步框架，以单控制器编排白盒/黑盒智能体，并提供轨迹新鲜度控制、拓扑感知权重同步和低显存多教师管理。

Details

Motivation: 前沿 LLM 训练从静态推理转向工具调用、多轮互动和长程决策，现有 RL 框架日益复杂，增加接入、使用和调试负担。

Result: 框架核心约 14K 行，基于 Megatron 分布式训练、SGLang 高吞吐推理；单一控制进程协调，GPU 组件封装为 Ray actors。AgentFlow 统一白盒与黑盒智能体执行；滑窗策略限制异步轨迹陈旧程度，TransferMesh 按接收方拓扑选择最快权重传输方式。Teacher Manager 将教师模型扩展与 GPU 内存解耦，并支持多教师在线蒸馏。作者评估四次完整训练运行，涵盖复杂 RL 和多教师蒸馏，报告任务奖励持续增长、训练—推理差距受控，异步管线消除同步空闲并保持硬件利用率。摘要未给吞吐或利用率数值。

Insight: 该工作聚焦训练基础设施，设计目标是将异步执行、权重同步和教师扩展作为统一 agent-native 系统处理。实验覆盖四次完整运行但未披露具体训练任务增益、硬件利用率和相较同步框架的成本，因此效率优势目前在摘要中为定性结论。

Abstract: Reinforcement learning (RL) has emerged as a critical phase in the post-training pipeline of large language models (LLMs). As the frontier of LLM capabilities shifts from static reasoning to agentic tasks–characterized by external tool invocation, multi-turn interactions, and long-horizon decision-making–existing mainstream RL frameworks have grown increasingly unwieldy, raising the overhead of adoption, usage, and debugging. To address these challenges, we present LoongSage, an agent-native, fully asynchronous RL framework with a lightweight core of roughly 14K lines of code. Built upon Megatron for distributed training and SGLang for high-throughput inference, LoongSage centralizes orchestration within a single controller process and encapsulates all GPU-resident components as Ray actors. It introduces AgentFlow, a unified execution paradigm that orchestrates both white-box and black-box agents. For asynchronous execution, it integrates sliding-window strategies that bound trajectory staleness, and employs TransferMesh, a topology-aware synchronization mechanism that routes weights after each step over the fastest transport available to each receiver. Beyond agentic RL, LoongSage incorporates a memory-efficient Teacher Manager that decouples the scaling of teacher models from GPU memory constraints, natively supporting Multi-Teacher On-Policy Distillation. Extensive evaluations across four complete training runs, spanning complex reinforcement learning and multi-teacher distillation, demonstrate that task rewards improve steadily while the train-inference gap is kept under continuous control, and the fully asynchronous pipeline removes synchronous idle bubbles, sustaining hardware utilization.

[374] Attention-Graph Curvature for the Verification of Mathematical Reasoning cs.LG | cs.CLPDF

Kanghyeon Kim, Johannes F. Lutzeyer, Yanzhu Guo

TL;DR: 该研究将有向 Forman 曲率作为逐步数学推理验证器的注意力图特征，检验局部注意结构是否能补充步骤正确性信号。

Details

Motivation: 自然语言数学证明尚难交由证明助手自动验证，错误步骤会污染后续推理；已有方法利用推理模型前向特征做逐步错误检测，研究探索注意图局部几何能否提供额外线索。

Result: 作者计算注意力权重构成的有向图边上的 Forman 曲率，并在 ProcessBench 上比较四种逐步验证器加入曲率前后的表现。四种验证器总体平均表现均提升；将一步所有 token 作为输入的 ReProbe 增益最大。加入曲率使语言模型前向时间平均增加 18%。摘要未提供各验证器的准确率增益或误报变化。

Insight: 注意力图周围边结构的局部曲率可能携带语言表示之外的推理质量信息；收益在使用完整步骤 token 的验证器上最明显。额外前向成本约 18%，具体增益和是否适用于其他推理数据集需看正文结果。

Abstract: Language models have recently solved open Erdős problems and achieved gold-medal level performance on International Mathematical Olympiad problems. But their solutions, written in natural language, cannot yet be verified by proof assistants. An erroneous reasoning step can invalidate every step that comes after, so it is advantageous to detect errors early. Several methods have been proposed to detect erroneous reasoning quickly based on the features of the forward pass of the language model that performs reasoning. In this work, we study whether the directed Forman curvature of that model’s attention graph, a local quantity computed at an edge from the attention weights around it, carries information about a step’s correctness. To do so, we evaluate four per-step verifiers on ProcessBench with and without the curvature appended to their input. Appending curvature values improves the overall mean performance of all four verifiers. ReProbe, the verifier that takes every token of a step as input, shows the largest improvement. Adding the curvature increases the time of the language model’s forward pass by 18% on average.

[375] TPPO: Tracking-Value PPO For Language Model Reasoning cs.LG | cs.CLPDF

Futing Wang, Chenxi Li, Jianhao Yan, Yun Luo, Zhi Wang

TL;DR: TPPO 将逐位置标量输出累加为价值 logits，以贝叶斯对数几率更新刻画推理轨迹中的 value 演变，改善稀疏终局奖励下 PPO 的 token 级信用估计。

Details

Motivation: 带可验证奖励的语言模型 PPO 依赖 critic 在轨迹中提供价值演变，但稀疏终局奖励下标准 critic 往往无法准确跟踪各前缀价值。

Result: 作者分析相邻前缀价值关系，给出其在对数几率空间中的可加贝叶斯分解，并以此构建 Tracking-Value PPO；位置标量累加形成价值 logits，再用重构价值上的终局结果监督学习。数学推理汇总分相较 PPO，在 Qwen3-4B-Base、Qwen3-8B-Base 分别提高 7.0、8.4 分，并改善分布外泛化。增益还扩展到 Qwen3-4B、Qwen2.5-Math-1.5B、Qwen3.5-9B，以及倒计时和科学编码任务。分析显示价值演变跟踪更准、优势估计更可靠。

Insight: 以结构化累积参数化强制 critic 表达相邻前缀价值变化，而不是独立拟合每个状态，有助于稀疏奖励下的信用分配。摘要未提供不同基准汇总分定义、训练成本及各任务单项增益。

Abstract: Proximal Policy Optimization (PPO) is widely used in reinforcement learning with verifiable rewards for language-model reasoning. While \emph{value evolution along trajectory} characterizes relative token-level credit, yet standard critics often fail to track it under sparse terminal rewards. To model this evolution, we analyze the relation between successive prefix values and show that they admit a Bayesian decomposition with additive updates in log-odds space. We use this additive structure to construct Tracking-Value PPO (\method), a cumulative value parameterization that sums position-wise scalar outputs into value logits. These outputs are learned through terminal-outcome supervision on the reconstructed values. \method improves aggregate mathematical reasoning scores over PPO by 7.0 and 8.4 points on Qwen3-4B-Base and Qwen3-8B-Base, respectively, while improving out-of-distribution generalization. Gains extend to Qwen3-4B, Qwen2.5-Math-1.5B, Qwen3.5-9B and to countdown, and scientific coding tasks. Further analysis shows more accurate tracking of value evolution and more reliable advantage estimates. Our code is publicly available in https://github.com/LINs-lab/TPPO.

[376] Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments cs.LG | cs.AI | cs.CLPDF

Xiaoyang Yuan, Ji Zhao, Xingyu Hu, Wenhan Wang, Yutao Hu

TL;DR: Davinci-Tool 将可恢复环境状态、执行轨迹、任务结果和可执行验证器统一为监督数据底座，通过示范 SFT 和状态内动作偏好 DPO 扩展工具使用训练。

Details

Motivation: 可靠的智能体工具使用需把多步流程与可执行成功证据配对；只扩展轨迹覆盖或验证器都不够，因为新环境还包含独特状态、规则和依赖。

Result: 数据包含 30,395 条验证轨迹，覆盖 2,260 种调用工具。EGTS 从独立编写系统生成已验证示范；EES 合成可审计环境、状态锚定任务和可执行验证器，覆盖单体及协同领域。先监督微调，再复用 EES 状态与验证器比较备选行动并构造结果锚定偏好进行 DPO。Davinci-Tool 将 Qwen3-Coder-30B-A3B-Instruct 在 τ²-bench 从 52.4 提至 88.2、BFCL v4 multi-turn 从 30.4 提至 62.8、MCP-Atlas 从 16.9 提至 56.2；三项均超过参数总量为其 16 倍的 Qwen3-Coder-480B-A35B-Instruct。

Insight: 可恢复状态和可执行验证器使偏好标签能对应具体行动结果，并将环境多样性与成功判定一起扩展。基准分数提升显著，但摘要未说明跨环境训练与测试隔离、模型推理成本或三项任务的真实工具失败率。

Abstract: Reliable agentic tool use requires training data that pair diverse multi-step workflows with executable evidence that the requested outcomes were achieved. Scaling either trajectory coverage or verification alone is insufficient as new environments introduce distinct states, rules, and dependencies that must also be reflected in the success criteria. To address this challenge, we introduce Davinci-Tool, a 30B tool-use model, and Davinci-ToolData, an execution-grounded dataset designed to scale environment coverage and outcome verification together. Davinci-ToolData preserves restorable states, execution traces, outcomes, and verifiers as a unified substrate for supervision. Two complementary pipelines populate this substrate: Execution-Grounded Task Synthesis (EGTS) derives verified demonstrations from independently authored systems, and Executable Environment Synthesis (EES) synthesizes auditable environments, state-grounded tasks, and executable verifiers across single and coordinated domains. These pipelines yield 30,395 verified trajectories spanning 2,260 invoked tool names. We first train on these demonstrations with supervised fine-tuning, then reuse EES states and verifiers to compare alternative actions and construct outcome-grounded preferences for Direct Preference Optimization. Davinci-Tool improves Qwen3-Coder-30B-A3B-Instruct from 52.4 to 88.2 on $τ^2$-bench, from 30.4 to 62.8 on BFCL v4 multi-turn, and from 16.9 to 56.2 on MCP-Atlas, outperforming Qwen3-Coder-480B-A35B-Instruct on all three benchmarks with only 1/16 of its total parameters. Code is available at https://github.com/MassLab-SII/Davinci-Tool.

[377] The belief-action gap in language model agents cs.LG | cs.AI | cs.CLPDF

Ayse Sila Okcu, Weisheng Wang, Tongmeng Xie, Niall Dalton, Angelos Nalmpantis

TL;DR: 该研究区分语言模型智能体是否具备任务相关状态/价值信息与是否能将其落实为行动，发现内部信息往往比外显行为更可靠。

Details

Motivation: 智能体失败可能来自知识缺失，也可能来自决策时未能使用已有信息；若不区分“知道什么”和“如何选择行动”，就难以解释失败及设计干预。

Result: 作者在文本网格世界中研究状态信念、行动价值和偏好之间的关系，以报告式和表征式探针估计任务相关信念，并用逆强化学习估计行动价值。任务信息和价值常比行为中体现得更可靠；决策偏好通常在承诺点附近逐渐明确，报告的信念不确定性也在承诺前下降，但个别信念变化并不能稳定解释偏好变化。将其他决策的推理前缀移植到强制行动时，行动会随前缀改变；对推理窗口的激活编辑控制有限。作者强调这些测量区分信息可得性与信念—行动一致性，但本身不能识别每次失败原因。

Insight: 行为错误并不总意味着模型缺少相关信息；问题也可能发生在信息到行动的转换过程。当前证据来自文本网格任务与有限激活干预，尚未揭示普遍因果机制或适用于真实工具智能体的纠错方法。

Abstract: Language model (LM) agents can fail for different reasons, requiring different explanations and interventions. We study how state beliefs, action values, and action preferences interact as LM agents reason and commit to decisions in text-rendered grid worlds. We estimate task-relevant beliefs using report-based and representation-based probes, and action values using inverse reinforcement learning. We find that task-relevant state and value information is often represented more reliably than it is reflected in behaviour. Tracking action preferences throughout reasoning further shows that decisions sharpen around a commitment point and reported belief uncertainty decreases near commitment, although individual belief changes do not consistently explain shifts in action preference. Finally, forced actions follow reasoning prefixes from other decisions when transplanted, whereas the tested reasoning-window activation edits provide limited control over the action. Together, these results suggest that some failures arise not from missing task-relevant information, but from how available information is translated into action. Our measurements distinguish information availability from belief-action consistency without by themselves identifying the cause of each failure. Our code is available at: https://anonymous.4open.science/r/belief-action-gap-F745/

[378] RELAY: Selective Self-Distillation with Adaptive Hand-off for Rubric-Based Reinforcement Learning cs.LG | cs.AI | cs.CLPDF

Kunwang Zheng, Xinyu Sun, Yu Cao, Shuocheng Li, Lei Cheng

TL;DR: RELAY 在量规 RL 中先对正优势和低分零优势轨迹做选择性自蒸馏，再用平台检测决定何时移除冻结教师并仅继续 GRPO。

Details

Motivation: 量规 RL 将多个文字标准压成单个奖励，丢失细节且组内响应打平时 GRPO 无奖励梯度；冻结教师提供 token 级量规监督虽有帮助，但增益可能早期进入平台期。

Result: RELAY 用奖励/优势选择正优势及低分零优势 rollout，监测蒸馏损失并决定教师退场。第一阶段结合冻结教师选择性蒸馏与 GRPO，检测到平台后移除教师，第二阶段只用 GRPO。平台检测器在一个模型—领域配对上校准后原样用于另外三组，触发步数为 122–171。四种设置中，RELAY 的域内成绩最高；RaR-Medicine 上 Qwen3-1.7B 和 Qwen3-4B 分别比 GRPO 高 4.62、6.02 分。Qwen3-1.7B/Medicine 的共享检查点消融显示移除教师优于保留、削弱或刷新教师；蒸馏约一半响应 token 也优于在匹配平均系数下蒸馏全部 rollout。

Insight: 教师监督可在训练早期提供细粒度量规反馈，但持续蒸馏未必有利；阶段式退场和选择性 token/轨迹能提高效率。平台检测器跨三组复用有一定稳健性证据，但摘要未列四种设置完整成绩和检测阈值敏感性。

Abstract: Rubric-based reinforcement learning grades a response against many written criteria but collapses them into one reward, which discards the detail the rubric carries and gives no GRPO reward gradient when a group’s responses tie. A frozen teacher that reads the rubric as privileged context provides token-level supervision, but we observe an early plateau in the gains over GRPO. RELAY uses advantages and rewards to select positive-advantage and low-scoring zero-advantage rollouts, and monitors the distillation loss to decide when to retire the teacher. Stage 1 combines selective distillation from a frozen teacher with GRPO; a plateau detector then removes the teacher, and Stage 2 continues with GRPO alone. Calibrated on one model-domain pair and reused unchanged in three others, the plateau detector fires between steps 122 and 171. RELAY scores highest in-domain among the compared methods in all four settings, ahead of GRPO on RaR-Medicine by 4.62 points with Qwen3-1.7B and 6.02 with Qwen3-4B. On Qwen3-1.7B/Medicine, removing the teacher from a shared checkpoint outperforms keeping, weakening or refreshing it. Distilling selected rollouts covering about half the response tokens also outperforms distilling all rollouts at a matched average coefficient.

[379] DeltaTTG: Training Multi-Turn Agents with a Token-to-Go Progress Meter cs.LG | cs.AI | cs.CLPDF

Xuyan Ye, Xiyan Fu, Shengda Fan, Jiawei Cui, Yankai Lin

TL;DR: DeltaTTG 用共享进度计预测每个步骤到任务完成还需生成的 token 数，并据此重分配终局奖励优势，训练多轮智能体时不改动原始奖励。

Details

Motivation: 可验证奖励的多轮智能体 RL 通常只有稀疏终局奖励，难判断单步贡献；组内步骤比较依赖准确状态匹配和多样结果，直接评分又需外部评估器、人工标注或任务定制启发式。

Result: 进度计从多策略记录轨迹获得低成本监督，目标 Token-to-Go 是到任务完成仍需生成的 token 数，作为剩余完成工作量代理。actor 训练时冻结进度计，根据其估计变化将基于结果的优势在步骤间重新分配，但不改奖励，保留轨迹优势符号和 token 平均尺度。在 ALFWorld、PaperSearchQA、BFCLv3 上，Qwen3-4B、Llama-3.1-8B-Instruct 的平均分分别比 GRPO 高 5.7、7.2 点，整体超过所比较 RL 基线；actor 训练 GPU 小时接近 GRPO。作者还报告样本效率、域外表现改善，进度计可跨 actor 规模区分步骤质量并指导测试时动作选择。

Insight: 用可观测的剩余生成 token 数做代理进度信号，可在不改原奖励的情况下给长轨迹步骤更细信用分配。TTG 与真实工作量并非必然一致；摘要未提供进度预测误差、不同任务的 token 长度偏差及额外推理成本。

Abstract: Reinforcement learning with verifiable rewards (RLVR) for multi-turn agents relies on sparse terminal rewards that say little about individual steps. Within-group step comparisons depend on reliable state matching and sufficiently diverse sampled outcomes, while direct step scoring may require costly external evaluators, human annotations, or task-specific heuristics. We introduce DeltaTTG to learn a reusable progress signal through a shared progress meter, using low-cost supervision from diverse policies’ recorded trajectories without task-specific progress rules. Its supervision target, Token-to-Go (TTG), the number of tokens remaining to be generated until task completion, serves as an observable proxy for remaining completion effort. During actor training, changes in the frozen meter’s estimates redistribute outcome-derived advantages across steps without altering the reward, preserving each trajectory’s advantage sign and token-average scale. On ALFWorld, PaperSearchQA, and BFCLv3, DeltaTTG improves the average score over GRPO by 5.7 and 7.2 points with Qwen3-4B and Llama-3.1-8B-Instruct, respectively, outperforms all compared RL baselines on average, and keeps actor-training GPU-hours close to GRPO. Further experiments show improved sample efficiency and out-of-domain performance, and support the meter’s ability to distinguish step quality and guide test-time action selection across actor sizes. The code is available at https://github.com/Xuyan923r/DeltaTTG.

