cs.CL [Total: 64]
cs.CV [Total: 177]
cs.AI [Total: 26]
cs.CR [Total: 1]
cs.DC [Total: 1]
cs.HC [Total: 2]
cs.IR [Total: 2]
cs.IT [Total: 1]
cs.LG [Total: 22]
cs.MA [Total: 1]
cs.RO [Total: 9]
cs.SD [Total: 3]
eess.IV [Total: 2]
stat.ML [Total: 1]

cs.CL [Back]

[1] The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections cs.CL | cs.AI | cs.CRPDF

Haoxuan Jia, Yang Liu, Bin Chong, Yingguang Yang, Yancheng Chen

TL;DR: 该研究用分支重采样区分智能体何时决定服从提示注入，发现推理开启后决定更晚形成且模型差异显著；单次运行标签可能不稳定。

Details

Motivation: 运行时监控和单次基准评分常假设智能体读到注入内容时就已决定是否服从，但模型之后的推理可能改变结果；若决策持续演化，监控时机和单次评测都会失真。

Result: 作者在智能体读到注入内容的步骤处分支重放完全相同的前缀，多次重采样后续，分解结果方差中读入时已固定和之后才形成的部分。在四个AgentDojo测试套件上，关闭思考的Qwen3结果大多在读入时决定，状态间解释方差占0.91；开启思考后降至0.62，下降的95%置信区间为[-0.38,-0.22]，三个攻击模板和两种成功标签方向一致。思考降低Qwen3攻击成功率，却提高另外两个模型的成功率。对推理前缀重采样表明决定逐步形成。基于较小模型的一次性CoT监控漏掉不少已决定内容；更强监控和预测下一动作的前瞻监控接近前缀所能揭示的信息。思考模型的单次标签与重跑最多35%不一致。

Insight: 是否服从注入可能在推理中逐步形成，且“思考”对不同模型的影响方向不同；评测应考虑重复运行。摘要结果来自特定模型与AgentDojo任务，方差分解不是对所有智能体固定属性。

Abstract: LLM agents that read untrusted content can be hijacked by prompt injections. Runtime monitors that try to catch a hijack before the agent acts, and benchmarks that score each task with a single run, share an assumption: whether an agent follows an injection is settled when the injection is read. We test this assumption with branching rollouts that replay an episode exactly up to the step at which the agent reads an injection and then resample the rest of the episode many times, splitting the outcome variance into a part fixed at read time and a part decided afterwards. Across four AgentDojo suites, a Qwen3 agent without thinking is largely decided at read time (0.91 of the outcome variance lies between states); with thinking on, this share drops to 0.62 (95% CI of the drop [-0.38, -0.22]), in the same direction for three attack templates and two success labels. How much is fixed at read time is a property of the model that the same measurement reveals, and thinking lowers the attack success rate of Qwen3 but raises it for two other models. Resampling from reasoning prefixes shows that the decision forms gradually as the agent reasons. Zero-shot chain-of-thought monitors built on cheaper models miss much of what is already decided, whereas stronger monitors, and a lookahead monitor that judges simulated next actions, come close to what the prefix determines. Finally, single-run labels of thinking agents disagree with a rerun up to 35% of the time, so benchmarks need repeated runs per task.

[2] Preservation and loss of algorithmic structure in experimental diffusion of joint-action norms cs.CL | cs.MAPDF

Dhara Yu, Bill D. Thompson

TL;DR: 该研究通过大规模社会规范传播实验观察隐式共同计划算法如何在人群间演化，发现跨变体互动会简化复杂程序，而互动修复可抵消这种压力。

Details

Motivation: 共享空间、轮流和排队等日常交互规范包含结构化程序，但这些隐式联合计划的“算法”在人群传播过程中如何变化尚不清楚，尤其复杂结构是否会在早期传播阶段丢失。

Result: 作者让人类参与者在多个独立群体中传播新社会规范，并利用程序合成刻画各群体独立形成的程序。继承的程序在具有不同规范变体的参与者互动中会结构简化并发生丢失；然而，提供互动修复机会可抵消这种趋势，使复杂算法得以保留。实验据此表明，互动推理会影响程序性规范文化演化中的结构复杂度选择压力。