[380] Checking the Checkers: A Synthesis of Verifier Noise and Robustness in RLVR cs.LG | cs.AI | cs.CLPDF

Savyasachi Gupta, Aniruddha Srinivas Joshi

TL;DR: 该综述汇总 58 项研究中的 RLVR 验证器错误、训练影响与缓解方法，指出错误类型及策略能否利用错误，比单看验证器错误率更关键。

Details

Motivation: 可验证奖励强化学习依赖检查器判断回答正确，但规则检查器可能因格式错误拒绝正确答案，单元测试可能接受错误代码，LLM 裁判也可能被空或对抗回答诱导。

Result: 综述比较四类问题：验证器会犯什么错、如何影响训练、何种缓解对哪些错有效，以及发现是否适用于自生成、量规/裁判或无验证器奖励。证据显示，随机错误主要减慢学习，但在固定预算或 KL 正则下也会降低最终准确率；策略可学会触发的假阳性则可能导致平台或崩溃，即便初始错误率低。错误方向与被检查内容和构建者有关，不由验证器家族单独决定；策略变强时错误分布也会改变，现有缓解多只对一种错误测试。作者整理比较表、调和文献冲突并提出报告协议。

Insight: 验证器错误率本身不足以预测训练影响，关键在于错误是否可被策略系统性利用及错误结构。综述的结论受纳入研究的任务和验证器设计影响；其建议是更充分报告错误机制与训练动态，而非认为某种验证器天然可靠。

Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a principal recipe for improving the reasoning of large language models (LLMs). It rests on the assumption that the checker scoring each response is itself correct. In practice, rule-based checkers reject correct answers in unexpected formats, unit tests accept incorrect programs, and LLM judges can be led to approve empty or adversarial responses. This work surveys 58 studies that measure, analyze, or mitigate such verifier errors. We ask what errors verifiers make, how they affect training, which mitigations work against which errors, and how the findings extend to self-generated, rubric- and judge-based, and verifier-free rewards. The evidence suggests that the structure of an error matters more than its rate. Random errors mostly slow learning, though under a fixed budget or KL regularization they also lower final accuracy. False positives that the policy can learn to trigger can cause plateaus or collapse, sometimes even with a low initial error rate. Whether a verifier wrongly rejects or accepts depends on what it checks and who built it, not on its family. Its errors shift as the policy improves, and most mitigations have been tested against only one kind of error. We compile the findings into comparison tables, reconcile apparent conflicts in the literature, and propose a reporting protocol to make future results comparable.

[381] HaloBlock: Faster and Better Decoding for Diffusion Large Language Models cs.LG | cs.AI | cs.CLPDF

Zhaohui Zheng, Chenhang He, Jinrui Zhang, Ming-Ming Cheng, Lei Zhang

TL;DR: HaloBlock 根据双向累积熵估计扩散语言模型每块解码难度，再自适应分配块边界两侧的 halo 上下文，以缓解块间信息断裂并提升质量与收敛速度。

Details

Motivation: 扩散语言模型支持双向注意和任意顺序解码，但块内解码难度主要集中在块右侧 token，且块式生成会产生边界上下文不连续。

Result: HaloBlock 用双向累积熵估计区块难度，并据此为两侧动态分配 halo 区域，平滑跨块上下文传播、降低难 token 解码难度。作者在 LLaDA-8B-Instruct、Dream-7B-Instruct、LLaDA-1.5 上的数学推理和代码生成基准中，报告准确率一致提升且推理效率更高。摘要未列出具体准确率、速度或 halo 大小。

Insight: 不同区块需要的上下文范围和计算难度不一致，固定边界处理可能导致信息断裂；不确定性指导的 halo 分配让资源集中于困难 token。摘要未说明额外上下文计算开销及自适应区域对生成长度的影响。

Abstract: Diffusion large language models (dLLMs) have attracted increasing research attention due to their bidirectional attention and arbitrary order decoding. However, there are two issues in dLLMs inference: the decoding difficulty is dominated by the tokens on the right side of the block, and the block-wise decoding inherently suffers from boundary contextual discontinuity. To tackle these limitations, we propose HaloBlock, a difficulty-aware adaptive halo mechanism for dLLM inference. We first leverage bidirectional cumulative entropy to characterize block decoding difficulty. Then, based on the bidirectional decoding difficulty, we adaptively allocate bidirectional halo regions on both sides of each block to mitigate boundary contextual discontinuity. The difficulty-aware design of HaloBlock smooths cross-block context propagation and lowers decoding difficulty for hard tokens, yielding better generation quality and faster convergence. Evaluated on LLaDA-8B-Instruct, Dream-7B-Instruct and LLaDA-1.5 across mathematical reasoning and code generation benchmarks, HaloBlock demonstrates consistent accuracy improvements and higher inference efficiency. Source codes are publicly available at https://github.com/PolyU-VCLab/HaloBlock.

[382] Match the Canvas, Not Just the Answer: Predictability-Aware SFT for Diffusion Language Models cs.LG | cs.AI | cs.CLPDF

Miao Li, Hanyang Jiang, Hengyu Fu, Baihe Huang, Yating Liu

TL;DR: PACS 根据扩散语言模型自身可预测性选择训练时可见上下文长度，并在参考 token 即将按置信度揭示前监督，以缓解随机遮盖与推理解码的不匹配。

Details

Motivation: 扩散语言模型按可预测性而非固定从左到右顺序生成，因此 SFT 需要决定哪些参考 token 已可见、给多少生成空间；独立于推理过程的随机遮盖可能误导监督。

Result: 作者分析发现随机遮盖可能隐藏模型预测错误或惩罚解码本会解决的不确定性；注意力可见输入长度也会改变模型偏好的回答表达，即使解法相同。PACS 用模型自身可预测性为每个样本选定可见长度，并在参考 token 由置信度引导揭示前监督。在推理和代码生成基准上，PACS 平均准确率高于所比较 SFT 方法，单项基准相较 vanilla SFT 最高提高 12.84 个百分点。摘要未列基准、模型和各项具体结果。

Insight: 对齐训练时的可见上下文与推理时 token 揭示顺序，比随机遮盖更符合扩散语言模型的生成机制。最高单项提升不能代表所有任务增益；摘要未报告训练代价和不同解码顺序下的稳定性。

Abstract: Diffusion language models (DLMs) enable flexible parallel generation by resolving tokens according to the model’s evolving predictability rather than a fixed left-to-right order. This flexibility, however, introduces training choices absent from standard autoregressive teacher forcing: which reference tokens are visible when a target is supervised and how much generation space is allocated. Standard supervised fine-tuning (SFT) typically makes these choices through random masking, independently of inference. Our analysis shows how this training-inference mismatch can misdirect supervision. Random masking can conceal prediction errors or penalize uncertainty that decoding would otherwise resolve, while the attention-visible input length can change which response formulations the model favors, even when the underlying solution is unchanged. Motivated by these findings, we propose Predictability-Aware Canvas Selection and Supervision (PACS), an SFT method. PACS uses the model’s own predictability to select each example’s attention-visible input length and supervises reference tokens immediately before their confidence-guided revelation. Across reasoning and code-generation benchmarks, PACS consistently achieves the highest mean accuracy among the compared SFT methods, with gains of up to 12.84 percentage points over vanilla SFT on individual benchmarks. Code will be released upon publication.

[383] inLoRA-MINT: Internal Representations for Auditing Training Data in Domain-adapted LLMs cs.LG | cs.AI | cs.CLPDF

Gonzalo Mancera, Julian Fierrez, Aythami Morales, Ruben Tolosana, Daniel DeAlcala

TL;DR: inLoRA-MINT 利用 LoRA 微调模型的中间激活及参数相关信号进行成员推断，判断特定样本是否参与训练，不必依赖完整模型参数访问。

Details

Motivation: 成员推断测试可用于审计训练数据是否泄漏隐私；既有 LoRA-MINT 仅在严格黑盒下依赖困惑度，新方法探索内部表示是否能提供更强审计信号。

Result: 方法将激活和参数依赖信号输入 MINT 分类器，以区分训练样本与非训练样本。作者在四个 LLM（Qwen、Phi-4、DeepSeek-V3、LLaMA 3.2-3B）上测试三个覆盖医疗、数学和学术推理的数据集。报告的检测分数在不同架构和领域间为 95%–98%，并对模型配置和数据分布变化具有鲁棒性。摘要未说明分数具体指标、攻击假设或隐私成员推断的误报率。

Insight: 比单一困惑度更丰富的内部表示可能让数据来源审计不依赖完整参数访问；结果跨模型与任务域较一致。95%–98% 的分数需结合阈值、训练/测试相似度和误报设置解读，摘要未提供这些细节，不能等同于正式隐私泄漏风险量化。

Abstract: This work investigates the application of the Membership Inference Test (MINT) to large language models (LLMs) fine-tuned with LoRA, aiming to determine whether specific samples were used during training and address concerns about data privacy, leakage, and transparency. Building on our previous LoRA-MINT framework, which relies exclusively on perplexity under a strict black-box setting, we propose an approach based on intermediate model representations, including activations and parameter-dependent signals. These representations enable a MINT classifier to distinguish training from non-training samples without requiring full access to model parameters. We evaluated the method on four LLMs (Qwen, Phi-4, DeepSeek-V3, and LLaMA 3.2-3B) fine-tuned on three datasets covering medical, mathematical, and academic reasoning tasks. Results show consistent performance across architectures and domains, achieving scores between 95% and 98%, with robustness to variations in model configurations and data distributions. The proposed approach enables efficient auditing of sensitive training data and supports data provenance and compliance with emerging privacy and AI governance regulations. Although our experiments focus on LoRA-based fine-tuning, the methodology can be extended to other LLM adaptation techniques and, more broadly, to other AI models.

[384] LoOPSD: On-Policy Self-Distillation with Looped Teachers cs.LG | cs.AI | cs.CLPDF

Wei Fan, Qingyue Yang, Chao-Hong Tan, Qian Chen, Wen Wang

TL;DR: LoOPSD 搜索同一语言模型的层执行顺序与重复层，构造更强的循环教师，再用其对学生自身前缀的预测交替蒸馏，利用额外推理计算生成自监督信号。

Details

Motivation: on-policy 自蒸馏需要比学生更有用的教师预测，既有方法常通过额外上下文或指令强化教师；研究探索能否仅改变同一模型的层执行方式获得更强教师。

Result: LoOPSD 保持权重不变，搜索应重复执行哪些层，使用循环教师在学生采样前缀上的预测指导学习；每轮蒸馏后根据更新学生重新构造教师并继续搜索。在数学和逻辑推理任务上，作者报告多个模型均改善。GSM8K、400 token 思考预算下，Qwen3-1.7B 准确率从 42.30% 升至 79.61%，后续蒸馏轮次还能在首轮基础上继续提升。摘要未提供其他任务数值和搜索开销。

Insight: 模型自身不同层执行路径可产生额外计算和更强自教师信号，不需外部大模型或改动权重结构。GSM8K 提升很大，但摘要未说明是否与相同额外推理预算的其他方法对照，也未报告层序搜索成本。

Abstract: On-policy self-distillation improves a model using guidance from the model itself, but it still requires a teacher that provides more useful predictions than the student. Existing approaches obtain such guidance by giving the teacher additional context or instructions. We ask whether it can instead come from changing how the same model uses its own layers. To address this, we introduce LoOPSD, which searches over a model’s layer execution order to construct a stronger teacher and distills its advantage into the student. Specifically, we keep the model weights unchanged while searching for layers to repeat, and use the resulting teacher’s predictions on student-generated prefixes to guide learning. We then alternate search and distillation, rebuilding the teacher from the updated student in each round so that the student progressively absorbs the gains from alternative layer execution. Experiments on mathematical and logical reasoning tasks show improvements across models. On GSM8K with a 400-token thinking budget, LoOPSD raises Qwen3-1.7B accuracy from 42.30% to 79.61%; subsequent distillation rounds yield further gains beyond the first round. Together, these findings show that by searching layer execution to construct teachers, we can turn a model’s additional computation into improved student capabilities, opening a new avenue for self-distillation to derive supervision from the model’s own computation.

[385] Teacher-Guided Tool Learning for Small Language Models via Bidirectional KL Constraints cs.LG | cs.CLPDF

Haoxuan Gao, Qin-Wen Luo, Xiang Chen, Yu Tian, Zequn Sun

TL;DR: SeTPO 结合低奖励轨迹上的反向 KL 教师 token 反馈与持续正向 KL 示范监督，帮助小语言模型学习工具使用并缓解师生行为分布不匹配。

Details

Motivation: 小模型仅靠强化学习难以探索有效工具行为；直接模仿教师又可能因师生分布差异传播教师偏好错误。

Result: SeTPO 将教师反向 KL 推导的逐 token 反馈加入低奖励轨迹的策略优势，同时通过辅助正向 KL 在训练过程中持续保留教师示范行为。作者在 BFCL v4 和 API Bank 上比较监督微调、强化学习及蒸馏基线，并报告该方法在三种学生规模上整体表现优越。理想化双向 KL 目标的理论分析表明，最优策略期望奖励不低于教师，同时保留一定比例的教师概率质量。摘要未给学生规模、具体成绩或概率保留比例。

Insight: 用不同方向的 KL 项分别鼓励探索改进和保留可用示范，是应对小模型工具探索不足与纯模仿偏差的折中。理论保证依赖理想化目标；摘要未量化教师错误传播是否实证减少及额外训练成本。

Abstract: External tools can extend the capabilities of small language models (SLMs), enabling their deployment in resource constrained environments. However, their limited capabilities make it difficult for SLMs to discover effective tool use behaviors through reinforcement learning alone, while direct teacher supervision suffers from teacher student behavioral distribution mismatch and may propagate erroneous preferences. To address these challenges, we propose Selective Teacher guided Policy Optimization (SeTPO), which coordinates student exploration with selective teacher guidance and persistent demonstration supervision. Specifically, SeTPO incorporates reverse KL derived token level teacher feedback into policy advantages on low reward trajectories, while auxiliary forward KL supervision preserves teacher demonstrated behaviors throughout training. Under an idealized bidirectional KL objective, we theoretically establish that the optimal policy achieves expected reward no lower than the teacher’s while retaining a guaranteed fraction of the teacher’s probability mass. Extensive experiments on BFCL v4 and API Bank demonstrate that SeTPO achieves superior overall performance over supervised fine tuning, reinforcement learning, and distillation baselines across all three student scales.

[386] Bayesian Test-Time Inference of Task-Aligned Similarity from Weak Interactive Feedback cs.LG | cs.CLPDF

Weida Liang, Kenji Kawaguchi

TL;DR: BASIS 用贝叶斯后验从弱交互反馈中推断当前决策应采用哪种相似性标准，并联合更新候选目标与标准选择，而非固定相似度或静态混合权重。

Details

Motivation: 检索、排序和回答修订都依赖相似性判断，但词面、语义、结构和风格相似性可能导向不同候选，任务对“相关”的定义还可能随实例改变。

Result: 训练免除的 BASIS 在给定任务专用标准库中维护候选目标与标准的联合后验，可根据标量分数、成对偏好或验证器输出更新，并按期望信息增益选择探测。在有限噪声模型中分析目标—标准可识别条件，并说明标准切换时固定标准或固定混合会失败。概念恢复、推理示例选择和回答修订任务中，6 次探测后 Success@1 从调优静态混合的 56.0% 提至 68.1%；推理准确率 72.3%→73.1%，修订胜率 61.4%→63.8%。它还超过同算力 best-of-N 在推理（72.1%）和修订（61.2%）的结果。收益在标准不匹配和候选池较大时最大。方法只在给定标准库中选择，不会发现新标准。

Insight: 自适应确定“按什么标准相似”可比固定融合处理实例间标准切换，主动探测也能更快收敛；候选目标和标准联合推断比仅调权重进一步有益。其能力受预先提供的标准库限制，无法自动发现新的相关性维度。

Abstract: Language model systems repeatedly make test-time decisions by comparing similarity: they retrieve exemplars, rank candidates, and revise responses. Yet similarity is not unique: lexical, semantic, structural, and stylistic resemblance can point to different candidates, and the task-aligned notion can change across instances. We introduce BASIS (Bayesian Axis Selection via Interactive Signals), a training-free Bayesian method that uses weak test-time feedback to infer which criterion in a supplied, task-specific library governs the current decision. It maintains a joint posterior over candidate targets and criteria, updates it from scalar scores, pairwise preferences, or verifier outputs, and selects probes by expected information gain. In a finite noisy model, we characterize when the target-criterion pair is recoverable and show why fixed-criterion and fixed-mixture rules can fail under criterion shift. Across concept recovery, reasoning exemplar selection, and response refinement, BASIS improves six-probe concept-recovery Success@1 from 56.0% to 68.1%, reasoning accuracy from 72.3% to 73.1%, and refinement win rate from 61.4% to 63.8% over tuned static mixtures. It also exceeds compute-matched best-of-N selection on reasoning (72.1%) and refinement (61.2%). Gains are largest under criterion mismatch and with larger candidate pools, and matched controls indicate that adaptive criterion weighting explains much of the benefit, with joint target-criterion inference adding further value. BASIS selects only among supplied criteria and does not discover new ones.