Insight: 文化传播不仅复制规范内容，也可能改变其程序结构；参与者能否协调和修复分歧会影响复杂规则存续。摘要未提供参与者数量、具体任务规则和量化结构变化，因此结论限于该实验设置。

Abstract: Everyday norms of social interaction - e.g. norms of sharing space, taking turns, or queuing - are essential to large societies, offering structured procedures for common interactions. However, it remains unclear how the underlying algorithms of these implicit joint plans evolve as they diffuse through populations. Here we show that this process of diffusion can select against complex algorithms in the earliest stages. We conducted a large-scale study of experimental diffusion of novel social norms among human participants, leveraging advances in program synthesis to characterize the procedures people developed in independent populations. Inherited procedures became structurally simplified and lost during interactions between participants who inherited different variants of joint-action norms; however, opportunities for interactive repair offset this pressure, preserving complex algorithms despite misalignment. These results demonstrate that interactive reasoning mediates selective pressures on structural complexity during cultural evolution of procedural norms.

[3] Reasoning with Context Management and External Memory cs.CLPDF

Man-Ling Sung, William Hartmann, Jiangying Zhou

TL;DR: RCMEM用外部记忆扩展推理模型的上下文管理，减少复杂任务中重复计算中间步骤的耗时，并在部分数学问题上提升表现。

Details

Motivation: 推理模型依靠上下文内存跟踪推理状态，复杂问题需要更多上下文和记忆。若不使用缓存而压缩上下文，可能丢弃后续推理所需信息。

Result: RCMEM在上下文管理外增加外部记忆，保存额外信息而不扩大推理模型当前上下文。作者在部分数学问题上测试模型规模从0.4到2,500万参数的模型及GPT-2，报告该方法显著缩短需要反复计算中间步骤的问题的计算时间，并改善复杂问题表现。

Insight: 外部记忆可复用中间计算，减少在有限窗口内重复生成相同推理步骤。摘要未给出问题集、运行时和准确率数值，也未说明外部记忆检索成本，因此结果属于范围有限的概念验证。

Abstract: Reasoning models maintain information about the status of their reasoning process through in-context memory. Complex problems require larger context windows and memory requirements. In this scenario, context management without caching risks discarding useful information for reasoning. We propose RCMEM, an approach that extends context management with an external memory to maintain additional information without burdening the reasoning model with additional context. RCMEM significantly reduces computation time on problems that require repeated computation of intermediate steps, while achieving better performance on complex problems. We demonstrate our approach on a few mathematics problems using model sizes ranging from 0.4 to 25 million parameters, and GPT-2.

[4] Preserving Reasoning Diversity in Reinforcement Learning with Verifiable Rewards via Adaptive Replay cs.CL | cs.LGPDF

Wen-Zhi Li, Sainyam Galhotra, Claire Cardie

TL;DR: TAR在可验证奖励强化学习中自适应回放基础策略已有的多样正确解法，只在这些行为模式开始被压制时触发，以减轻推理多样性遗忘。

Details

Motivation: RLVR可提升推理能力，但有限的在线采样可能逐渐压低语义不同却同样正确的推理模式；某个模式样本不足后，标准在线强化学习很少能自行恢复。

Result: TAR建立包含基础策略多样正确解法的紧凑记忆库，并采用自适应机制，只在RLVR训练正在压制已存行为时选择性回放。在六个数学推理基准和两个Qwen2.5模型规模上的实验中，TAR较既有保多样性后训练方法更好平衡推理成绩与多样性保留，并显著减少对基础模型正确解法的遗忘。

Insight: 保留有价值的正确推理模式并在其趋于消失时重放，可缓解有限样本训练造成的模式坍缩。摘要未给出具体准确率、多样性指标及记忆库成本，也未说明策略在非数学任务上的效果。

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) improves the reasoning capabilities of large language models, but often reduces the diversity of their reasoning behaviors. We study this phenomenon from the perspective of finite-sample learning dynamics and show that limited on-policy rollouts can progressively suppress semantically distinct but equally rewarded reasoning modes. Once a mode becomes sufficiently under-sampled, standard on-policy RLVR rarely recovers it. Motivated by this observation, we propose TAR, a Targeted Adaptive Replay framework that preserves diverse correct reasoning behaviors already present in the base policy. TAR constructs a compact memory bank of diverse base-model solutions and selectively replays them through an adaptive mechanism that activates only when stored behaviors are being suppressed during RLVR. Across six mathematical reasoning benchmarks and two Qwen2.5 model scales, TAR consistently achieves a better balance between reasoning performance and diversity preservation than existing diversity-preserving post-training methods, while substantially reducing the forgetting of correct base-model solutions. Code is available at https://github.com/wenzhilics/tar.

[5] Encoder-Free Multimodal LLMs Learn Implicit Encoders cs.CL | cs.AI | cs.CV | cs.LG | eess.ASPDF

David Harwath, Desh Raj, Xavier Velez, Sandra Arcos-Holzinger, Willie Jeng

TL;DR: 该研究分析Gemma-4 12B和Inkling-Small等无专用编码器多模态模型，发现其早期Transformer层会形成类似音频和视觉编码器的隐式表征。

Details

Motivation: 传统多模态大模型用模态专用编码器把音频或图像变成token嵌入；新模型直接把采样窗口、像素或频谱特征输入语言模型。需要了解早期语言模型层是否自行学会类似专用编码器的处理功能。

Result: 作者实证分析两种无编码器多模态模型的早期层，发现其确实充当隐式编码器，以较少受跨模态上下文影响的方式编码音频和视觉输入。在多模态推理任务中，先在早期层独立编码各模态，与从第一层起联合编码相比准确率相同。逐层探针还显示，这些隐式编码器捕捉信息的行为类似多种常用专用编码器。

Insight: 模态编码功能可能内化到Transformer早期层，显式前置编码器并非唯一架构选择。结论基于两种模型及论文测试任务，摘要未说明准确率数值、模态覆盖范围和模型规模迁移性。

Abstract: We empirically analyze two recently proposed encoder-free multimodal LLMs, Gemma-4 12B and Inkling-Small. Unlike previous multimodal LLMs, these models do not use modality-specific encoders to extract tokenized embeddings representing audio and image input; instead, they directly feed windows of samples, pixels, or spectral features to the Transformer language model. We analyze whether the early layers of these models learn to function similarly to dedicated audio and visual encoders. We find that these early layers do in fact act as ``implicit encoders,’’ encoding audio and visual inputs in a way that is largely invariant to their cross-modal context. Furthermore, when performing multimodal reasoning tasks, we show that independently encoding the modalities throughout these early model layers achieves the same accuracy as jointly encoding them from the first layer onwards. Finally, we use layer-wise probing to analyze the salient information captured by the implicit encoders and find their behavior to be similar to several popular encoder models.

[6] Toward a Universal Local Speech Feature Extractor Through Distillation cs.CLPDF

Jessica Yang, Marcelo Beramendi Caballero, Ju-Chieh Chou, Karen Livescu

TL;DR: 该研究通过蒸馏从多种自监督语音模型学习两层通用局部特征提取器，以约3%的CNN乘加运算量保留大部分教师性能。

Details

Motivation: 自监督语音模型通常由计算开销较大的卷积局部特征提取器和后续Transformer构成。不同模型的CNN部分相似，并与简单梅尔频谱特征有共同点，因此可能用统一的小型特征器替代。

Result: 作者将HuBERT、data2vec和WavLM作为一个或多个教师，通过蒸馏训练两层特征提取器。SUPERB及SUPERB-SG任务上的实验显示，该特征器大体保留教师模型表现，也可用于训练时未见的wav2vec 2.0。计算量仅为CNN特征提取器的3%乘加运算，使整体推理运行时间减少超过20%。

Insight: 语音模型的局部特征提取部分可能共享可迁移结构，适合跨模型蒸馏复用。摘要未给出各下游任务的单项性能损失、具体延迟和教师组合差异，因此“大体保留”仍需结合完整表格判断。

Abstract: Self-supervised speech models commonly consist of a set of convolutional neural network (CNN) layers used as a local feature extractor, followed by transformer layers. The CNN layers are computationally demanding, and are also similar across models and resemble the much simpler mel spectral features. We hypothesize that the CNN layers can be replaced by a simple, universal model. To test this hypothesis, we learn a two-layer feature extractor via distillation from one or more teacher models (HuBERT, data2vec, and WavLM). Experiments on SUPERB(-SG) benchmark tasks show that our model largely retains the performance of the teacher models and can also be used in an unseen model (wav2vec 2.0). Our feature extractor requires only 3% of the MACs of the CNNs, resulting in a $>$20% overall inference runtime reduction.