[387] Does self-distillation improve reasoning? cs.LG | cs.AI | cs.CLPDF

Jason Ken Adhinarta, Raina Wu, Ashutosh Sharma, Avinash Balakrishnan, Rameswar Panda

TL;DR: 该研究重新评估 SDPO 与 OPSD 等自蒸馏方法，发现其能力提升未能稳定泛化到未见问题，收益多可由过拟合或已有推理行为迁移解释。

Details

Motivation: 自蒸馏可利用正确答案或环境反馈等特权信息，有研究报告其样本效率和泛化优于强化学习；但这些收益是否超出域内评测、是否确由特权信息带来仍不明确。

Result: 作者在代码、科学、工具使用和数学推理上重测 Self-Distillation Policy Optimization、On-Policy Self-Distillation。结果显示已报告增益并不稳定迁移到未见问题，更多可由过拟合或已有推理行为转移解释，而非有效利用特权信息。自蒸馏梯度与直接最大化期望奖励的策略梯度对齐较弱，质疑其是否提供有效策略改进方向。实验覆盖 Qwen3 1.7B–14B 和 OLMo-3-7B。摘要未给各任务具体数值。

Insight: 自蒸馏提升域内表现并不自动证明特权监督被有效利用；应单独测试未见问题并比较梯度方向。作者认为有效利用特权信息仍是开放问题，结论针对所评方法和模型，不能概括所有自蒸馏变体。

Abstract: Self-distillation is an emerging approach that trains a model to imitate its own predictions when conditioned on privileged information, such as correct answers or environmental feedback. Recent methods have reported better sample efficiency and generalization than reinforcement learning (RL), but it remains unclear whether these gains extend beyond in-distribution evaluations and whether they stem from privileged distillation. In this paper, we revisit two representative methods, Self-Distillation Policy Optimization (SDPO) and On-Policy Self-Distillation (OPSD), across code, science, tool use, and math reasoning tasks. We find that their reported gains do not consistently generalize to unseen problems and are largely explained by overfitting or transferring existing reasoning behavior, rather than effectively using privileged information. Moreover, self-distillation gradients show weak alignment with policy gradients that directly optimize expected reward, questioning whether they provide an effective direction for policy improvement. Our experiments across Qwen3 models (1.7B-14B) and OLMo-3-7B suggest that effectively leveraging privileged information to improve reasoning remains an open challenge. Our code is available at https://github.com/sleepypandu/does-sd-improve-reasoning

[388] On the Monitorability of Natural Language Autoencoders: Identification vs Discrimination cs.LG | cs.CLPDF

Kevin Yandoka Denamganaï

TL;DR: 该理论与实验研究指出，自然语言自编码器训练让激活解释能指认来源激活，却未充分训练其排除相似错误解释，因此人工监督时的真实性判别能力有限。

Details

Motivation: 自然语言自编码器通过激活 verbaliser 将内部状态写成解释，再由 reconstructor 将解释还原为向量；监督价值取决于人类能否判断某段解释真/假，但生成式训练与这种判别式监督任务不同。

Result: 作者将训练形式化为生成式指称博弈，将监督部署形式化为多模态、部分可观测且一对多/多对多的判别式博弈，并推导两者差异会导致模型能识别解释对应哪个激活，却不能有效区分真实解释与只差一项主张的流畅假解释。用同一 NLA 的 AR 作偏袒性代理监控者测试一个公开检查点：320 次比较中真实解释均胜过其他激活的解释；但与只差一项主张的近似假解释比较，真实解释仅 60.2% 胜率（标准误 6.7%，n=35）。不相关模型的保义改写仅使其到其他激活解释的距离变化 2.1%，因此失败并非仅因内部私有用语。实验仅检查对话可确定的 transcript 级真值，未测试模型内部处理真实性。

Insight: 激活解释能“指向”来源状态，不代表能在细微近似说法中可靠判真；训练时需让解释具备排除能力。实验用同模型重构器作代理监控且只测一个检查点，作者明确称这是较容易的 transcript 真值情形，内部过程主张仍未验证。

Abstract: Natural Language Autoencoders (NLAs) are proposed as oversight tools for language models: an Activation Verbaliser (AV) writes an English explanation of an activation vector, and an Activation Reconstructor (AR) maps the explanation back to a vector. Their value for oversight depends on a human monitor being able to tell a true account of an activation from a false one, given the Av’s explanation. We formalise NLA training and NLA oversight deployment as two different referential games from Emergent Communication: training is a generative referential game, while oversight is a discriminative, multi-modal, partially-observable referential game played over a many-to-many relation between the monitored language model’s activations and the accounts that are true of them. The gap between the two games predicts that AV’s explanations will identify which activation they describe without discriminating its true account from closely matched false ones, since training never asks explanations to rule anything out. Using the NLA’s own AR as a maximally-sympathetic surrogate monitor model, we experimentally confirm this theoretical insight on one released checkpoint: true accounts beat accounts of other activations in all 320 comparisons, but beat fluent rivals differing in a single claim only 60.2% of the time (std.err. 6.7%, chance at 50%, $n = 35$). A meaning-preserving rewrite by an uninvolved model costs only 2.1% of the distance to another activation’s account, so the failure is not due to the NLA pair having developed private conventions in the AV’s explanations. Our rival accounts test only transcript-level truths, which the monitored language model’s conversation settles, so these results bound the easiest case and leave claims about the monitored model’s own processing untested. We state conjectures that would make both results consequences of the framework.

[389] SkillScaffold: Scaffolding Agentic Reinforcement Learning with Policy-Adaptive Skill Credit Assignment and Selective Injection cs.LG | cs.CLPDF

Wenlin Zhong, Hao Wang, Kun Yuan, Fan Yang, Jinluan Yang

TL;DR: SkillScaffold 用已验证训练轨迹产生候选技能并以结果感知似然比较估计其效用，不需候选专属环境 rollout；仅在预计有帮助时向当前策略注入技能。

Details

Motivation: 文本技能可把历史经验转成可复用指导，但现有 skill-RL 常假定注入必然有益，或用间接反馈和大量 rollout 评估技能；技能收益还随候选与策略对当前查询的能力而变。

Result: 框架根据 outcome-verified 轨迹生成候选技能，再进行低成本效用估计，并选择注入哪些技能；若没有候选预计有帮助，则保持无技能执行。Generator 与 Executor 联合训练，使技能生成适配不断演化的执行策略；技能支架也促使推理时无外部技能的内化。在 ALFWorld、WebShop、ScienceWorld、τ²-Bench 上，作者报告相较强 agentic-RL 和 skill-RL 基线，在短程与长程任务均有提升；消融和分析支持效用选择及允许拒绝注入的重要性。摘要未提供具体数值。

Insight: 将技能注入作为可选择的策略决策，允许对无益技能弃权，避免“有技能就一定好”的假设；基于已验证轨迹的似然估计则省去逐候选环境 rollout。摘要未报告效用估计校准、任务级收益和系统额外开销。

Abstract: Agentic reinforcement learning (RL) enables large language models (LLMs) to learn from multi-turn interactions with verifiable outcomes. Textual skills can improve exploration by distilling prior interactions into reusable guidance. However, existing skill-based RL approaches often either assume skill injection is beneficial or rely on indirect feedback or rollout-intensive evaluation to assess its utility. Our pilot studies show that rollout-based utility estimation for skill credit assignment is costly and unstable under limited sampling, while skill benefits vary across candidates and depend on the current policy’s competence on each query. To address these challenges, we introduce SkillScaffold, a policy-adaptive training scaffold that combines low-cost skill-utility estimation with selective skill injection. By reusing outcome-verified training trajectories, SkillScaffold generates candidate skills and estimates their utility through outcome-aware likelihood comparisons, without candidate-specific environment rollouts. The resulting utility guides whether and which skill to inject, retaining skill-free execution when no candidate is estimated to help. Joint training enables the Generator to produce skills tailored to the evolving Executor. Scaffolded RL encourages skill internalization for skill-free inference. Experiments on ALFWorld, WebShop, ScienceWorld, and $τ^2$-Bench demonstrate gains over strong agentic-RL and skill-based RL baselines across both short- and long-horizon tasks. Ablations and further analyses support the importance of utility-based selection and abstention. Our code is available at https://github.com/Vaidurya00/SkillScaffold.

[390] Barycentric Anchored Weight Search for Robust Model Merging cs.LG | cs.AI | cs.CVPDF

Antoni Buades, Jose-Luis Lisani

TL;DR: Barycentric Anchored Search 使用预训练权重与两个微调检查点构造重心候选，并在锚定权重空间做小规模免梯度搜索，以提升分布外鲁棒性且不增加推理成本。

Details

Motivation: 微调视觉模型在分布偏移下可能退化；固定权重平均或模型合并规则有限，且常需多个微调检查点，而额外梯度训练又会增加成本。

Result: 方法沿候选重心权重与预训练初始化之间的方向进行少量无导数搜索。候选用一个或多个 OOD 数据集的小型验证子集选择，再在其留出部分、其他 OOD 基准和 ImageNet 验证集评估鲁棒性—准确率权衡。CLIP ViT-B/32 实验中，作者报告在 ImageNet-A、ImageNet-R、ImageNet-Sketch 等留出 OOD 集上表现显著改善，同时保持 ImageNet 准确率具有竞争力；对未用于候选选择的 OOD 集也能迁移，并超过原微调检查点、朴素 Soup 和 Model Stock。摘要未给具体分数。

Insight: 只需两个微调检查点并借助 OOD 验证反馈即可自适应调整合并，比固定平均更灵活；不增加推理成本。评估在验证子集上进行选择，需谨慎区分这些已用于选型的数据与额外 OOD 泛化集；摘要未量化搜索预算和具体泛化增益。

Abstract: We propose Barycentric Anchored Search, a post-hoc weight-space procedure for improving fine-tuned vision models under distribution shift without additional gradient training or increased inference cost. Given a pre-trained initialization and two checkpoints fine-tuned on a source dataset, the method constructs barycentric candidate weights and performs a small derivative-free search along directions connecting each candidate to the pre-trained model. Candidate weights are selected using a small validation subset of one or more out-of-distribution (OOD) datasets; the selected model is then evaluated on the held-out portions of those datasets, on additional OOD benchmarks, and on the ImageNet validation set to quantify the robustness-accuracy trade-off. The method is inspired by simplex-based derivative-free optimization, including Nelder-Mead and Torczon multidirectional search, but is specialized to the geometry of neural network weight merging. Unlike Model Soup, which typically benefits from many fine-tuned checkpoints, and unlike Model Stock, which applies a fixed geometric merge, our approach uses only two fine-tuned checkpoints and iteratively adapts the merge using validation feedback from OOD data. Experiments with CLIP ViT-B/32 show that the proposed search substantially improves performance on held-out OOD data, including ImageNet-A, ImageNet-R, and ImageNet-Sketch, while keeping ImageNet accuracy competitive. The resulting weights also improve transfer to OOD datasets not used during candidate selection, outperforming the original fine-tuned checkpoints, naive model soups, and Model Stock in our evaluations. These results suggest that lightweight optimization in anchored weight space provides a practical alternative to fixed weight-averaging rules, geometric merging, or single-model selection.

[391] PLDiv: Measuring Dataset Diversity and Semantic Collapse via Persistence Landscapes cs.LG | cs.AI | cs.CV | stat.MLPDF

Yang Ba, Michelle V. Mancenido, Rong Pan

TL;DR: PLDiv 用持久景观刻画数据点邻域随连接尺度变化的拓扑结构，以单个指标衡量语义多样性并检测语义坍缩。

Details

Motivation: 不断扩大的训练集可能只是增加语义重复样本而未扩展覆盖面；词汇变化、熵或成对相似度等现有多样性指标难以反映底层分布的整体几何结构。

Result: PLDiv 从数据的几何表示构建持久景观，追踪邻近样本随连接尺度增加逐步形成的结构，并汇总成无需任务监督的标量多样性度量。作者在文本生成多样性、图像嵌入、指令数据和推理轨迹语义坍缩等任务上评估，报告与已测基线竞争，并在语义坍缩检测中与多样性标签高度对齐。摘要未给相关系数、数据集或基线数值。

Insight: 拓扑结构可补充词面差异与局部相似度，帮助识别样本集合是否覆盖多个语义群体；指标不依赖任务标签，适用于数据监控和生成模型评估。摘要未量化尺度选择、嵌入模型偏差或与具体基线差距。

Abstract: Modern data pipelines increasingly combine human-created and human-curated data with AI-generated data for model training and evaluation. As datasets grow, they often add semantically redundant samples rather than expanding semantic coverage, which deteriorates the quality of the whole dataset. Existing diversity metrics primarily summarize lexical variation, entropy, or pairwise similarity, but provide limited characterization of the shape of the underlying data distribution. In this paper, we propose PLDiv, a diversity metric derived from persistence landscapes that quantifies the structural organization of a dataset from its geometric representation. PLDiv summarizes how data neighborhoods evolve as nearby samples progressively connect, producing an interpretable scalar measure of structural diversity without requiring task-specific supervision. Whether measuring intrinsic diversity in text generations or image embeddings, or detecting semantic collapse in instruction datasets and reasoning trajectories, PLDiv performs competitively against evaluated baselines and aligns strongly with diversity labels during semantic collapse. Consequently, it effectively supports applications in synthetic data monitoring, dataset construction, data augmentation, and generative model evaluation.

[392] Design Choices for Multimodal Transformer Learning with Missing and Longitudinal Healthcare Data cs.LG | cs.CVPDF

Rita Cordeiro Mendes, Maria Manuel Pereira, Carlos Santiago, Catarina Barata

TL;DR: 该研究系统比较纵向癌症医疗数据中多模态 Transformer 的 token 粒度、缺失模态处理和时间编码策略，指出多模态收益会被缺失数据显著影响。

Details

Motivation: 医疗数据常有整个模态缺失或模态内部字段缺失；纵向序列虽适合 Transformer，但 token 如何组织、缺失如何处理、时间信息如何编码都会影响预测。

Result: 作者在三个癌症数据集上比较模态级与特征级 tokenization，以注意力掩码和插补处理自然及合成缺失，并评估生存二分类与离散时间生存预测的时间编码策略。实验显示多模态总体优于单模态，但缺失数据会显著影响预测，合理 token 粒度和模态内缺失处理对可靠预测至关重要。摘要未列出各设计的具体性能或最佳配置。

Insight: 多模态并非必然稳健，缺失机制和输入粒度是架构选择的核心因素；比较自然缺失与人工缺失也有助于区分数据机制影响。摘要没有提供结果数值和各癌种细节，无法给出统一的最佳策略。

Abstract: Multimodal learning offers the potential to improve predictions by integrating complementary information from heterogeneous data sources. However, multimodal datasets are frequently incomplete, raising important questions about how models should handle missing information. Although recent approaches increasingly rely on Transformer-based architectures for multimodal fusion, incomplete multimodal data introduces challenges such as missing modalities and intra-modal missingness, which motivate key architectural design choices, such as tokenization granularity. In addition, many datasets are longitudinal, with observations collected across multiple time points. The sequential nature of the Transformer architectures provides a natural framework for modeling such data, but introduces additional considerations, such as how temporal information should be represented. In this study, we systematically investigate these design choices using three different cancer datasets. We compare modality-level and feature-level tokenization and evaluate attention masking and imputation as strategies for handling missing information, both naturally occurring in the datasets and synthetically induced. We further investigate approaches for incorporating temporal information in our training process and how to model it. Experiments are conducted with unimodal and multimodal configurations, as well as different levels of missingness for binary survival classification and multiple time encoding strategies for discrete-time survival prediction. Our results show that while multimodal outperforms unimodal results, missing data heavily impacts predictions and thus effective design choices such as tokenization granularity, and intra-missing handling is essential for reliable prediction.

[393] CIViC: Evidence Gates for Audio-Conditioned Visual Prediction on Proxy Tasks cs.LG | cs.CV | cs.SD | eess.ASPDF

Mengzhe Geng, Max W. Y. Lam, Yuewen Cao, Shoukang Hu, Jing Xu

TL;DR: CIViC 提出四道检查门，将模型解释与干预、留出任务比较和终点控制连接起来，评估音频条件是否真正影响视觉预测及其影响是否有益。

Details

Motivation: 移除音频导致视觉预测改变，并不能说明任务使用了预期音频信息；需要区分输入敏感性、任务相关收益和终点行为控制。

Result: 协议在演员隔离的人脸栅格任务、合成音频特征混合和视频隔离的 SpeakerVid 代理任务上演示。静态人脸任务中跨演员替换音频满足预注册检测规则但效应较小；单独时间终点上，四个匹配种子中恒定音频拟合优于音频训练拟合。合成混合任务中加入条件的拟合在 1,000 步时 RMSE 更高（0.1807 对 0.1707），差距随训练缩小。SpeakerVid 代理任务中，五个低层音频摘要置零、保留身份/对话/条件输入的模型，在四种留出条件和五个配对种子上全帧灰度 RMSE 更低，而对齐音频在片段漂移误差上更低。结果表明不同代理任务中输出激活、留出误差与终点控制支持不同解释。

Insight: 音频输入对输出有影响，不等于模型利用了目标音频信息或得到更好任务预测；不同终点和干预会导向不同结论。实验为代理任务，研究强调多门控诊断，不能据此推断真实说话视频生成或音视频因果耦合能力。

Abstract: Removing audio can shift a visual predictor’s output without showing whether the task uses the intended audio information. We introduce Conditional Input and Visual Interpretation Checks (CIViC), a four-gate protocol that links an interpretation to an intervention, held-out task comparison, or endpoint control. We demonstrate CIViC on actor-disjoint face-raster tests, a synthetic audio-feature mixture, and source-video-disjoint SpeakerVid proxy tasks. In the static face-raster test, cross-actor audio replacement meets the registered detection rule with a small effect; on a separate temporal endpoint, constant-audio fits score better than audio-trained fits in four matched seeds. In the synthetic mixture, the condition-augmented fit has higher 1,000-step RMSE than the audio-only fit (0.1807 vs. 0.1707), with the gap shrinking during training. In the SpeakerVid proxy, the five-feature-zero fit (five low-level audio summaries zeroed; identity, dialogue, and condition inputs retained) has lower whole-frame grayscale RMSE than aligned audio across four held-out conditions and five paired seeds, while aligned audio has lower segment-drift error. The results show that output activation, held-out error, and endpoint controls support different interpretations across these proxy tasks.

[394] SYNPRED: A Synergistic Approach to Multimodal Learning for Clinical Prediction cs.LG | cs.CVPDF

Ivana Janíčková, Yen Y. Tan, Thomas H. Helbich, Philipp Seeböck, Julieta Di Marco

TL;DR: SYNPRED 用多模态 VAE 将任务相关的单模态与跨模态信息同实例特异、任务无关变化分离，以便在模态缺失时仍利用互补预测信号。

Details

Motivation: 对比学习主要利用模态间共享信息，可能忽略对临床任务有用但模态特异的互补线索；实际应用中某些模态在测试时也可能缺失。

Result: 作者提出协同多模态 VAE，在表示中区分任务相关的单模态/多模态信息和实例级无关变化，并用于医疗影像与 RNA 的治疗响应预测。与对齐式方法和多模态 VAE 基线相比，作者报告多模态预测表现提升；即使仅用单一模态预测，多模态训练仍带来稳定收益。摘要未给数据集规模、任务指标及缺失模态比例。

Insight: 共享信息之外的模态专属信号也可能互补；联合训练所学表示在单模态推理时仍有迁移价值。结果来自单一治疗响应任务，具体预测增益及不同缺失机制下的性能未见摘要。

Abstract: Multimodal medical data captures complementary aspects of disease. It provides a comprehensive basis for diagnosis and the identification of biological associations linked to progression or treatment response. Established contrastive learning approaches exploit shared information across modalities, but fall short in exploiting complementary, modality-specific signals relevant for a task. Here, we propose a synergistic multimodal variational autoencoder (VAE) framework that separates task-relevant unimodal and multimodal information from instance-specific variation task-irrelevant, allowing complementary signals across modalities to be leveraged for prediction. The method yields benefits even if modalities are missing at test time. We evaluate the method on a treatment response prediction task using medical imaging and RNA data. Empirical comparison with alignment-based and multimodal VAE state-of-the-art approaches demonstrates improved multimodal prediction performance, and consistent benefits for prediction from a single modality after multimodal training. The full code can be found here: https://github.com/cirmuw/SYNPRED.