[7] When Opposite Premises Translate Identically: A Source-Conditioned Diagnostic for Multilingual Commonsense Items cs.CL | cs.AIPDF

Shuai Wang, Rui Hu

TL;DR: 该研究用源语言反事实极性对比检查常识题的多语言翻译是否保留关系，发现斯瓦希里语中出现大量相反前提翻译成相同文本的碰撞。

Details

Motivation: 多语言推理比较依赖翻译后的前提保留英语源文关系；自动翻译分数可以排序候选，却不能证明某个反事实基准条目本身忠实。

Result: 作者对ACCORD前提采用源对比诊断：反转英语前提极性，独立翻译两种形式，统计目标语言输出完全相同的情况，并分别与各自源句评分。测试覆盖71个匹配问题身份、两个包含项目深度、五种目标语言及三个翻译检查点。斯瓦希里语每个检查点的710组中分别出现63、113和67组完全碰撞。在人类参考校准阈值下，针对NLLB-3.3B碰撞，往返自然语言推理仅有3/113组同时通过两侧，而LaBSE通过78组。每次碰撞证明至少有一条翻译不忠实，但无法确定是哪一侧。

Insight: 相反前提译成相同文本是翻译不忠实的明确警报，但该诊断不能识别具体错误译句。研究未验证对自然发生翻译错误的筛查准确率，也没有母语等价标注或答案模型评测，不能推断语言跳数导致的推理差异。

Abstract: Multilingual reasoning comparisons depend on translated premises preserving the relations in their English sources. Automatic translation scores can rank source-target pairs, but an acceptance decision alone cannot certify a counterfactual benchmark field. We test a source-contrast diagnostic for ACCORD premises: reverse the English polarity, translate both arms independently, count identical target outputs, and score each arm against its own source. Across 71 matched question identities at two containing-item depths, five target languages, and three translation checkpoints, Swahili has 63, 113, and 67 exact collisions among 710 pairs per checkpoint. At a human-reference-calibrated threshold, round-trip NLI passes both arms of 3 of the 113 NLLB-3.3B collisions, while LaBSE passes 78. Each collision proves at least one unfaithful translation, but does not identify its arm. These controlled measurements do not establish screening accuracy on spontaneous translation errors or a language-by-hop reasoning effect; the study has no native equivalence labels or answer-model evaluation.

[8] Modeling Discourse Profiles in Climate Misinformation Discourse cs.CLPDF

Ge Gao, Zhengyang Shan, Bryan A. Plummer, Margrit Betke, Derry Wijaya

TL;DR: 该研究以TikTok气候讨论语料中的重复修辞模式构造话语画像，并将其融合进细粒度气候错误信息分类，发现画像能改善多数文本模型的宏F1。

Details

Motivation: 气候错误信息检测常用文本单独分类评论或主张，忽略社交媒体中反复出现于评论和话题串的修辞模式，以及这些模式对错误信息传播和接受的作用。

Result: 作者从大规模TikTok气候讨论数据集中聚类气候相关评论，形成重复修辞话语画像，再通过后期融合将画像加入分类器。在按视频隔离、三个训练随机种子的评测中，画像增强模型在词袋、TF-IDF和MPNet表示上均显著超过纯文本基线的宏F1；RoBERTa-large平均提升为正但波动更大，差异未达到统计显著。进一步分析发现，依据话语画像条件化的语言模型智能体，其错误信息接受模式与ClimateDisc中画像层级的错误信息比例一致；错误信息比例与互动持续性的关系也随主导话语画像而异。

Insight: 话语层级的修辞模式可能提供单条评论无法体现的传播背景，适合与文本特征融合建模。摘要未给出语料规模、具体宏F1增益和画像类别解释，且RoBERTa-large结果不显著。

Abstract: Prior work on climate misinformation detection typically classifies individual comments or claims using text alone. This framing overlooks a key property of online climate discourse: misinformation is often produced, reinforced, and sustained through recurring rhetorical patterns that reappear across comments and threads. Using a large-scale TikTok climate discourse dataset, we construct discourse profiles by clustering climate-related comments into recurring rhetorical patterns. We then incorporate these profiles into fine-grained climate misinformation classifiers using a late-fusion architecture. Under video-disjoint evaluation across three training seeds, discourse-profile-aware models significantly improve macro-F1 over text-only baselines for bag-of-words, TF-IDF, and MPNet representations. RoBERTa-large exhibits a positive mean improvement but substantially greater variability, and the difference is not statistically significant. Beyond classification, our results show discourse profiles shape misinformation reception and propagation. Discourse-profile-conditioned language model agents exhibit misinformation acceptance patterns aligned with profile-level misinformation prevalence in ClimateDisc, and the relationship between misinformation prevalence and engagement persistence differs across dominant discourse profiles.

[9] EGIP: An Evidence-Grounded Interaction Protocol for Clinical Agents cs.CLPDF

Zecheng Ren, Ruichen Cong, Yi Min Chen, Jing Zhang, Feiyang Chen

TL;DR: EGIP为临床智能体规定角色信息边界、诊断动作和结果获取方式，将证据采集与最终诊断预测分开，提升多个临床智能体组合的准确率及证据支持度。

Details

Motivation: 交互式临床智能体通过问诊和诊断检查获取证据，但信息何时暴露、可选动作及检查结果呈现方式也会影响表现。若交互条件不明确或不一致，最终诊断准确率会混杂推理能力和协议差异。

Result: EGIP为不同角色设置信息边界，展示可执行诊断动作但不预先泄露结果，结果从存储的临床记录中检索，并将收集证据组织用于最终判断。分离证据获取与终端预测后，不同决策策略可基于同一观察比较。作者评估两个临床智能体框架、五个数据集和六种语言模型，EGIP在所有主要比较中提升诊断准确率。临床医生复核显示，自动判为正确且同时被医生评为有证据支持的回答比例从10.20%升至36.55%。

Insight: 统一并显式控制交互协议，可更公平地比较临床智能体推理与证据采集能力。摘要未提供各数据集准确率和协议细节；医生复核比例提升也不等于临床部署已安全有效。

Abstract: Interactive clinical agents acquire evidence through patient dialogue and diagnostic examinations, yet their performance also depends on how the interaction environment exposes information, available actions, and examination results. When these conditions are implicit or inconsistently controlled, final-answer accuracy can conflate diagnostic reasoning with the effects of the interaction protocol. We propose the \textbf{E}vidence-\textbf{G}rounded \textbf{I}nteraction \textbf{P}rotocol (EGIP), which enforces role-specific information boundaries, exposes diagnostic actions without revealing their results, retrieves examination results from stored clinical records, and organizes acquired evidence for final decision-making. By separating evidence acquisition from terminal prediction, EGIP enables different decision strategies to operate on the same observations. We evaluate EGIP across two clinical-agent frameworks, five datasets, and six language models. EGIP improves diagnostic accuracy over native baselines across all primary comparisons. A clinician review further shows that the proportion of automatically scored correct responses jointly judged evidence-supported rises from 10.20% to 36.55%. These findings support the effectiveness of EGIP as a unified interaction protocol for improving clinical-agent performance through controlled evidence access and acquisition.

[10] From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability cs.CL | cs.AI | cs.LGPDF

Dmitry Scherbakov, Oleg Chernikov, Victoria Khamash, Kirill Afendulev, Alexey Dontsov

TL;DR: 该研究用合成的概念—对立概念故事对构造可解释特征向量，扩展至1,036个AI安全相关概念，并展示跨模型检测、行为引导和表示分析能力。

Details

Motivation: 稀疏自编码器等无监督方法试图恢复语言模型完整特征空间，但可能牺牲单个特征质量和可解释性。研究采用更目标明确的路线，针对安全与对齐相关概念构造特征。

Result: 作者生成概念及其对立概念的合成故事对，取模型两类平均激活之差作为特征向量；该方法扩展至1,036个对齐相关概念、200万条跨语言和文体故事，并公开数据。六个模型、三个模型家族上的实验显示，这些特征可迁移到聊天对话和网页文本检测对应概念；特征在不同模型间形成可解释的共享几何，并可通过注入激活因果性地引导模型行为。作者还用其分析越狱攻击、数学推理和智能体推理中的表征。

Insight: 针对性合成对比数据可比完整无监督特征恢复更直接地构造目标概念方向，但效果依赖故事对能否代表概念边界。摘要未提供检测准确率、引导副作用和概念质量人工评价；扩展至任意领域或数百万特征仍是潜在方向。