[395] A Controlled Evaluation of Curvature Inspired Scalar Feature Scaling for Domain Generalization cs.LG | cs.AI | cs.CVPDF

Sunny Gupta, Amit Sethi

TL;DR: 该研究以严格限定的正标量特征缩放实现一个曲率启发训练包，并通过 384 次受控实验检验其是否改善域泛化；结果未发现其优于 ERM 的调整后证据。

Details

Motivation: “曲率学习”可指不同计算机制，需先明确实现再评估其对域泛化的影响，避免把欧氏特征缩放误称为双曲表示或曲率优化。

Result: 作者审计的 HyperDG 实际是在训练时以学习到的正标量缩放欧氏编码器特征，推理时使用未缩放欧氏特征和普通线性分类器；不含 Lorentz/Poincaré 表示、测地线预测或 SAM 内循环。在 PACS、VLCS、Office Home、TerraIncognita 的六种条件下共完成 384 次运行，检查点只按三个源域验证集选择。完整实现相较 DomainBed ERM 的等数据集宏平均点估计低 1.61 个百分点，95% 聚类 CI 为 [-3.17,-0.04]；四个数据集点估计均偏向 ERM，但精确符号翻转检验 p=.125、Holm 校正 p=.250，不支持校正后差异。更干净的领域特定缩放对比为 -0.13 点（95% CI [-0.53,0.27]），在预先设定 ±1 点实际意义界内达到等效性检验；该结论仅限源域选择的套件宏平均。

Insight: 这是对特定标量训练方案的有界负结果，不能据此否定真正的双曲表示学习。作者仔细区分整体包与领域特定缩放的控制对比，并明确结论不覆盖每个数据集、目标域或最终训练步。

Abstract: Domain-specific curvature is an appealing idea for domain generalization (DG), but curvature learning can denote substantially different computations. This paper evaluates one deliberately narrow implementation: the code calls it HyperDG, but it rescales Euclidean encoder features during training by learned positive scalars and uses an ordinary linear classifier on unscaled Euclidean features at inference. It does not implement a Lorentz or Poincaré representation, geodesic prediction, or a Sharpness-Aware Minimization inner loop. We audit the executed objective and evaluate six conditions on PACS, VLCS, Office Home, and TerraIncognita. The completed design contains 384 runs: six conditions, four held-out targets, and four run seeds per dataset. Checkpoints are selected using only the three source-validation environments. Inference first averages paired seed differences within each held-out target; targets, rather than 16 seed-by-target observations, are the per-dataset clusters. The full implementation’s equal-dataset macro point estimate is 1.61 percentage points below native DomainBed ERM (95% clustered CI [-3.17, -0.04]). All four dataset point estimates favor ERM, but an exact four-dataset sign-flip test gives p = .125 and Holm-adjusted p = .250; we therefore claim no adjusted evidence of a difference. The cleaner domain-specificity contrast, separate learned source scales without their variance penalty versus one shared learned source scale, is -0.13 pp (95% CI [-0.53, 0.27]). Its 90% CI lies within a revision-stage $\pm 1$ pp smallest effect of practical interest (TOST p = .003), supporting equivalence only for the source-selected suite macro, not every dataset, target, or final-step endpoint. The study thus supplies a bounded negative result for this scalar training package; it does not test or refute genuine hyperbolic representation learning.

cs.NI [Back]

[396] Semantic Edge-Cloud Communication for Real-Time Urban Traffic Digital Twins cs.NI | cs.CV | cs.DC | eess.SPPDF

Murat Arda Onsu, Arman Dogru, Poonam Lohan, Burak Kantarci, Aisha Syed

TL;DR: 该框架用边缘智能从交通视频提取与任务相关的语义信息并传给云端，重建实时交通数字孪生，以减少直接传输原始视频的通信负担。

Details

Motivation: 城市数字孪生需要摄像头持续同步高保真道路状态，用于监测、碰撞预测和交通流管理；原始视频流带来显著带宽和延迟压力。

Result: 提出语义通信边缘—云框架，由边缘侧提取任务相关交通语义而不是上传整段原始视频，以重建数字孪生所需的关键动态。实验报告传输量节省超过 99%，平均轨迹重建误差为 0.12 米。摘要未说明基准道路、相机数量、比较的原始带宽或延迟测量值。

Insight: 若传输只保留数字孪生重建所需语义，通信成本可大幅下降，同时保持轨迹精度。99% 节省与 0.12 米误差需结合具体场景和数据编码协议理解，摘要未给高峰负载、漏检事件或云端时延。

Abstract: Urban traffic systems are becoming increasingly complex, demanding real-time traffic perception, proactive collision prediction, and intelligent management to improve safety and reduce congestion. Although AI-driven traffic management has shown strong promise, its success relies on diverse, high-quality data spanning critical scenarios such as collisions, near-collisions, and traffic rule violations. Digital twins offer a powerful solution by enabling synchronized virtual replicas of physical traffic environments for real-time monitoring, collision prediction, traffic-flow optimization, and diverse realistic scenario generation. However, continuously synchronizing a high-fidelity urban traffic digital twin through edge surveillance cameras requires substantial edge-cloud communication, and transmitting raw video streams leads to severe bandwidth and latency constraints. To overcome this challenge, this paper proposes a semantic communication-based edge-cloud framework for real-time urban traffic digital twin applications. By leveraging edge intelligence to extract and transmit only task-relevant semantic traffic information instead of full video feeds, the proposed framework significantly reduces communication overhead while preserving the essential dynamics needed for accurate real-time digital twin reconstruction. Experimental results demonstrate more than 99% transmission savings and an average trajectory reconstruction error of only 0.12 m, confirming the proposed framework as an effective, scalable, and low-latency solution for AI-enabled urban traffic digital twin applications.

cs.RO [Back]

[397] SEAR: Self-Evolving Agentic Robotics cs.RO | cs.AI | cs.CLPDF

Xinyi Wang, Bangzheng Li, Hanyu Wang, Lin Zhang, Andrea Yaoyun Cui

TL;DR: SEAR 研究机器人智能体通过写控制代码、在线控制或自动研究微调 VLA 来自我进化，并用基准区分真实经验迁移与重复尝试、熟悉任务或额外预算带来的表面提升。

Details

Motivation: 前沿语言模型智能体可编写机器人代码、测试并把有效技能和记忆用于后续任务，但需要验证它们是否真正学习，以及知识能否迁移到新任务、模拟器和实体机器人。

Result: 框架含 code-as-policy、agent-as-policy 和 auto-research 三种模式；基准控制重复尝试、熟悉任务、特权模拟状态、更多预算及运气。对 7 个前沿 LLM、330 个模拟任务、5 个实体任务及累计四年以上智能体运行时间的研究发现：学习曲线同时依赖总时长和任务难度/模型能力，困难任务更受益于长运行，简单任务多次短尝试即可；控制代码和共享程序可跨任务/模拟器迁移，任务特定方案不行；实体机器人上 agent 设计练习与新的人类设计场景存在性能差距。不同模型偏好的进化模式不同；文本模型可追平相近多模态模型，感知有助于错误评估，但推理和工具使用可能更关键。

Insight: 该工作强调把学习与重复试验、预算和任务熟悉度分开评估，也揭示模拟器内转移不保证真实场景迁移。实体机器人仅 5 个任务，且摘要未提供具体成功率，文本与多模态模型的比较需限于所测模型和场景。

Abstract: Frontier language-model agents can now improve robots on their own: they write control code, test it, and keep what works as skills and memories for the next task. We introduce SEAR (Self-Evolving Agentic Robotics), a framework and benchmark built around two questions. Is the agent really learning, or just trying more often? And does what it learns still help on new tasks, in a new simulator, and on a real robot? The framework runs an agent in three modes: it writes code that becomes the policy (code-as-policy), acts live in the control loop (agent-as-policy), or works as a researcher that fine-tunes a vision-language-action model (auto-research). The benchmark adds controls that separate retained experience from repeated attempts, familiar tasks, privileged simulator state, larger budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second, control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception helps with error assessment, but agentic reasoning and tool use can matter more. We release the benchmark and codebase to support future research.

[398] SenseWeather: Multimodal Sensing for Robust Robotic Perception under Extreme Conditions cs.RO | cs.CVPDF

Fucai Ke, Sukai Huang, Junyi Hao, Kal Backman, Naim Rastgoo

TL;DR: SenseWeather 评测极端天气下的机器人多模态感知，并用深度与红外传感融合重建任务兼容的 RGB 观测，使冻结下游模型无需改动也能从辅助传感器受益。

Details

Motivation: 雨、雪、尘、雾和黑暗会严重退化主 RGB 输入，但机器人已配备深度和红外等传感器；现有 RGB 中心感知栈与设备实际多模态观测不匹配。

Result: 基准提供对齐 RGB、深度、红外、干净参考图和高层任务标注。Fusion-to-RGB 有两种参数高效实现：扩散 Transformer 的 token 空间融合，以及 inpainting U-Net 的残差潜变量条件。冻结的开放词汇 grounding 和 VLM 下游比较中，RGB+D+IR 融合优于对应 RGB-only 恢复器的所有配对设置，十组检测器比较的 AP50 均提升。感知提升不总与传统图像保真指标一致，天气越严重也不一定导致下游退化单调加重。摘要未提供 AP50 的具体数值。

Insight: 辅助传感器的价值应以任务级感知表现衡量，而不只看重建图像像素质量；天气严重程度和下游性能也非简单单调关系。结论来自受控基准和冻结模型，真实部署中的传感器配准误差与动态环境影响仍需测试。

Abstract: Robotic perception must remain reliable when extreme environmental conditions severely degrade the primary RGB observation. Although robotic platforms commonly carry complementary sensors such as depth and infrared, much of the modern perception stack remains RGB-centric, creating a mismatch between multimodal onboard sensing and pretrained downstream models. We study whether complementary sensing can be used upstream to recover a task-compatible RGB observation that preserves existing perception capabilities without modifying or retraining the downstream stack. To this end, we introduce SenseWeather, a controlled multimodal benchmark for robotic perception under severe rain, snow, dust, fog, and darkness, with registered RGB, depth, and infrared observations, clean references, and grounded high-level perception tasks. We further develop Fusion-to-RGB, with two parameter-efficient sensor-fusion realizations based on token-space fusion for a diffusion transformer and residual latent conditioning for an inpainting UNet. Across frozen open-vocabulary grounding and vision-language models, RGB+D+IR fusion improves all paired downstream comparisons over the corresponding RGB-only restorers, including AP50 gains across all ten detector comparisons. Importantly, these perception gains do not consistently follow conventional image-fidelity metrics, and increasing weather severity does not induce monotonic downstream degradation. These results show that complementary sensing can preserve reusable robotic perception under extreme conditions, while highlighting task-level perception robustness as an objective distinct from visual restoration quality.

[399] RAIN: Robotic Task Generalization with Region-Aware Interaction Networks cs.RO | cs.CVPDF

Seunghun Lee, Kyuhwan Shim, Yeogyeong Song, Seung Hwan Kim, Sihaeng Lee

TL;DR: RAIN 将机器人任务分解为由动作类型和目标区域掩码指定的子任务，以跨视角目标条件特征、轨迹规划和完成判定支持技能跨物体与场景复用。

Details

Motivation: VLA 模型从大规模图文预训练受益，但可能记忆训练轨迹，难以把互动技能迁移到新的任务和场景组合。

Result: Target-adaptive Cross-View Encoder 根据目标掩码调节多视角特征并进行跨视图信息交换；PlanDiT 预测面向目标的轨迹和动作，Transition Head 判断子任务完成以触发自动切换。Reference-based Retargeting Strategy 利用已有轨迹为不同目标区域合成平滑接近动作，无需新增数据。在新建 LIBERO-Analogy 和真实桌面任务上，作者报告 RAIN 比所比较 VLA 基线具有更稳健的泛化。摘要未列具体成功率或提升幅度。

Insight: 目标区域掩码让子任务技能从具体物体轨迹中解耦，参考轨迹重定向也减少新增演示需求。结果包含模拟新基准和真实桌面任务，但摘要未量化各场景成绩、训练示范数量或未见组合的成功率。

Abstract: Existing vision-language-action (VLA) models benefit from large-scale vision-language pretraining, but tend to memorize training trajectories, limiting their ability to adapt learned interactions to new task and scene configurations. To address this problem, we propose \textbf{Region-Aware Interaction Networks (RAIN)}, which formulates a task as a sequence of goal-conditioned subtasks, each specified by an action type and a target-region mask. This formulation allows learned skills to be reused across different objects and scene layouts and composed into task sequences not seen during training. Its Target-adaptive Cross-View Encoder (TCE) conditions multi-view features on target masks and exchanges information across views, while PlanDiT predicts target-directed trajectories and actions. A Transition Head estimates subtask completion to enable automatic transitions. We also introduce a \textbf{Reference-based Retargeting Strategy (RRS)}, which uses recorded trajectories to synthesize smooth approaches to reference interactions with alternative target regions, without collecting additional data. Experiments on \textbf{LIBERO-Analogy}, our new benchmark for robotic task generalization, and real-world tabletop tasks demonstrate that RAIN generalizes more robustly than existing VLA baselines.

[400] ChunkAct: Action Chunking for Agentic Robotics cs.RO | cs.AI | cs.CL | cs.CVPDF

Bingxin Xu, Wenrui Bao, Xinxin Liu, Jiawei Zhou, Yuzhang Shang

TL;DR: ChunkAct 将记忆中已知的常规原子动作编译成无需 LLM 逐步监督的动作块，仅在决策或执行失败时将控制权交回规划器。

Details

Motivation: 智能体机器人每执行一个原子动作都要让 LLM 根据完整交互历史重新推理，episode 成本随动作数增长；作者观察到子目标确定后，很多动作序列已由记忆中的常规程序预先规定。

Result: ChunkAct 在线通过一次调用启动一组串联原子动作，动作块自主执行并仅在失败时返还控制；离线将记忆编译为动作块，使测试执行阶段无需持续读取记忆，规划器只处理需推理的决策。LLM 计算因此随决策数而非原子动作数增长。在 LIBERO-Pro 上，任务成功率比当时 agentic robotics 最佳结果高 5.0 分，同时 LLM 调用减少 2.9 倍、输入 token 减少 4.4 倍、输出 token 减少 2.3 倍。

Insight: 将常规子目标执行编译成可恢复动作片段，可减少重复调用 LLM；失败触发返还控制保留了异常处理通道。摘要未给出不同失败类型下的恢复率、动作块覆盖范围及低层控制安全性。

Abstract: Agentic robotics places a large language model (LLM) agent in the robot control loop, where it composes perception, motion, and vision-language-action (VLA) tools under the guidance of a memory and thereby generalizes to new tasks without retraining. This generality comes at a steep price in efficiency. Every primitive requires a round of LLM reasoning over the full interaction history, so the cost of an episode grows with the number of primitives. We find this reasoning largely redundant: once a sub-goal is chosen, its primitives follow a routine already prescribed by memory. We present ChunkAct, which introduces action chunking to agentic robotics. Online, a single call launches a chunk of chained primitives, which executes autonomously and returns control only upon failure. Offline, ChunkAct compiles the memory into these chunks, so execution is memory-free at test time and the planner is reserved for decisions that require reasoning. LLM computation therefore scales with the number of decisions rather than the number of primitives. On LIBERO-Pro, ChunkAct improves task success by 5.0 points over the state of the art in agentic robotics while requiring 2.9x fewer LLM calls, 4.4x fewer input tokens, and 2.3x fewer output tokens.

[401] Disentangling Task-Relevant Latent Dynamics for Reward Modeling cs.RO | cs.CVPDF

Dayou Li, Jionghao Wang, Peihao Wang, Jiuzhou Lei, Hao Wang

TL;DR: DiTaR 从多帧视频潜变量学习连续的任务相关表示，再用历史和语言条件的 rectified-flow 模型描述任务动态，并以动态兼容性变化作为稠密奖励。

Details

Motivation: 从视频学习机器人奖励可减少手工稠密奖励工程，但通用视觉空间会混淆任务相关状态变化和无关视觉差异，使奖励可能响应背景干扰而非任务执行。

Result: DiTaR 以非对称两阶段过程抑制任务无关视觉变化，学习任务相关连续潜表示；在此空间训练历史及语言条件的 rectified-flow 动力学模型，并根据时间变化的动态兼容性构造方向性稠密奖励。在 MetaWorld-v3 和 ManiSkill 上，作者报告策略学习优于视频奖励基线；将奖励重新标注用于真实世界离线 RL 时也能迁移。奖励分析显示其给分更可靠。摘要未提供具体成功率或迁移数值。

Insight: 通过先解耦任务相关状态，再预测条件动力学，奖励可关注行为是否朝目标演化，而非单纯视觉相似。真实迁移使用离线 RL 奖励重标注，摘要未说明真实机器人规模、任务结果或与人工奖励的比较。

Abstract: Reward design remains a major bottleneck for reinforcement learning in robotic manipulation. Handcrafted dense rewards can compensate for the limited guidance provided by sparse rewards, but require substantial task-specific reward engineering. Learning rewards from videos provides a scalable, data-driven alternative to manual reward design. However, existing video-based rewards are often computed in entangled visual spaces where task-relevant state changes are confounded with extraneous visual variation. As a result, the learned reward may respond to visual distractors instead of actual task execution. To address this limitation, we propose DiTaR (Disentangling Task-Relevant Latent Dynamics for Reward Modeling), a framework for learning dense rewards from heterogeneous videos. DiTaR learns a continuous task-relevant representation from multi-frame video latents through an asymmetric two-stage procedure that suppresses task-irrelevant visual variation. In the task-relevant latent space, we train a history- and language-conditioned rectified-flow model to capture task dynamics and derive a directional dense reward from temporal changes in dynamics compatibility. Together, these components enable reward modeling directly over task-relevant dynamics rather than generic visual variation. DiTaR substantially improves policy learning on MetaWorld-v3 and ManiSkill over video-based reward baselines, and transfers to real-world offline RL through reward relabeling. Reward analyses further show that DiTaR provides more reliable rewards.

Renbin Shu, Jing Yang, Xiangyuan You, Guolin Qiu, Yufei Xia

TL;DR: Axiom-0 通过中间表示一致性精炼和动作生成前的条件递归更新，改善 VLA 在陌生任务组合上的操作成功率。

Details

Motivation: VLA 结合预训练多模态理解与控制能力，但在不熟悉的任务组合上仍受表示质量及其传递到动作生成方式的限制。

Result: IVRR 在训练时用学习到的特征投影一致性监督强化选定中间表示；VAIR 在动作生成前递归更新条件。VLM 监督层、VAIR 和动作生成器联合优化。在 RoboCasa365 上，Axiom-0 总体成功率为 62.32%，高于已发表最高榜单分数 Paimon-0 的 58.1%（高 4.22 个百分点），也比 Xiaomi-Robotics-1 参考系统高 4.92 点。在作者评测的 profiling 设置下，平均动作块生成延迟仅比 Xiaomi-Robotics-1 高 1.7%。

Insight: 同时改善 VLM 内部表示及表示到动作模块之间的条件接口，针对理解—控制落差而非仅扩大动作模型。62.32% 是 RoboCasa365 评测结果，延迟差仅在指定 profiling 设置测量，摘要未给真实机器人验证。

Abstract: Reliable manipulation requires policies that generalize across tasks, layouts, and object interactions. Vision-language-action (VLA) models combine pretrained multimodal understanding with learned control, but their performance remains limited on unfamiliar task compositions. We present Axiom-0, a VLA policy that addresses both the quality of its representations and their use in action generation through hierarchical self-refinement. Intra-VLM Representation Refinement (IVRR) strengthens selected intermediate features through consistency supervision with learned feature projections during training. VLM-Action Interface Refinement (VAIR) recursively updates the resulting conditioning before action generation. Joint optimization of the supervised VLM layers, VAIR, and the action generator connects the two components. On RoboCasa365, Axiom-0 achieves 62.32% overall success in our evaluation, exceeding the highest published leaderboard score (Paimon-0, 58.1%) by 4.22 percentage points and the Xiaomi-Robotics-1 reference by 4.92 points. Under the evaluated profiling setup, mean action-chunk generation latency increases by 1.7% over Xiaomi-Robotics-1. These results demonstrate improved task completion with modest inference overhead.

[403] ClothCode: Generative Modeling for Garment State Estimation and Dynamics Prediction cs.RO | cs.CVPDF

Yingdong Ru, Lipeng Zhuang, Shiyu Fan, Zhuo He, Florent P. Audonnet

TL;DR: ClothCode 用材质对齐的残差向量量化 token 表示衣物状态，并以掩码生成 Transformer 并行预测状态和动作条件下的动力学变化。

Details

Motivation: 柔性衣物状态估计与动作后预测对机器人操作重要，需要共享表示连接当前构型和未来动态，同时保持推理效率。

Result: 框架在统一离散构型空间中建模衣物状态及动作条件动态。三个衣物模板上，相较最强基线，状态估计平均位移误差降低 26.6%，动力学预测降低 37.0%。动力学预测每窗口仅需 16.7 ms，对比扩散模型 556.2 ms。在相同模型预测控制器中，真实折叠和铺平成功率从 56.7% 提高到 78.9%。

Insight: 共享离散服装表示可同时支持状态估计和条件未来预测，速度优势使其能融入实时控制。真实实验覆盖折叠与铺平，但摘要未报告硬件、衣物模板变化和长时多步预测误差。

Abstract: We introduce ClothCode, a unified generative framework for garment state estimation and action-conditioned dynamics prediction in a shared discrete configuration space. A material-aligned tokenizer encodes garment configuration as residual vector-quantized patch tokens, while task-specific Masked Generative Transformers predict these tokens through parallel confidence-guided refinement. Across three garment templates, ClothCode reduces mean displacement error by 26.6% for state estimation and 37.0% for dynamics prediction relative to the strongest baselines. Dynamics prediction requires only 16.7 ms per window, compared with 556.2 ms for diffusion. Integrated into the same model-predictive controller, ClothCode improves real-world folding and flattening success from 56.7% to 78.9%. Further results and videos are available at: https://anonymous.4open.science/r/ClothCode-DE32

Yuxuan Liu, Yating Luo, Yunfei Luan, Xinyao Zhou, Yuyang Zhuge

TL;DR: 该方法将手术器械语义识别和连续运动估计作为互补线索，与历史轨迹共同融合，用于显微吻合操作的长时 2D/3D 轨迹预测。

Details

Motivation: 显微吻合要求在狭小空间精准操纵器械；轨迹预测对场景理解、术中规划和机器人控制重要，但既有方法多只利用视频特征和历史轨迹，忽略器械—组织交互的语义及运动信息。

Result: 框架包含多模态表示提取器和跨模态融合网络，将手术对象语义识别、连续运动估计作为互补视觉引导。粗到细渐进融合结合 LSTM 逐步捕获时序依赖。作者在自建机器人平台及覆盖仿体和离体手术场景的数据集上训练评测，定量和定性结果均称优于既有方法，并对不同手术场景表现出更好的鲁棒性。摘要未报告数据规模、预测时域、误差指标或提升幅度。

Insight: 同时建模对象语义和连续运动可为器械—组织互动提供空间与时间线索，优于只依赖历史视觉/轨迹的设计。摘要结论限于自建平台及仿体、离体场景，未提供真实临床手术验证或具体精度。

Abstract: Microsurgical anastomosis is one of the most challenging and sophisticated procedures for surgical robotics, requiring accurate manipulation of surgical instruments within a limited space to achieve precise operations. Accurate surgical trajectory anticipation is essential for surgical scene understanding, intraoperative planning, robotic manipulation, and reliable behaviour analysis. However, existing approaches mainly focus on features extracted from surgical videos and historical trajectories, while overlooking rich semantic and motion information that can provide more discriminative multi-modal representations for complex surgical instrument-tissue interactions. In this work, we propose a cross-modal feature fusion framework for long-term 2D and 3D microsurgical anastomosis trajectory anticipation. The proposed framework consists of a multi-modal representation extractor and a cross-modal feature fusion network. Semantic recognition of surgical objects and consecutive motion estimation are incorporated as multi-modal visual cues, providing complementary spatial and temporal guidance for instrument-tissue interaction characterization. A coarse-to-fine progressive fusion strategy with a long-short-term memory mechanism is introduced to capture temporal dependencies step by step. The training and evaluation are conducted on our established robotic platform and a comprehensive dataset containing both phantom and ex-vivo surgical scenarios is constructed. Extensive quantitative and qualitative experiments demonstrate that the proposed framework consistently outperforms existing methods and exhibits improved robustness across different surgical scenes, highlighting its potential for both intelligent robot-assisted and autonomous surgical systems.

[405] Flow-Support-Weighted Proposals for Test-Time Verified VLA Policies cs.RO | cs.CVPDF

Chen Zhao, Zipeng Ye, Haitao Shen, Guanlin Li, Yuchen Xie

TL;DR: ADS 在训练期按 flow 支持度对当前 VLA 策略采样出的轨迹加权，以扩展可供验证器选择的动作候选多样性，且不增加部署时成本。

Details

Motivation: 测试时多次采样可让验证器从多种可执行动作中选择，但 VLA 微调可能使 flow-matching 解码器集中到主导演示动作，重复采样高度相似，候选预算增大也未必带来更多选择。

Result: Action Dynamic Sampling 每步从当前策略在新基础噪声下积分 K 条额外轨迹，用 Hutchinson trace estimator 计算各轨迹的 flow 支持度，并通过停止梯度的有限集合 KL 项调整局部动作分布。训练成本增加，部署成本不变。在模拟与四项真实机器人任务上，ADS 提高候选分散度而不产生高斯扰动的成功率惩罚；RoboCasa oracle best-of-8 成功率从 57.6% 提至 63.0%，固定验证器在模拟中提高 2.4–4.9 点，真实操作 TokenPB 最高提高 13.0 点。

Insight: 提升 test-time scaling 效果的关键可能在训练时形成不冗余的候选分布，而非单纯增加采样次数。oracle best-of-N 与固定验证器结果不同，后者更接近可部署选择效果；摘要未列每项实体任务成功率和新增训练成本。

Abstract: Test-time sampling promises to turn one action prediction into a set of distinct executable candidates that a verifier can choose among. Downstream fine-tuning works against that promise. A vision-language-action model trained to solve one task distribution concentrates its flow-matching decoder around the dominant demonstrated action, so repeated samples differ only marginally and the verifier receives many copies of one decision. We call this proposal contraction and identify it as a key constraint on test-time scaling: repeated samples can remain highly redundant as the candidate budget grows. Action Dynamic Sampling (ADS) alleviates the constraint on the proposal side, during training. At each training step, ADS integrates $K$ additional trajectories from the current policy under fresh base noise, scores each by the flow support along its path with a Hutchinson trace estimator, and corrects the policy’s local action distribution toward this weighted set through a detached finite-set KL term. Training cost rises; deployment cost is unchanged. Across simulation and four physical-robot tasks, ADS raises proposal spread without the success penalty that Gaussian perturbation incurs, lifts oracle best-of-8 success on RoboCasa from 57.6% to 63.0%, improves fixed verifiers by 2.4 to 4.9 points in simulation, and raises TokenPB by up to 13.0 points on real-world manipulation.

[406] Remembering What Matters for Long-Horizon Robotic World Models cs.RO | cs.CVPDF

Hao Wu, Tianyu Shao, Shengju Qian, Yizhou Zhao, Fan Zhang

TL;DR: ReCAP 在冻结的机器人视频世界模型中保留初始场景锚点和近期完整视觉—动作块，淘汰更旧生成历史，以缓解长 rollout 的上下文记忆冲突。

Details

Motivation: 自回归世界模型即使保留全部生成历史也不一定预测更好；持久场景信息、近期动作状态和陈旧帧可能争夺注意，导致长时生成退化。

Result: 训练免除的 ReCAP 保持 tokenizer、世界模型和视觉解码器冻结，按滑动上下文日程保留初始锚点和连续近期块、淘汰较旧历史；保留块内结构与时间顺序，且活跃上下文增长不依赖 rollout 长度。在 RT-1、BridgeV2、LIBERO-90、CALVIN 的冻结自回归世界模型评测中，视觉保真、动作条件一致性和机器人结构保持均优于 Full Context。BridgeV2 的 LPIPS 降低 63.3%；H=64 系统基准中，逻辑 KV 占用降低 73.3%，p50 延迟下降 6.51%。注意力分析显示模型更多关注保留的近期状态，但作者指出该现象尚不能与上下文长度和位置处理效应分离。

Insight: 完整历史可能引入陈旧信息而损害预测，保留场景锚点及近期完整状态提供了简单推理期日程。性能变化与上下文截断、位置效应也可能相关，作者未声称已单独证明记忆角色冲突假说。

Abstract: Autoregressive robot video world models based on discrete visual tokens remain unreliable over long rollout horizons. Retaining the entire generated history does not necessarily improve prediction quality. We investigate a possible memory-role conflict in Full Context decoding: persistent scene information, recent visual-action states, and stale generated history compete for attention despite their different predictive roles. This perspective motivates ReCAP (Receding Context with Anchor Preservation), a training-free context-scheduling method for long-horizon robot video world models. ReCAP keeps the pretrained tokenizer, world model, and visual decoder frozen. It preserves the initial scene anchor and a contiguous window of recent complete visual-action blocks while evicting older generated history. The schedule preserves within-block structure and temporal order, requires no additional modules or parameter updates, and bounds active-context growth independently of rollout length. We evaluate ReCAP on RT-1, BridgeV2, LIBERO-90, and CALVIN using frozen autoregressive world models. Across evaluated environments and horizons, ReCAP improves visual fidelity, action-conditioned consistency, and robot structure preservation relative to Full Context. On BridgeV2, it reduces LPIPS by 63.3%. In the system benchmark at H=64, it reduces the logical KV footprint by 73.3% and p50 latency by 6.51%. Attention analyses show increased attention to retained recent states, while history interventions associate larger amounts of stale context with higher prediction error. These observations are consistent with the memory-role-conflict hypothesis, although they do not isolate it from context-length and position-handling effects. Our results support inference-time context scheduling as a practical approach to improving long-horizon robot video prediction and computational efficiency without retraining.

[407] Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning cs.RO | cs.CVPDF

Seongmin Park, Jaewon Lee, Yujun Heo, Yoonsik Kim, Seheon Choi

TL;DR: QCRL 针对低比特 VLA 的失败，在离线强化学习中结合进度优势估计和量化条件化参考动作选择，利用全精度 rollout 超越单纯动作蒸馏。

Details

Motivation: 量化可降低 VLA 部署成本，但 PTQ 和 QAT 在长程任务、视觉/语义分布偏移下仍与全精度存在差距，且仅模仿目标不直接利用 rollout 成功或部分完成信号。

Result: QCRL 用全精度 rollout 提供超出动作级蒸馏的参考经验，以随进度缩放的终局惩罚区分部分完成，并根据量化条件从参考行为中选择有利策略。评估包含 π0.5 和 GR00T N1.5 的权重 INT4 W4A16，以及 π0.5 的权重/激活 NVFP4 W4A4。仅用分布内数据时，π0.5 的 INT4 W4A16 在三个视觉偏移条件上的平均分达到全精度的 99.3%，模型内存节省 42.9%；NVFP4 W4A4 达到同一全精度基线的 100.7%，以 NVFP4/FP8 部署相较 Jetson Thor 上 FP16 加速 1.73 倍。

Insight: 进度信号让部分成功也能影响低比特策略优化，而条件化参考行为提供量化策略可学习的较好轨迹。报告性能为全精度均分比例，硬件收益比较基于 Jetson Thor FP16；摘要未给各任务成绩和失败案例。

Abstract: Quantization lowers deployment costs of vision-language-action (VLA) policies but can compromise task completion. Across five VLA models and 4,050 real-world trials, we systematically reveal that quantization-aware training (QAT) generally improves on post-training quantization (PTQ) but leaves substantial gaps to full precision (FP) on long-horizon tasks and under visual and semantic shifts, while its imitation-based objective does not explicitly use rollout success or partial completion. To exploit these outcome signals, we introduce Quantization-Conditioned Reinforcement Learning (QCRL), an offline RL method coupling progress-based advantage estimation with quantization-aware source conditioning. FP rollouts supply reference experience beyond action-level distillation, progress-scaled terminal penalty distinguishes partial completion, and quantization-aware conditioning selects advantageous reference behavior within the quantized policy. We evaluate weight-only INT4 W4A16 on $π_{0.5}$ and GR00T N1.5 and joint weight-activation NVFP4 W4A4 on $π_{0.5}$. Using only in-distribution data, QCRL reaches 99.3% of FP’s mean score across three visual-shift conditions for $π_{0.5}$ with INT4 W4A16, alongside 42.9% model-memory savings. With NVFP4 W4A4, it reaches 100.7% of the same FP baseline, alongside a $1.73\times$ speedup using NVFP4/FP8 deployment. Hardware gains are relative to FP16 on Jetson Thor.

[408] BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries cs.RO | cs.CVPDF

Jiaxin Fang, Ruitong Liu, Zeyu Lai, Xingyu Wang, Mingyuan Jia

TL;DR: BridgeWAM 用 32 个潜桥接查询聚合视觉、语言和本体感觉先验，在视频专家与轻量动作专家之间选择性传递任务相关信息。

Details

Motivation: 密集逐层 MoT 注意力会把无关视觉特征带入动作预测，并强制视频/动作专家深度匹配，造成信息利用不选择及冗余计算。

Result: Bridge-of-Experts 的 BridgeWAM 将信息压缩至 32 个 Latent Bridge Queries；联合视频—动作监督使动作梯度调整视频专家表示，动作专家再以交叉注意力选择性读取相关信息。模型总计 5.06B，视频专家 30 层、动作专家 2 层。LIBERO 成功率为 98.25%，LIBERO-Plus 汇总成功率 72.24%。相较 Fast-WAM，模拟 profiling 单样本 FLOPs 降低 28.72%，真实机器人延迟为 77.1 ms、速度快 2.47 倍；定性实验显示扰动下注意力偏向目标物体和任务目的地。

Insight: 潜查询为不同深度专家提供有限容量的信息接口，使动作学习能反向塑造视频表示，并减少动作头计算。LIBERO-Plus 72.24% 表明更复杂泛化任务仍有提升空间；摘要未列不同扰动类别的定量成绩。

Abstract: World Action Models (WAMs) commonly couple video and action experts through dense, layer-wise Mixture-of-Transformers (MoT) attention, exposing action prediction to irrelevant visual features while enforcing matched expert depths. This coupling hinders selective visual-feature use and introduces redundant computation. We introduce Bridge-of-Experts (BoE), instantiated as BridgeWAM, which aggregates visual, language, and proprioceptive priors through 32 Latent Bridge Queries (LBQs) into a shared, limited latent space. Joint video-action supervision lets action gradients guide the Video Expert toward representations better suited for action prediction, while the Action Expert learns to selectively retrieve task-relevant information through cross-attention. Decoupling expert depths and reusing LBQs during action denoising yield a 5.06B model with a 30-layer Video Expert and a two-layer Action Expert, achieving 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus, demonstrating effective task execution and strong generalization with an ultralight action head. Ablations support joint representation adaptation and unified LBQ conditioning. Relative to Fast-WAM, BridgeWAM reduces single-sample inference FLOPs by 28.72% in simulation profiling and real-robot latency to 77.1 ms (a 2.47-fold speedup). Qualitative experimental results further indicate preferential attention to target objects and task destinations across perturbations.