Abstract: We study feature vectors: directions in a language model’s activation space that correspond to human-interpretable concepts, such as honesty or deception. Existing unsupervised methods like sparse autoencoders attempt to recover the whole feature space, often at the cost of feature quality and interpretability. We take a more pragmatic route: focusing on concepts relevant to AI safety and alignment, we generate synthetic concept-antagonist story pairs and use the difference of their mean model activations as a feature vector. We scale this approach to 1,036 alignment-relevant concepts using 2M synthetic stories across multiple languages and genres, which we publicly release. Across six models from three model families, our feature vectors reliably generalize to detecting their corresponding concepts in chat dialogues and web text. They form an interpretable geometry shared across models and can causally steer model behavior when injected into activations. We further use these features to analyze model representations during jailbreak attacks, mathematical and agentic reasoning. Our targeted feature construction thus offers a scalable route to studying concepts in language models, potentially extending to millions of features across arbitrary domains.

[11] Muse Glimmer 30B: An On-Device Agentic Model cs.CL | cs.AIPDF

Arash Ebrahim Abtahizadeh, Muna Aghamelu, Adil Ahmed, Juan C. Angeles Acuna, Alexis Audran-Reiss

TL;DR: Muse Glimmer是面向消费级设备的开源30B代理模型，结合图像感知、多步推理、工具调用和失败恢复，并通过约4比特量化支持单GPU推理。

Details

Motivation: 自主智能体能力通常依赖云端大模型，成本和数据外发限制本地部署；消费硬件上的模型还需在多步规划、可靠工具使用和失败恢复上具备足够能力。

Result: Muse Glimmer是带感知编码器的因果语言模型，支持文本及图像输入、文本输出，训练数据覆盖100多种语言，并量化至约4比特以适配单GPU推理。作者报告该模型在其规模档的智能体基准上达到先进表现，并在工具使用、多步规划和任务完成方面超过Gemma4-31B及Qwen3.6-27B等同规模模型。模型以Apache 2.0许可证开源。

Insight: 该模型的目标是把多步推理、工具操作和本地运行结合，降低消费设备运行智能体的门槛。摘要未列出基准名称、成绩、硬件要求和安全评估结果，无法独立判断其相对优势或设备适配范围。

Abstract: We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable tool use and failure recovery into a single model that runs locally on consumer devices without requiring cloud infrastructure. The model accepts text and image inputs and produces text outputs, was trained on data from over a hundred languages, and is quantized to approximately 4-bit precision for efficient single-GPU inference. Muse Glimmer achieves state-of-the-art performance for its size class on agentic benchmarks, setting a new standard compared to the models of its size: Gemma4-31B and Qwen3.6-27B on tool use, multi-step planning, and agentic task completion. We release the model under the Apache 2.0 license to foster open-source development of capable, safe, local AI agents.

[12] What RL Brings to Reasoning: A Few Decisions the Base Model Would Not Make cs.CL | cs.LGPDF

Siyu Zhang, Yuquan Lu, Yoshua Y. Li, Yifan Ding, Kun Zeng

TL;DR: 该研究通过稀疏替换强化学习模型与基础模型分歧的token，分析RLVR增益来自生成过程中的哪些决策，发现少数后续选择可恢复大部分收益。

Details

Motivation: 可验证奖励强化学习能提高数学推理，但相较基础模型究竟新增了什么仍不清楚：收益可能来自随机采样、开头风格、后续推理选择或生成过程中的少数关键位置。

Result: 作者在Qwen3-4B-Base上对120道AIME题逐步干预生成，并在MATH-500及三组模型对上检查。相同8,192 token上限的温度采样并未比贪心解码带来增益；只在基础模型与RL模型贪心选择不同时（约4%的生成位置）借用RL模型token，可恢复pass@1增益的90.5%。强制以“Alright”开头单独恢复35%增益，但在相同开头后继续替换后续token还能增加5.2个百分点；固定插入“Wait”等反思词规则没有重现收益。稀疏替换保留了基础模型较多样性，并使pass@1接近RL模型。在k=256时，其解题率67.5%，高于RL模型61.4%和pass@1匹配的整回答混合61.45%。比较使用相同尝试次数而非相同计算量，因为替换策略每步查询两个模型。