[409] Hand2Robot: Object-Centric 4D Hand Reconstruction from Human Demonstrations for Robotics cs.RO | cs.CVPDF

Zetao Du, Weiqing Wang, Hanbing Zhang, Yimiao Zhou, Ye Shi

TL;DR: Hand2Robot 从人类演示视频重建物体与手部 4D 交互，再通过手—物优化和运动重定向生成机器人动作监督。

Details

Motivation: 人类演示能提供任务相关操作，但手部运动不能直接成为目标机器人的动作标签；单独重建手运动也无法保证物理一致的手—物互动。

Result: Hand2Robot 提取交互片段、重建物体几何并生成机器人动作监督。物体跟踪结合观测置信度和时间运动约束，从逐帧候选中选择全局一致位姿；手部细化以物体几何和运动为参照，结合分阶段 2D 对齐及物体局部接触约束优化手轨迹和手指关节。在 DexYCB、HO3D-v3 上作者报告物体位姿精度和手—物交互质量提升；三项真实机器人灵巧操作任务显示重建并重定向的人类演示可有效补充有限机器人数据。摘要未列量化提升。

Insight: 动作监督需同时重建物体轨迹、手部运动和接触关系，再映射到机器人动作空间；只重定向手轨迹可能违背物理接触。摘要结果覆盖真实机器人任务，但未说明动作重定向成功率、额外人工校正需求或基线成绩。

Abstract: Human demonstrations can capture task-relevant manipulation without using the target robot, but the recorded hand motions cannot directly serve as robot action supervision. Hand reconstruction and motion retargeting provide a direct route to such supervision, yet recovering hand motion alone does not ensure physically consistent hand-object interactions. To address this issue, we introduce Hand2Robot, an object-centric 4D hand reconstruction framework for robot policy post-training. Given videos of humans performing the target task, Hand2Robot extracts interaction clips, reconstructs object geometry, and generates robot action supervision through hand-object interaction optimization and motion retargeting. For object tracking, we combine observation confidence with temporal motion constraints to select a globally consistent object pose trajectory from per-frame pose candidates. For hand refinement, we use object geometry and motion as a reference and combine staged 2D alignment with object-local contact constraints to refine hand trajectories and finger joint configurations. Experiments on DexYCB and HO3D-v3 show improved object pose accuracy and hand-object interaction quality. Experiments on three real-robot dexterous manipulation tasks further show that reconstructed and retargeted human demonstrations can effectively supplement limited robot data for target-task policy post-training.

[410] Language2Drive: Do Vision-Language-Action Models Really Follow the Instructions in Autonomous Driving? cs.RO | cs.CVPDF

Zihan You, Zehua Fan, Wenqi Lyu, Aoqi Wang, Ni Wang

TL;DR: Language2Drive 用初始场景相同、只改变指令的配对反事实评测，检查自动驾驶 VLA 是否真的按乘客语言选择可行行为。

Details

Motivation: 闭环驾驶基准通常看路线完成、安全和规则遵守，但车辆即便安全到达，也可能忽略乘客具体指令；孤立指令成功无法区分指令跟随与模型始终执行偏好的固定行为。

Result: 基准的 paired counterfactual unit 固定初始场景、只改变指令，并涵盖四种指令范围；评测结合闭环 CARLA、重建真实场景和开环驾驶基准，发布专用 CARLA 模拟器。对代表性驾驶 VLA 的评估发现，在成功 rollout 中，93.2% 没有在替代可行指令下对应成功的反事实轨迹；主要失败为单侧完成，即同一模型反复执行一种偏好行为，却无法在同场景按反事实指令执行另一行为。

Insight: 驾驶能力、单条轨迹成功和语言控制能力是不同指标；配对反事实测试揭示传统单次成功率会高估指令遵循。93.2% 是成功 rollout 未具备另一指令成功配对的比例，摘要未提供不同指令类型的具体率。

Abstract: Recent Vision-Language-Action (VLA) models increasingly use natural-language instructions to condition autonomous driving behavior. Yet existing closed-loop benchmarks primarily measure route completion, safety, and rule compliance, leaving a fundamental question unresolved: does language actually determine which feasible behavior the vehicle executes? A trajectory may be safe and successful while still violating or even ignoring the passenger’s instruction. To address this, we introduce \ltd, a benchmark designed to directly test this capability. Its key evaluation unit, the paired counterfactual unit (PCU), holds the initial driving scene fixed while changing only the instruction. This design enables us to distinguish isolated instruction success from counterfactual paired success across four instruction-following scopes. To test whether this distinction persists across evaluation regimes, we instantiate Language2Drive in complementary closed-loop CARLA simulation, reconstructed real-world scenes, and open-loop driving benchmarks, including our released CARLA-Language2Drive simulator with purpose-built maps. Evaluating representative driving VLAs across these settings reveals a striking mismatch between strong driving competence and weak language control: 93.2% of successful rollouts do not have a successful counterpart under the alternative feasible instruction. This failure is dominated by one-sided completion, where a model repeatedly executes one preferred behavior but fails when the same scene requires the counterfactual alternative. These results expose two fundamental limitations of current evaluation: isolated success and driving scores, which are widely used by conventional driving benchmarks, can substantially overstate language-following capability, while current driving VLAs themselves lack reliable instruction following under matched counterfactual tests.

[411] UniLAP: A Unified Latent Action Policy for Cross-Embodiment Manipulation cs.RO | cs.CVPDF

Lei Yao, Qiao Feng, Mutian Tong, Qihao Yun, Lap-Pui Chau

TL;DR: UniLAP 用 flow matching 建模连续潜在动作，并从无动作视频蒸馏潜动作教师；推理时直接生成可执行动作，不依赖未来帧规划。

Details

Motivation: 离散潜动作存在粒度瓶颈；连续潜动作方法虽能细化行为，却常作为世界模型代理，仍需额外规划阶段转成执行动作。

Result: 统一连续潜在动作策略以 flow-adaptive 目标蒸馏 LAM 教师，继承动作表示同时保留多种可行动作分布。作者用无标签运动目标隔离相关运动学线索，减少潜表示受无关外观动态影响，并构建与具身无关的潜在流形以适配新平台。1.1B 参数模型用多源视频预训练，在手姿预测、单臂及人形机器人操作和真实部署任务上，报告超过潜动作和 VLA 基线。摘要未给具体指标或平台适配成本。

Insight: 连续潜空间保持细粒度多模态行为，同时省去基于未来帧的额外规划；无标签运动目标有助于跨外观迁移。摘要未报告跨平台的具体成功率、动作延迟或动作潜变量可解释性。

Abstract: Latent Action Models (LAMs) can capture manipulation priors from action-free videos, but discrete latents usually impose a granularity bottleneck, limiting their ability to represent fine-grained action. Recent continuous alternatives mitigate this, but they are typically used as proxies for world modeling and require a further planning step to produce executable actions. We thus introduce UniLAP, a unified continuous latent-action policy that models the latent space via flow matching and bypasses reliance on future frames during inference. By distilling a LAM teacher with a flow-adaptive objective, UniLAP inherits the teacher’s action representation while retaining the multimodal distribution of feasible behaviors. To decouple latent representations from task-irrelevant appearance, we extract a label-free motion target to isolate relevant kinematic cues from spurious visual dynamics. With the established embodiment-agnostic latent manifold, UniLAP can be easily adapted to novel platforms. Pretrained on multi-source videos at a compact 1.1B parameter scale, UniLAP outperforms state-of-the-art latent-action and Vision-Language-Action (VLA) baselines across hand pose forecasting, single-arm and humanoid manipulation, and real-world deployment.

[412] From Pixels to Executable Physics: Behavior-Grounded World Models for Embodied Action Consequence Prediction cs.RO | cs.CVPDF

Ruo Chen, Jiahua Pang, Yingjie Niu, Hongyi Lin, Jun Tian

TL;DR: DR-WM 从 RGB 场景重建几何，并利用真实动作响应校准的地形接触先验模拟候选动作后果，以提升四足机器人户外规划。

Details

Motivation: 户外地形和接触条件变化会使相同控制指令产生不同运动结果；视觉世界模型虽然可生成逼真未来，但常未能建模机器人动力学和地形交互，导致动作后果预测过度乐观。

Result: DR-WM 将无量纲密集地形接触先验与几何结合，并由真实动作响应数据校准；可以模拟空间变化摩擦等接触性质。给定机器人状态和候选动作序列，模型先模拟未来运动，利用相机轨迹与渲染几何指导视频世界模型，再以生成观察更新地形重建，实现无新增真实图像的迭代 rollout。在 90 个挑战情景中，结合 DR-WM 的 8B VLM 在六项评测中的五项超过未增强的 32B VLM；物理指标上也优于仅几何和统一物理参数基线。摘要未列各物理指标数值。

Insight: 从视觉场景恢复接触属性并做物理 rollout，可补足仅视觉生成模型缺乏可执行动力学的问题。8B 增强模型在五项指标超过 32B 未增强模型，但结果仍依赖校准数据和模拟物理准确性；摘要未报告真实机器人部署。

Abstract: In outdoor environments, varying terrain and contact conditions can cause identical action commands to produce different motion outcomes for quadruped robots. Predicting these outcomes is essential for reliable navigation and informed path selection. Existing video world models can generate visually plausible futures, yet often struggle to capture robot dynamics and terrain interactions, leading to overly optimistic predictions of action outcomes. We present DR-WM, a physically consistent world model that incorporates robot-terrain contact dynamics to predict the consequences of quadruped actions. Starting from an initial RGB image, our method reconstructs scene geometry and integrates dense, dimensionless priors over terrain contact properties, calibrated using real-world action-response data. The resulting terrain model supports the simulation of physics with spatially varying contact properties such as friction. Given the initial state of the robot and candidate action sequences, we simulate its future motion and use the resulting camera trajectories and rendered geometry to guide the training of a video world model. The generated future observations are then used to update the terrain reconstruction, enabling iterative model rollouts without additional real-world images. Extensive experiments demonstrate that DR-WM substantially outperforms geometry-only baselines and baselines with uniform physical parameters on key physical metrics characterizing action outcomes, while achieving the best performance among the evaluated baselines. These physically grounded predictions support more informed action selection: across 90 challenging scenarios, an 8B vision-language model augmented with DR-WM outperforms an unassisted 32B vision-language model on five of six evaluation metrics.

[413] Code2Worlds-2: Physically Plausible Agent Harness for Embodied World Simulation cs.RO | cs.CVPDF

Haoyu Zhang, Siming Li, Zeyu Zhang, Yi Zhang, Tat-Seng Chua

TL;DR: Code2Worlds-2 将自然语言生成的 4D 视觉场景编译成 MuJoCo 可执行场景，并用编译—模拟—诊断—修改循环优化真实物理参数。

Details

Motivation: 代码模型能生成动态 4D 场景，但生成物可能只是视觉效果，缺少碰撞几何、质量、摩擦、惯性和关节等可供物理引擎运行的语义，动作也可能是预先脚本化而非物理执行结果。

Result: 系统保留原双流生成流程并添加 SimReady Scene Compiler：凸分解生成碰撞网格，VLM 估计物理属性，推断支撑图，设置接触、摩擦和动力学约束，重力沉降后导出 MuJoCo MJCF。Simulator-in-the-Loop 4D Refiner 自动循环编译、模拟、批评和修改物理设定；模拟器检查穿透深度、碰撞、稳定性、接触力、关节限制和机器人可达性。作者提出五种任务模板的状态式 SimCode4D 基准；实验覆盖自然、室内、工业、城市和服务场景，报告生成场景具物理合理性、模拟器兼容性和机器人交互能力。摘要未给具体指标。

Insight: 把生成目标从“看起来动态”推进到可装载、可步进的可执行物理世界，并以模拟器反馈修订参数而非只改代码。摘要未量化物理一致性、机器人成功率或与原版比较，状态基准及模拟环境范围也需结合正文理解。

Abstract: Recent coding large language models (LLMs) can synthesize dynamic 4D worlds from natural language, but the resulting content remains a visual artifact. It lacks physical semantics such as the collision geometry, mass, friction, inertia, and joint structure, required to enable it to simulate the physical world, and its dynamics are scripted to look plausible rather than designed to execute. What is needed is an executable world: one that a physics engine can load and step, resolving contact physically. We present Code2Worlds-2, an extension of Code2Worlds that turns 4D world generation into an embodied world simulation. We keep the original dual-stream generation pipeline intact and append two modules. First, a SimReady Scene Compiler that turns a generated visual scene into a simulator-compatible representation: it assigns collision meshes via convex decomposition, estimates physical attributes with a vision-language model, infers a lightweight support graph, instantiates the engine’s contact, friction, and dynamics constraints, and applies a gravity settling pass, before exporting the scene as MuJoCo MJCF. Second, a Simulator-in-the-Loop 4D Refiner that implements an automated agent harness-a compile, simulate, critique and refine loop-that converts visual dynamics into executable rigid and articulated dynamics and evolves the physical settings rather than only the code, replacing the purely visual motion critic with a simulator critic that inspects penetration depth, collision pairs, stability, contact forces, joint limits, and robot reachability. We further propose SimCode4D, a state-based benchmark with five task templates for evaluating embodied simulation. Experiments on SimCode4D and across natural, indoor, industrial, urban, and service scenes show that Code2Worlds-2 produces physically plausible, simulator-compatible, and robot-interactable scenes.

[414] DexStitch: Learning and Composing Dexterous Skills from Human Videos via Contact-Aware Robotic Inpainting cs.RO | cs.CVPDF

Byeonggyeol Choi, Seongmin Jung, Unghui Nam, Youngbo Song, Sunhong Min

TL;DR: DexStitch 从人类长程操作视频中切分技能及过渡，利用手指接触和遮挡顺序重建机器人兼容动作，再将技能拼接成目标序列。

Details

Motivation: 长时操作任务由基础技能及其转换组成，机器人演示两者采集成本高；人类视频规模更大，但需转换成机器人观测—动作，并保持手物接触结构。

Result: 技能分类器先分割技能与过渡，再将演示转成带技能标签的机器人数据；逐指人—物接触引导灵巧手重定向和机器人补绘，保留接触几何及手指—物体遮挡顺序。共享技能条件策略先在合成数据预训练，再用合成数据和少量单技能遥操作片段微调；当前/下一技能条件引导过渡，并用预测进度闭环切换。在留出录制场次上，技能分类与人工标注一致率 92.4%；四种家庭技能上，单技能成功率 80.0%，序列成功率 77.5%。

Insight: 接触约束和逐指遮挡排序使人类演示转换不只复制手部轨迹，也尽量保留物理互动；技能进度预测支持闭环串联。四种任务的序列成功率略低于单技能，摘要未说明技能数量扩展或不同机器人平台迁移。

Abstract: A long-horizon manipulation task consists of primitive skills and the transitions connecting them. Learning skills that can be executed in sequence requires supervision for both, which is costly to collect on a robot. Human videos offer a scalable source of demonstrations, but transferring them to robots requires extracting robot-compatible observations and actions together with their skill structure. We present DexStitch, a framework that cuts long-horizon human videos into skills and transitions, converts them into aligned robot observation-action-skill data, and stitches the learned skills into a requested sequence. A skill classifier first segments the videos into manipulation skills and transitions. DexStitch then converts the human demonstrations into robot observations and actions while retaining their corresponding skill labels. During conversion, per-finger human-object contacts guide dexterous hand retargeting and robotic inpainting, preserving contact geometry and finger-object occlusion ordering. To stitch the learned skills into a sequence, a shared skill-conditioned policy is pretrained on the resulting synthetic data and finetuned on a mixture of synthetic data and a small set of single-skill teleoperation episodes. Conditioning on the current and next skills guides each transition toward the following skill, while predicted skill progression enables closed-loop switching. On a held-out session, the classifier agrees with manual annotations on 92.4% of frames. Across four household skills, DexStitch achieves 80.0% single-skill success and 77.5% sequence success.

[415] Beyond the End-Effector: Learning Whole-Arm Contact Manipulation from Human Demonstrations cs.RO | cs.CV | cs.LGPDF

Bo-Ruei Huang, Daniel Seita, Erdem Bıyık

TL;DR: WACL 从人类裸手演示学习全臂接触轨迹，并用运动学优化器将语义关键点、接触轨迹和末端安全约束落到具体机器人连杆与关节。

Details

Motivation: 机器人操作常以末端执行器为中心，而人类会用整个手臂接触物体以提高效率和鲁棒性；机器人需决定哪段手臂何处接触、如何形成非抓取互动，同时满足末端任务或安全限制。

Result: WACL 用三个与机器人无关的表示描述任务：场景语义关键点、指定手臂接触位置/方式的参考轨迹，以及由 VLM 从任务描述生成的末端安全路径约束。准静态表面接触优化器根据具体运动学结构联合搜索接触连杆、接触位置和关节构型。每个任务只需 10 段无需遥操作的人类裸手示范学习接触轨迹；改变末端约束只需修改优化器输入，不必重训。7-DoF 机械臂四项真实全臂接触任务中，在末端约束下 WACL 成功率 78%，强 link-pose 基线为 33%，且可无重训适应新约束。

Insight: 将任务表示与机器人运动学解算分开，使同一接触示范可适配不同末端要求；只用少量非遥操作演示也能学习。实验证据为一台 7-DoF 机械臂四项任务，摘要未说明周期时间、人体安全或更广机器人形态迁移。

Abstract: Robot manipulation has traditionally relied on end-effector-centric interactions. In contrast, humans routinely exploit whole-arm contacts to improve efficiency, robustness, and adaptability, especially when the hands are occupied or when precise grasping is unnecessary. Enabling similar capabilities in robots remains challenging: the robot must select where on its arm to make contact, generate effective non-prehensile interactions, and satisfy task-specific pose or safety constraints on the end-effector. We propose Whole-Arm Contact Learning (WACL), a hierarchical framework that describes each whole-arm contact task with three robot-agnostic representations and hands them to a quasi-static surface-contact optimizer: semantic keypoints that summarize the scene, a reference contact trajectory that specifies where and how the arm presses on the object, and safety-aware path constraints that a vision-language model generates from the task description to protect the end-effector’s task. The robot’s kinematic chain enters only in this optimizer, which grounds these representations onto the arm by jointly searching over which link makes contact, where on that link, and the joint configuration that realizes it. This decomposition lets WACL learn the contact trajectory from 10 bare-hand human demonstrations per task collected without teleoperation, while new end-effector constraints change only the optimizer’s input rather than any trained component. Real-world experiments on a 7-DoF arm across four whole-arm tasks show that WACL sustains contact under end-effector constraints where a strong link-pose baseline fails (78% vs. 33% success) and generalizes to new constraints without retraining. Website: https://wacl-robot.github.io

[416] Data-Free Continual Learning for World Action Model with Rectified Generative Replay cs.RO | cs.CVPDF

Kaipeng Fang, Hangxu Liu, Chaojian Shi, Yunsong Zhou, Hongzi Zhu

TL;DR: R²-WAM 通过智能体评估器截断完成后的无效 rollout，并过滤动作与视觉不一致的自生成轨迹，实现无需旧训练数据的世界动作模型持续学习。

Details

Motivation: 视频生成预训练的世界动作模型可快速适配新机器人任务，但顺序微调会遗忘旧任务；传统经验回放需要保存过去数据，而机器人策略常仅以权重形式部署。

Result: 作者发现 WAM 能自回归重生成既学任务的完整轨迹，但原始回放不如真实数据，原因包括模型无法识别任务完成而继续生成无目的动作，以及开环动作逐渐偏离生成观察。Rectified Generative Replay 用智能体评估器在任务完成时截断轨迹，并筛除动作—观察物理不一致的 rollout。在四个 LIBERO 套件和五项真实任务上，相较顺序微调，遗忘减少超过 90%，且不访问旧任务数据时接近经验回放。分析发现遗忘集中在逆动力学：旧任务视觉仍合理，但生成动作失败。

Insight: 自生成数据可支持无数据访问的持续学习，但必须修正完成截断和动作—观察一致性，否则会回放错误经验。摘要称接近 ER 而非完全等同；真实任务及各套件具体成功率未提供。

Abstract: World action models (WAMs) built on pretrained video generation models adapt quickly to new robotic tasks, but they forget previously learned tasks under sequential fine-tuning. Experience replay (ER) mitigates such forgetting by rehearsing data of previous tasks, yet this data is often unavailable in robotics, where a policy is deployed as weights only and its training data is not shared along with it. In this work, we ask whether a WAM can achieve data-free continual learning by regenerating its own past. Our key observation is that a WAM can indeed autoregressively regenerate complete rollouts of previously learned tasks and replay them to mitigate forgetting. However, replaying such raw rollouts still lags far behind replaying real data, which we attribute to two issues. First, the WAM is unable to recognize task completion and keeps generating purposeless motion after the task is done. Second, under open-loop generation, the generated actions gradually drift away from the generated observations. To address both issues, we present rectified generative replay (R$^2$-WAM), which truncates generated rollouts at task completion with an agentic evaluator and filters out rollouts whose actions are physically inconsistent with the observations. On the four LIBERO suites and five real-world tasks, R$^2$-WAM reduces forgetting by over 90% relative to sequential fine-tuning and nearly matches ER without accessing any data from previous tasks. Our analysis further reveals that forgetting in WAMs concentrates in inverse dynamics modeling: a forgotten WAM still produces visually plausible observations of previous tasks while its generated actions fail.

[417] Perception Beyond the Visible: Incompleteness-Aware Scene Understanding for Robot Autonomy cs.RO | cs.CVPDF

Rohit Mohan

TL;DR: 该研究愿景提出不完整感知框架，从证据缺失、知识缺失和传感器可靠性不足三个方面组织机器人场景理解问题，并规划面向信念的 4D 表示。

Details

Motivation: 机器人需基于局部、噪声或退化观测理解环境；只建模可见区域、固定语义类别并假定传感器可靠的系统，会在遮挡、新物体和传感器退化时失效。

Result: 作者概述其在 amodal 感知、语义占用预测、开放世界场景理解、标签高效学习及不确定性感知多模态融合方面的研究贡献，并提出未来联合表示几何、语义、运动和不确定性的信念中心 4D 场景表示；同时讨论通过交互降低不确定性的主动感知与安全适配。本文属于研究愿景，未报告统一新实验或量化结果。

Insight: 将不完整证据、开放类别知识和传感器可信度作为互相关联的问题，可让场景表示同时包含“已知/未知”及其置信度。文中目标是研究方向总结，不能视作单一方法或已有端到端系统验证。

Abstract: Reliable robot autonomy requires reasoning about the world from partial and noisy observations. However, many perception systems model only visible scene regions, rely on fixed semantic categories, and assume reliable sensor measurements. These assumptions break under occlusion, novel object encounters, and sensor degradation. In this paper, I present a research vision for incompleteness-aware scene understanding. I organize this vision around three fundamental challenges: incomplete evidence, incomplete knowledge, and incomplete sensor reliability. I discuss my research contributions toward addressing these challenges through amodal perception, semantic occupancy forecasting, open-world scene understanding, label-efficient learning, and uncertainty-guided multi-modal fusion. Building on these contributions, I outline future research directions toward belief-centric 4D scene representations. These representations aim to jointly model geometry, semantics, motion, and uncertainty. I also discuss active sensing and safe adaptation strategies that allow robots to reduce uncertainty through interaction. The goal is to develop calibrated and actionable scene representations that support reliable autonomy in dynamic, open-world environments.

[418] EgoExplorer: A Vision-Motion-Touch Observatory for Human Loco-Manipulation In Situ cs.RO | cs.CVPDF

Yunsong Zhou, Xuekun Jiang, Haiqi Shen, Hangxu Liu, Chaojian Shi

TL;DR: EgoExplorer 发布 100 小时穿戴式视觉—运动—触觉人类全身操作数据，并提出稀疏身体传感下的校准、运动补全和几何约束流程。

Details

Motivation: 人形机器人从人类演示学习全身移动与操作，需要同步观察步态、姿态和物体交互；现有数据集只覆盖部分模态或动作细节，而跨环境采集需要轻便可穿戴设备。

Result: 数据将第一人称视频与重建全身运动、细粒度手指关节和触觉观测对齐，并提供步骤级描述，重点覆盖专业活动。Calibrate-Complete-Constrain 流程先校准参与者几何及传感器安装，再用时间先验补全未观测关节，并以测量一致性、时间连续性和几何一致性约束结果。相较 tracker 拟合参考，身体点误差降低 71.9%；匹配评测还显示手部稳定性和接触估计提升。初步实验探索多模态生成/条件控制，以及模拟中的全身控制人形机器人重定向。

Insight: 稀疏身体传感需要区分安装偏移和真实关节运动，单纯追求运动视觉合理性不足以保证测量一致；数据同步视觉、运动和触觉有助于研究完整互动。机器人结果仍是模拟初步研究，摘要未报告数据主体数和采集场景分布。

Abstract: Learning humanoid loco-manipulation from human demonstrations rests on observing how locomotion, posture, and object interaction evolve together as tasks unfold. Scaling collection across diverse environments, in turn, favors lightweight, portable capture systems. Yet, existing datasets provide only partial views of this coordination. Thus, we introduce EgoExplorer, a 100-hour vision-motion-touch observatory of human loco-manipulation built with a wearable system. With an emphasis on professional activities, its continuous recordings align egocentric video with reconstructed whole-body motion, fine-grained finger articulation, and tactile observations, accompanied by step-level descriptions. To achieve this coverage with sparse body sensing, we develop the Calibrate-Complete-Constrain pipeline. Concretely, reusable calibration fixes participant geometry and sensor attachments before motion inference to disentangle mounting offsets from articulation. A temporal prior then completes unobserved joint motion, while layered constraints promote measurement agreement, temporal continuity, and geometric consistency, properties that motion plausibility alone does not guarantee. Against tracker-fitted references, the method reduces body-point error by 71.9% relative to the strongest baseline; matched evaluations also show gains in hand stability and contact estimation. Preliminary studies explore multimodal generation and conditioning alongside humanoid retargeting with whole-body control in simulation.

[419] LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts cs.RO | cs.CVPDF

Wenxuan Guo, Yingping Liang, Zhaoguo Wang, Ziyuan Li, Angyuan Ma

TL;DR: LoopWAM 从自身及已部署策略产生的不同质量 rollout 中持续学习，用来源相关噪声、质量引导和加权监督避免盲目模仿失败动作。

Details

Motivation: 单纯增加专家演示不能保证通用机器人策略从部署中改进；真实 rollout 既有成功、次优也有失败，全部模仿会强化坏行为，删除失败则丢失物理交互及后果信息。

Result: LoopWAM 以 source-dependent 噪声采样统一行为学习与动作条件世界建模；视频和动作 token 在并行去噪过程中交互，使后果信息参与动作生成，并以质量引导和质量加权保留不同质量视频监督。RoboChallenge Table30-V2 上三个部署—训练轮次将成功率从 23.67% 提至 43.33%，提升 19.7 个百分点，最终分数 56.80。无额外具身预训练时，RoboTwin 2.0 平均成功率 94.5%，RoboCasa365 总体 49.5%，作者报告三个基准均优于比较方法。单项衬衫折叠任务经两轮部署改进后，从单任务后训练的 13.20% 提至 55.56%。

Insight: 混合质量经验需按价值加权而不是全盘模仿或丢弃，让世界后果和动作选择联合学习；部署 rollout 构成闭环改进数据。任务效果差异仍大，RoboCasa365 49.5% 显示复杂任务空间仍有显著提升余地。

Abstract: Scaling expert demonstrations alone does not enable generalist robot policies to improve from deployment. Real-world rollouts contain successes, suboptimal behavior, and failures: imitating every recorded action can reinforce poor behavior, while filtering out failures discards evidence of physical interactions and their consequences. We introduce \textbf{LoopWAM}, a closed-loop world-action model that learns directly from this full quality spectrum, using rollouts from both itself and other deployed policies. Source-dependent noise sampling unifies behavior learning and action-conditioned world modeling, while quality steering and quality-weighted supervision retain full video supervision across qualities and regulate action imitation. Video and action tokens interact throughout parallel denoising, allowing learned consequences to inform action generation directly. On RoboChallenge Table30-V2, three deployment-and-training rounds raise success from 23.67% to 43.33%, a 19.7-percentage-point gain, with a final score of 56.80. LoopWAM also achieves 94.5% average success on RoboTwin~2.0 and 49.5% overall success on RoboCasa365 without additional embodied pretraining, obtaining the best overall results among the compared methods across all three benchmarks. On challenging shirt folding task, two rounds of deployment-driven improvement after single-task post-training raise success from 13.20% to 55.56%. These results demonstrate closed-loop improvement from mixed-quality experience under both multi-task learning and single-task post-training. \href{https://loop-wam.github.io/}{\textcolor{orange}{Project Page}}

cs.SD [Back]

[420] Fish Audio Drama 3 Technical Report cs.SD | cs.CLPDF

Dawei Han, Shidong Li, Shijia Liao, Jiahua Liu, Songting Liu

TL;DR: Fish Audio Drama 3 是 350 亿参数、每 token 激活 30 亿参数的开源混合专家语音模型，支持多轮修改、多人场景配音、语音补写和流式生成。

Details

Motivation: 高表现力语音合成不仅要朗读文本，还需遵循自然语言导演指令、克隆声音、按反馈修改、控制多角色，并支持长音频及流式场景。

Result: Drama 3 基于混合线性注意力骨干，采用新的码本多 token 预测解码器，逐帧预测多个声学码本并通过专用层关注全序列。训练包括三阶段预训练、指令微调和大批量强化学习，奖励仅来自一个学习到的人类偏好模型。单张 H200 上推理引擎解码 271 帧/秒，高于 Fish Audio S2 的 177 帧/秒，尽管参数量超过其七倍。作者称多轮修订、场景配音、补写和流式文本输入是该版本新增的统一能力。

Insight: 大模型的MoE稀疏激活与码本级并行预测结合，旨在同时提供表现力和较高推理速度；速度比较是在单 H200 上进行，摘要未说明音质一致性、延迟口径或不同任务质量差异。

Abstract: We introduce Fish Audio Drama 3, an open-source speech generation model that turns text into directed performances, built on a 35B-parameter mixture-of-experts backbone that activates 3B parameters per token. It clones voices from multiple references, follows natural-language direction, revises a take when the next turn gives a note, voices multi-character scenes, edits existing speech by infilling, and synthesizes long-form and streaming audio. Multi-turn revision, scene dubbing, editing by infilling, and streaming text input are new in Drama 3, all in one model. Drama 3 pairs the hybrid linear-attention backbone with a new codebook multi-token-prediction audio decoder that predicts the acoustic codebooks of each frame in dedicated layers that attend over the entire sequence. The model is trained with a three-stage pre-training recipe, instruction fine-tuning, and large-batch reinforcement learning whose only reward is a single learned model of human preference. Its inference engine decodes a single stream at 271 frames per second on one H200, against 177 for Fish Audio S2, although Drama 3 has more than seven times as many parameters. The model weights are available on GitHub (https://github.com/fishaudio/fish-speech) and Hugging Face (https://huggingface.co/fishaudio/drama-3), and inference is available through SGLang-Omni (https://github.com/sgl-project/sglang-omni). We highly encourage readers to visit https://fish.audio to direct their own performances.

cs.SE [Back]

[421] AutoEnvScaling: Automating the Data Flywheel with Terminal Agents cs.SE | cs.CL | cs.LGPDF

Simon Yu, Baolin Peng, Bo Liu, Zhengyan Shi, Zhaoyang Wang

TL;DR: AutoEnvScaling 让提议智能体根据求解器 rollout 自动生成新的可验证终端任务环境，再用新 rollout 驱动下一轮环境生成，形成环境与求解能力共同演化的数据飞轮。

Details

Motivation: 终端智能体强化学习需要多样、可验证的环境，但环境主要由研究者手工设计，任务供给受人工速度限制。

Result: 相较仅凭提示生成，AutoEnvScaling 将有效环境比例最高提高 25 倍，并将每个有效环境成本最高降低 75%。用 GPT-5.6-sol 环境训练 Qwen3.5-35B-A3B 后，Terminal-Bench 2.1 比匹配的 Tmax 基线高 4.3 分。进一步让 Qwen3.6-35B-A3B 同时担任提议者和求解器，Terminal-Bench 2.1 提高 12.2 分，Terminal-Bench 4.0 从 1.6 升至 3.2。流程还生成软件工程、电脑使用、专业工作及 GPU kernel 编程等智能体环境。

Insight: 把环境生成变成可执行终端任务，可扩展训练题来源并让出题和求解相互反馈；同一模型兼任两种角色的实验提供递归自改进方向的初步证据。摘要未提供有效环境率绝对值、生成质量人工评估及不同角色是否存在评测污染。

Abstract: Terminal agents solve tasks in software engineering and scientific research by executing commands in a terminal. Training them via reinforcement learning (RL) requires diverse and verifiable environments, but these environments are still mostly designed by human researchers, so their supply grows only as fast as researchers can work. We introduce AutoEnvScaling, which treats environment design as a terminal task and uses agents to automate the data flywheel. A proposer agent builds new RL environments from the rollouts of the solver being trained, and the solver’s new rollouts guide the next round. By turning environment generation into a terminal task, AutoEnvScaling increases the valid-environment rate by up to 25 times compared to prompt-only generation, at up to 75% lower cost per valid environment. We show that training Qwen3.5-35B-A3B on AutoEnvScaling environments from GPT-5.6-sol improves its Terminal-Bench 2.1 score by 4.3 points over the matched Tmax baseline. The broader goal of AutoEnvScaling is to enable recursive self-improvement (RSI): by making both environment generation and solving terminal tasks, it lets one model co-evolve in both roles. We show that Qwen3.6-35B-A3B, acting in both roles, gains 12.2 points on Terminal-Bench 2.1 and raises its Terminal-Bench 4.0 score from 1.6 to 3.2. The pipeline also generalizes across agentic domains, generating environments for software engineering, computer use, professional work, and GPU kernel programming. Overall, these results show a path toward agents building their own training environments, reducing human reliance and taking a core step toward RSI.

[422] SPROUT: Scaling Agentic Reinforcement Learning for Long-Horizon Software Engineering cs.SE | cs.CL | cs.LGPDF

Mengyang Liu, Weiming Wang, Haixin Nan, Yu Wu, Peize Li

TL;DR: SPROUT 用文件系统和智能体上下文快照保存长程编码任务中间状态，并通过验证器引导回溯和分支复用失败尝试的早期进展，以提高强化学习 rollout 效率。

Details

Motivation: 长程软件工程任务中一个小错误就可能使完整 rollout 失败，重新从初始状态开始浪费计算；仅重置上下文也不能恢复依赖安装等执行环境状态。

Result: SPROUT 每个 assistant 回合后同时检查点保存文件系统和智能体上下文，由验证器引导策略决定从哪个中间步骤恢复及如何继续。在多种 LLM 的长程编码任务上，pass@k 最高提升 23.6%，token 用量仅为基线的 32%–52%。在 Qwen3.8-27B 强化学习训练中，SPROUT 在三个长程编码基准上比标准 GRPO 高 13.9 分，且每训练步耗时降低 39.5%。

Insight: 保存工作区和上下文的完整快照，使失败尝试可回到有效检查点分叉，而不重复完成全部前置工作。摘要的最高 pass@k 增益和训练比较针对特定任务/模型，检查点存储、回滚正确性及并行分支开销未量化。

Abstract: Reinforcement learning (RL) has driven substantial progress in coding agents by relying on verified successful trajectories as learning signals. However, such trajectories are scarce in long-horizon software engineering tasks, where a minor error can invalidate the entire rollout. Repeatedly restarting from the initial state is commonly used yet inefficient, and resetting only the context with working directory misses state changes such as installed dependencies. Therefore, branching from intermediate steps to preserve earlier progress requires restoring both the agent context and execution environments. In this paper, we present SPROUT (SnaPshot-based RollOUT), an agentic RL rollout framework that scales long-horizon coding tasks by resuming and branching from intermediate states. Specifically, SPROUT was built on top of a lightweight infrastructure that checkpoints both the filesystem and the agent context after each assistant turn. A verifier-guided backtracking policy then selects where to resume a failed attempt and how to proceed. Extensive experiments on long-horizon coding tasks show that SPROUT achieves up to 23.6% improvement over pass@k across various LLMs, while using $32$–$52%$ of their tokens. For RL training of Qwen3.8-27B, SPROUT stably outperforms standard GRPO by $13.9$ points on three long-horizon coding benchmarks while cutting the time per training step by $39.5%$. SPROUT thus makes RL on long-horizon coding tasks both more effective and cheaper to scale.