Insight: 强化学习收益并非只由推理开头决定，较大部分来自后续少量决策；可借用这些选择同时保留基础模型多样性。由于推理时双模型逐步查询，报告的高k结果不代表等计算成本优势。

Abstract: Reinforcement learning with verifiable rewards (RLVR) substantially improves mathematical reasoning, but what it adds to the base model remains unclear. We investigate this through a sequence of generation-time interventions on Qwen3-4B-Base, measuring how much of its RL-trained counterpart’s gain each intervention recovers on 120 AIME problems. Power sampling with the same 8,192-token output limit brings no gain beyond greedy decoding. Yet borrowing the RL model’s token only where the two models’ greedy choices differ, roughly 4% of generated positions, recovers 90.5% of the pass@1 gain, with similar recovery across three model pairs on MATH-500. We find that even a single opening token has a substantial effect: forcing the base model to begin with “Alright” recovers 35% of the RL gain. The opening alone is insufficient, however. Later replacements after the same opening add another 5.2 percentage points to pass@1, and the fixed rules we test for inserting reflective words such as “Wait” do not reproduce this benefit. Sparse replacement also retains much of the base model’s diversity while staying close to the RL model at pass@1. At k=256, it solves 67.5% of the problems, significantly more than the RL model (61.4%) and a mixture of whole responses matched in pass@1 (61.45%). The comparison is at equal attempt counts rather than equal compute, since rl-replace consults both models at every decoding step. These results show that the opening can recover a meaningful part of the RL gain, but a larger contribution comes from later choices, which can be borrowed while retaining much of the base model’s diversity. Code is available at https://github.com/CyrusZhang04/WhatRLBrings.

[13] DiffuPlan: Editable Planning State for Diffusion Language Model Agents cs.CL | cs.AIPDF

Qisheng Zhou, Zhen Xiong, Qiaoyu Tan

TL;DR: DiffuPlan利用扩散语言模型可重写先前文本的能力，维护可编辑规划状态并在新证据出现时只重构受影响步骤。

Details

Motivation: 扩散语言模型智能体大多沿用自回归智能体的追加式交互协议：一旦写下中间状态，后续新观察只能继续附加文本，无法利用扩散模型根据前后文重建旧内容的特性。

Result: DiffuPlan是扩散搜索智能体，保存单一可编辑规划状态；当新证据解析实体或改变搜索路线时，只重写受影响步骤。为处理不同长度的修订，作者提出Expansion-Partition Marginalization，对等价扩展路径求边缘化，并将扩散轨迹强化学习扩展至多轮搜索。在四个多跳问答基准上，平均F1为50.24，高于强自回归搜索智能体。相同数据和训练条件下，可编辑状态使扩散智能体F1提高2.44，而其自回归来源模型只提高0.47。

Insight: 允许智能体修改已提交的计划步骤，而非不断追加历史，是扩散模型可能具有的独特代理能力。摘要未给出四个基准的分项成绩、推理计算成本和编辑错误率，结论限于多跳问答搜索任务。

Abstract: Diffusion language models (dLLMs) are emerging as an alternative to autoregressive (AR) language models, yet dLLM agents largely inherit an interaction protocol designed for AR generation: once intermediate state is written, later observations can only be incorporated by appending more text. This is a poor match to a defining capability of dLLMs – reconstructing earlier text conditioned on context that appears both before and after it. We ask whether this capability can support a different agent-state paradigm in which intermediate planning state is directly revisable as evidence arrives. We introduce DiffuPlan, a diffusion search agent that maintains a single editable planning state and reconstructs only the affected steps when later evidence resolves an entity or changes the search route. To support variable-length revisions, we introduce Expansion-Partition Marginalization (EPM), which marginalizes over equivalent expansion paths. We further adapt diffusion trajectory RL to multi-turn search. Across four multi-hop QA benchmarks, DiffuPlan achieves 50.24 average F1, outperforming strong AR search agents. More importantly, under identical data and training, editable planning state improves the diffusion agent by 2.44 F1 versus only 0.47 for its AR source model. These results show that dLLMs need not imitate append-only AR agents: revising committed state is a distinct agentic advantage.