[423] PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards cs.SE | cs.AI | cs.CLPDF

Yunkai Zhan, Miaosen Chai, Shangshang Wang, Jike Zhong, Yuqing Yang

TL;DR: PreciseCoder 为单元测试奖励增加语义验证的局部编辑信号，抑制调试时过度修改和对正确代码臆造 bug，同时保持通过率。

Details

Motivation: 代码 RL 常用单元测试作为唯一奖励，只判断最终程序是否通过；作者发现该信号虽能提高通过率，也会让模型改动更多，并在原本正确的代码中臆造错误。

Result: PreciseCoder 奖励语义核验过的精确局部编辑。4B、9B、27B 模型平均编辑行数由 7–15 降至 3，且通过率与仅单元测试训练相当。只用单行 bug 训练后可迁移到仓库级修复，解决更多问题且补丁更小。9B PreciseCoder 推理轨迹还可帮助更强模型 GPT-5.6 Sol、Qwen3.8-Max 在无需额外训练时将编辑精度提高 4.6–12.4 分。思维轨迹分析显示该奖励鼓励逐行推理、指出 bug 行并形成一致的错误假设。

Insight: 奖励精确且语义核验的修改，可以在通过率相近时减少无谓代码改动；轨迹还可作为更强模型的提示信号。摘要未给训练数据规模、通过率绝对值及仓库级任务的具体比较基线。

Abstract: Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes. We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses.

eess.AS [Back]

[424] RUBATO: Decoupled Clocks Allow Full-Duplex Models to Think as Fast as They Can eess.AS | cs.AI | cs.CL | cs.SDPDF

Yi-Jen Shih, Puyuan Peng, Haohe Liu, David Harwath, Abdelrahman Mohamed

TL;DR: RUBATO 用独立时钟的多流位置编码解耦全双工语音模型的聆听、说话和内部思考速率，使模型可在推理时按硬件速度调整解码。

Details

Motivation: 全双工语音模型加入内部思考流后，常用固定解码速率保证实时处理，但这假设特定硬件时延；思考本身是异步的，固定速率可能限制答题速度或降低准确率。

Result: RUBATO 为不同流分配独立时钟，并以两种极端解码速率训练，目标是在中间速率零样本插值。五个口语推理任务上表现稳健。Spoken GSM8K 中，基线在插值速度下准确率降至 1.7%，RUBATO-Full 则在各速率维持 48.7%–51.1%，更快硬件上 time-to-answer 缩短 6.5 倍，同时首次响应时间不变。BIG-Bench Audio 上优势也保持，在用户于模型说话和推理时插话的复杂全双工场景，RUBATO-Full 准确率为 40.0%–45.1%，基线完全失败。

Insight: 独立时钟使思考、聆听和发声不必共享静态节奏，可利用硬件吞吐提高最终答案速度而不延迟首次响应。准确率和速度是作者报告的特定任务/设置，摘要未给出硬件型号、速率范围或实时流式延迟分布。

Abstract: Recent Speech Full-Duplex (FDX) models are capable of engaging in naturalistic conversation by listening and speaking at the same time. Building on this foundation, recent architectures introduce internal thinking streams to enhance their intelligence. To guarantee these streams process successfully in real time, existing methods enforce a fixed decoding rate, which assumes a hardware profile capable of a specific latency. We argue this static approach is severely suboptimal. Because internal thinking is intrinsically asynchronous, an ideal FDX model should sustain real-time conversational flow while allowing each stream to decode at its own hardware-adaptive rate. Thinking streams must process as fast as the hardware allows to deliver accurate answers faster. To solve this, we propose RUBATO, a novel multistream positional encoding that assigns independent clocks to entirely decouple the streams. This structural isolation empowers the model to zero-shot interpolate across any inference decoding speed. It drastically simplifies dataset construction by requiring training on only two extreme decoding speeds while smoothly adapting to any intermediate rate during deployment. RUBATO demonstrates consistent robustness across five spoken reasoning tasks. On Spoken GSM8K, a baseline model collapses to 1.7% accuracy at interpolated speeds. Conversely, RUBATO-Full sustains 48.7% to 51.1% accuracy across all rates, leveraging faster hardware to achieve a 6.5x shorter time-to-answer with a constant time-to-first-response. This advantage holds on BIG-Bench Audio and amplifies significantly during complex full-duplex interruptions. Specifically, when a user interjects while the model is actively speaking and reasoning, RUBATO-Full achieves 40.0% to 45.1% accuracy, whereas the baseline fails completely.

eess.IV [Back]

[425] Exploring the Latent Structure of AMD Severity Using Retinal Foundation Models eess.IV | cs.CVPDF

Niccolò Fineschi, Jovana Panic, Atsushi Watanabe, Hideaki Hayashi, Kazuya Nishimura

TL;DR: 该研究分析 RET-CLIP 中年龄相关黄斑变性严重度的潜在表示，发现疾病更像连续谱，且中期病变比健康与晚期阶段更难区分。

Details

Motivation: 视网膜基础模型在多种眼科任务上表现较强，但疾病信息如何组织在其嵌入空间中，以及这种结构如何影响下游预测，仍知之甚少。

Result: 作者用 AREDS 双眼眼底图像，分析 RET-CLIP 相对非眼科编码器的严重程度表示、三个临床粒度分类任务及有影响的潜在维度。RET-CLIP 将 AMD 表征为连续谱而非离散类别，健康和晚期病变分离较清楚、中间阶段较难。下游平衡准确率从二元 AMD 的 0.844、Late-AMD 的 0.814，降至六分类分级的 0.446。重要潜在维度与玻璃膜疣、色素变化及晚期视网膜病灶等表型统计相关。

Insight: 任务标签越细粒度，严重度连续谱中的中间状态越难判别；嵌入结构影响临床任务难度，也提供可解释分析线索。摘要未报告队列规模、外部验证或不同编码器在全部任务上的数值比较。

Abstract: Retinal foundation models have demonstrated strong performance across a wide range of ophthalmic tasks, yet relatively little is known about how disease-related information is represented within their learned embeddings and whether this distribution influences downstream prediction performance. In this work, we use Age-related Macular Degeneration (AMD) as a case study to investigate how disease severity is represented within the embedding space of RET-CLIP, a retinal vision-language foundation model, using bilateral fundus images from the AREDS dataset, benchmarked against non-retinal encoders. We first characterize the distribution of AMD severity within the learned representation space and then evaluate how this structure influences classification performance across three clinically relevant task formulations of increasing granularity. Finally, we analyze the most influential latent dimensions combining feature attribution, statistical phenotype association, and qualitative image analyses. Our results show that RET-CLIP represents AMD severity as a continuous spectrum rather than a set of discrete categories, with healthy and advanced disease stages exhibiting clearer separation than intermediate ones. This organization was reflected in downstream performance, which decreased from a balanced accuracy of 0.844 and 0.814 for binary AMD and Late-AMD detection respectively to 0.446 for six-class grading. Furthermore, the most influential latent dimensions were statistically associated to distinct retinal phenotypes, including drusen accumulation, pigmentary alterations, and advanced retinal lesions. These findings suggest that the structure of retinal foundation-model representations influences the difficulty of clinically relevant prediction tasks and demonstrate the potential of representation-level analysis for understanding disease organization and improving interpretability.

[426] Deficit-Driven Stratified Acquisition for Building STRATA-S12: An Open SAR-Optical Corpus for Self-Supervised Learning eess.IV | cs.CV | cs.LGPDF

Christian Sanchez-Leon, Dongming Chen, Dongqi Wang

TL;DR: DDSA 将 SAR—光学卫星配准语料构建视为分层覆盖采集问题，按类别、季节和区域缺口动态排序，并以人工在环修补失败配对。

Details

Motivation: 自监督地球观测配准语料往往机会式采集，没有分层配额控制；不均匀云覆盖会导致某些地理区域数据缺失。

Result: DDSA 构建 Sentinel-1/Sentinel-2 配对时校正半球因素，按类别×季节×区域共 192 个层分层，执行严格时间配对和多阶段质检，并在采集中闭环修补缺口。前几个百分点预算就填充全部 192 层，而丰度优先顺序在后期前让大部分层仍为空。缺陷修复率取决于成因持续性：几何 nodata 等失败可修复率 41.4%，云失败为 18.8%。修复环贡献发布配对的 39.5%，质量无可测下降，人工复核通过率 89.8%。总计 2,928,370 次采集尝试释放 740,746 对。数据集 STRATA-S12 提供 STAC/PDS4 对齐元数据；预注册线性探针支持其用于自监督预训练。

Insight: 缺口驱动采集使预算早期覆盖所有地理/季节层，云问题更难修复是因为季节性持久，而非缺陷标签本身。发布配对约占尝试数四分之一，说明高质量门控有采集成本；跨语料和跨方法系统比较留待后续。

Abstract: Building a co-registered SAR-optical corpus for self-supervised Earth Observation is above all an acquisition problem, yet the main multi-modal datasets are collected opportunistically, with no per-stratum quota control and no mechanism against the geographic dropout that non-uniform cloud cover leaves behind. We present Deficit-Driven Stratified Acquisition (DDSA), an algorithm that treats building a co-registered Sentinel-1/Sentinel-2 corpus as a constrained maximum-coverage problem over a hemisphere-corrected (class x season x region) stratification, enforces strict temporal pairing and a multi-stage quality gate, and closes the loop during acquisition with a human-in-the-loop deficit-repair stage. Measured on its own product, the algorithm yields three results a single-pass acquisition cannot report. Deficit-driven ordering makes a partial run a covering one: it holds every one of the 192 strata populated from the first few percent of the budget, where an abundance-first order leaves most strata empty until the end. Whether a failed attempt can be recovered depends on the temporal persistence of its cause, not the kind of defect: geometric failures such as nodata are among the most repairable (eta = 41.4%) while cloud is among the least (eta = 18.8%), because a cloudy location is cloudy on most dates of its season. The repair loop supplies 39.5% of the released pairs at no measurable quality cost, with 89.8% surviving manual review. The quality guarantee has a price: 2,928,370 acquisition attempts produced 740,746 released pairs. We release the corpus, STRATA-S12, with STAC/PDS4-aligned FAIR sidecars, and a pre-registered linear-probe check confirms it is fit for self-supervised pre-training. A metadata-level coverage comparison against SSL4EO-S12 and SEN12MS situates the corpus over a shared frame; systematic cross-corpus and cross-method comparison is left to follow-up work.

Table of Contents

cs.CV [Back]

[1] How Vision Transformers Represent Position: A Mechanistic Comparison of Absolute and Rotary Positional Encodings cs.CV | cs.LGPDF

[2] Scenes as an Attack Surface: Uncovering Scene-Grounded Safety Failures in MLLMs cs.CVPDF

[3] Human Pose Estimation by Probabilistic Mixtures of Gaussian and Uniform Distributions cs.CV | cs.LGPDF

[4] TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback cs.CVPDF

[5] Video-Arena: A Scalable Agentic Video Benchmark for Spatiotemporal Capability cs.CVPDF

[6] CAPTOR: a vision-language foundation model for Chest-Abdomen-Pelvis Oncologic CT cs.CV | cs.AIPDF

[7] Token-Budgeted Candidate Retrieval for 3D Scene-Grounded VLMs: Category Names vs. Pairwise Geometry cs.CV | cs.LGPDF

[8] Latent World Action Model for GUI Agents cs.CV | cs.AI | cs.HCPDF

[9] Image Diffusion Models Are Good 3D Human Motion Generators cs.CVPDF

[10] BigCell: Generating Gigapixel Whole-Slide Images cs.CV | cs.LGPDF

[11] Thinking with Images without Coordinates cs.CVPDF

[12] RoboFake: Are Deepfake Detectors Ready for Humanoid Face Robots? cs.CV | cs.CR | cs.ROPDF

[13] GeoFix: Geometry Fixing for 4D Consistent Generation cs.CV | cs.AI | cs.LGPDF

[14] Real-Time Traffic Density Estimation: A Hybrid Computer Vision and Predictive Machine Learning Pipeline cs.CVPDF

[15] RealityBench: Simulated Environments to Evaluate a Vision Language Model’s Ability to Document Human Physical Work cs.CV | cs.AI | cs.CLPDF

[16] Learning Transferable World-Action Models from Task-Paired Human Videos cs.CV | cs.ROPDF

[17] SceneForesee: Learning Scene Evolution for Spatiotemporal Reasoning cs.CV | cs.AIPDF

[18] VisualCoding: Learning Executable Skills for Reliable Visual Reasoning cs.CVPDF

[19] FreeCHROMA: Generative Chroma Keying from Unaligned Background References cs.CVPDF

[20] TartanMatch: Towards Universal Dense Correspondence Across Modalities cs.CV | cs.LG | cs.ROPDF

[21] Adversarial Gradient Forcing for Robust Few-Step Autoregressive Video Generation cs.CVPDF

[22] FrameFold: Response-Guided KV Cache Consolidation for Autoregressive Long Video Generation cs.CVPDF

[23] Attribute-Centric Vision-Language Alignment for Fine-Grained Gait Understanding cs.CVPDF

[24] Bidirectional Adaptive Transformers for Multimodal Anomaly Detection cs.CVPDF

[25] EviScene: Beyond “Is It AI?” for Agentic Image Factuality Assessment cs.CVPDF

[26] EarthAlign: Learning Language-Aligned Representations of Earth System Data cs.CV | cs.AI | cs.LGPDF

[27] TAG: Calibrating SAM3 for Universal Image Segmentation without Dense Supervision cs.CVPDF

[28] Prospective Forcing: Learning to Look Ahead for Autoregressive Long Video Diffusion cs.CV | cs.AIPDF

[29] TGuide: Faster Autoregressive Video Diffusion via Historical Trajectory Guidance cs.CVPDF

[30] TETRIS-4DGS: Efficient Post-Training Compression of 4D Gaussian Splatting cs.CVPDF

[31] SPARE: Shared Perception and Adaptive Retrieval of Evidence for Streaming Video Understanding cs.CVPDF

[32] Competition Domains Matter: Selective Head Globalization for Training-Free Sparse Attention in Video Diffusion Transformers cs.CVPDF

[33] CAST-ASD: Causal Attention over Speakers and Time for Online Active Speaker Detection cs.CV | cs.SDPDF

[34] HumanPhys: A Human-Centric Physical-World Perception Benchmark for Real-World Visual Reasoning cs.CVPDF

[35] Demiurge: Forging Navigable 4D Worlds via Autoregressive Video Rerendering cs.CVPDF

[36] PURE: Refining Feed-forward 3D Reconstruction with Unlabeled Post-training cs.CVPDF

[37] LightVAE: Towards Compact and Efficient Video Autoencoders cs.CVPDF

[38] Align-VAE: Aligning Attention and Geometry in VecSet VAEs for 3D Reconstruction cs.CVPDF

[39] YOUDrive: Driving Style as a Composable Direction for End-to-End Autonomous Driving cs.CV | cs.AI | cs.ROPDF

[40] [Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning cs.CVPDF

[41] Getting Space Right: Agentic Rewards for Image Generation cs.CV | cs.AIPDF

[42] CAST: Real-Time Motion Capture for Any Skeleton Topology cs.CVPDF

[43] MORE: Learning Motion-Oriented Registers for Driving World Models and Planning cs.CVPDF

[44] Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations cs.CVPDF

[45] CatPortrait: Taming Video Generative Models for Portrait Animation via In-context Conditioning cs.CVPDF

[46] Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning cs.CV | cs.LGPDF

[47] Retraining-Free Depth Pruning of Vision Transformers: A Cross Domain Study of Layer Redundancy cs.CVPDF

[48] CowCare: A Unified Visual Framework for Viewpoint-Robust Cattle Tracking and Spatiotemporal Heat-Stress Analysis cs.CVPDF

[49] What Really Works for Universal Multimodal Embeddings? A Controlled Study of Contrastive Training Strategies cs.CVPDF

[50] Use Of Convolutional Neural Networks In Detection Of Branch Dieback Disease In Pequi Trees cs.CV | cs.CE | cs.DBPDF

[51] BrainAid: Efficient Brain Tumor Detection Using Lightweight CNNs and Explainable AI for Resource-Constrained Environments cs.CVPDF

[52] TIDE: Trajectory-Informed Skill Distillation for Efficient Text-to-Video Prompt Optimization cs.CVPDF

[53] ExeWorldBench: Benchmarking Code-Based Executable 3D World Modeling cs.CVPDF

[54] GeoBridge: Evolving Transferable Decoder Topologies for Geometric Vision cs.CV | cs.AIPDF

[55] TempoKV: Predictive KV Cache Quantization for Autoregressive Video Generation cs.CVPDF

[56] BaT-SAM: Introducing Inductive Bias into SAM via a Mixture of Convolutional Experts for Brain Tumor Segmentation cs.CVPDF

[57] Across Viewpoints and Time: Trajectory-Generalizable Physical Attacks on World Models and Video Understanding Models cs.CV | cs.AIPDF

[58] Video-Based Sexual Harassment Detection Using Fine-Tuned Video Transformers cs.CVPDF

[59] VideoEvoHarness: Self Evolving Agent Harness from Unlabeled Videos for Video Understanding cs.CVPDF

[60] OmniRSI: Taming Latent Diffusion for Unified Multimodal Remote Sensing Image Generation cs.CVPDF

[61] Harnessing SPAD Noise: Test-Time Gated Attention for Indoor Depth Estimation cs.CV | cs.LG | cs.RO | eess.SPPDF

[62] Rethinking Supervision in Document Unwarping: A Self-consistent Flow-free Approach cs.CVPDF

[63] EmoTrace-Bench: Diagnosing Emotional Expression in Text-to-Video Generation through Controlled Cues and Intensity cs.CVPDF

[64] Semantic Saliency from Multi-Modal Large Language Model Scene Understanding Maps cs.CVPDF

[65] Structured Stateful Relational Scripts with a High-Quality Dataset for Multi-Shot Audio-Visual Video Generation cs.CVPDF

[66] A Plausible Future Is Not a Causal Future: Can Video Models Generate How the World Unfolds? cs.CVPDF

[67] Genesis: Visual Augmentation Primitives for Multimodal Chain-of-Thought Reasoning cs.CVPDF

[68] FireNav: Decision-Oriented Reconstruction Evidence Routing for Vision-Language Navigation in Indoor Fire Environments cs.CVPDF

[69] Lesion-Aware Attention for Mammography Segmentation with Heterogeneous Supervision cs.CVPDF

[70] SurgCast: Action-Conditioned Future Skeletons for Controllable Surgical World Models cs.CVPDF

[71] Sparse All-Layer Connector for Domain Generalised Semantic Segmentation cs.CVPDF

[72] A Persian Vision-Language-Action Dataset and Benchmark for Ambiguity- and Infeasibility-Aware Instruction-to-Plan Grounding cs.CV | cs.CLPDF

[73] ProtoPur: Prototype-Purity Pseudo-Labeling for Weakly Supervised Instance Learning in Pathology cs.CV | cs.LGPDF

[74] Rethinking Human Intent-to-CAD: Parametric CAD Model Generation via Cooperative Multi-Task Alignment and Spatial-Aware Reinforcement Learning cs.CV | cs.AIPDF

[75] PhySimCode: A Benchmark and Evaluation Method for Physics Video to Code Generation cs.CVPDF

[76] Visually Legible but Adversarial: Watermark-Constrained Attacks on End-to-End VLMs for Document Understanding cs.CVPDF

[77] ScenePlacer: Reformulating Scene-Level Object Pose Estimation as Visibility Field Learning cs.CVPDF

[78] RSAJEV: Training-Free Abstaining Prototype Memory for Eliminating Silent Hallucination in JEV-Based Remote Sensing Classification cs.CVPDF