[14] ECHO: Efficient Causal Self-Recursion Beyond Auxiliary Multi-Token Prediction cs.CLPDF

Yuxuan Zhou, Junyuan Shang, Tianyu Wang, Chunrui Zheng, Zichang Liu

TL;DR: ECHO让自回归模型自身最终层预测多个未来token，并用动态深度采样控制训练成本，在推理时还能充当草稿模型。

Details

Motivation: 多token预测常作为辅助目标训练独立模块，每增加一个预测步长就增加模块和计算，且主要训练模块而非生成下一token的骨干参数。

Result: 高效因果自递归ECHO由骨干最终层同时预测不同未来跨度，使多步预测直接进入自回归目标；动态深度采样以随深度衰减的概率抽样训练深度，降低更深预测监督成本。预训练中增加深度在辅助模块收益饱和后仍可转化为下游准确率提升，尤其在推理和代码任务，训练开销约为训练全部深度MTP的一半。推理时模型可作为自身草稿器，并能超出训练深度起草；在开放式、高熵生成中保持草稿与目标模型对齐，而后训练草稿器会失去这种对齐。优势在token预算翻倍、重复预训练和监督微调设置下仍存在。

Insight: 把未来token预测转化为骨干本身的训练和推理能力，可减少辅助模块成本并形成自草稿。摘要未提供各下游基准数值、模型规模和加速实测，所称开销优势指训练多深度MTP的比较设置。

Abstract: Existing Multi-token prediction (MTP) methods are auxiliary by design: the multi-horizon loss mostly trains modules attached to the backbone, and each added horizon costs another module and another pass. We propose Efficient Causal Self-Recursion (ECHO), which removes both costs. The backbone’s own final layer predicts every horizon, so the supervision lands on the parameters that generate the next token, and Dynamic Depth Sampling (DDS) keeps deeper horizons affordable by drawing the training depth with probabilities that decay along depth. Future prediction thus becomes part of the autoregressive objective rather than a task attached to it, and the gains follow on both sides of it. In pretraining, added depth converts into downstream accuracy where auxiliary modules saturate, concentrating on reasoning and code at half the MTP training overhead of training every depth. At inference, the pretrained model is its own drafter: it drafts beyond the depth it was trained for, and keeps its draft-target alignment on the open-ended, high-entropy generation where post-hoc drafters lose theirs. The advantage survives a doubled token budget, a repeated pretraining run, and supervised finetuning.

[15] Ask, Don’t Guess: Targeted Clarification for Speech Recognition Errors cs.CLPDF

Sankalp Parashar, Preethi Jyothi, Sunita Sarawagi

TL;DR: 该方法通过流式识别可能受更清晰声学条件改善的语音错误片段，并只请求用户重说这些片段，以减少实时语音识别错误。

Details

Motivation: 远场实时对话语音识别仍会出现较多错误。如果能在流式过程中找出最可能通过重新发音纠正的片段，就可请说话者在更清晰环境下重复，避免整句重录。

Result: 方法用干净音频估计识别器在更好声学条件下的输出变化，将噪声和干净解码通过pair-HMM概率对齐生成软监督。反事实收益评分器结合声学和语言特征生成token级分数，再聚合为需澄清的片段；用户重复后的ASR转录由LLM只用于修订选中范围。在精确率70%时，该方法达到最高召回，相比最强基线最多多识别52%的错误token。定向重复在同时计入打断频率和重说片段长度的预算下，最高使相对词错误率下降24%。

Insight: 请求重说的价值取决于错误片段是否能在更好声学条件下改变，而非仅看识别器当前置信度。摘要未提供绝对打断次数、用户体验或重复语音质量变化，WER收益为最高值且受交互预算约束。

Abstract: Speech recognition in real time in far-field conversational settings still incurs significant errors. Identifying potential error spans during streaming ASR permits a recourse mechanism where the speaker can be asked to repeat the affected content under cleaner acoustic conditions. We propose \sysname, a new framework for online error localization, that identifies spans likely to benefit from this intervention. Our supervision uses clean audio to estimate how the recognizer’s output might change under improved acoustic conditions. We probabilistically align noisy and clean decodes using a pair-HMM to obtain soft supervision. These soft targets train a counterfactual gain scorer by combining a suite of acoustic and language features to produce token-level scores, that are further grouped into spans for targeted recourse. An LLM uses the ASR transcript of the repetition to revise only the selected span. At a precision of $70%$, \sysname achieves the highest recall by identifying up to 52% more erroneous tokens compared to the strongest baseline. Targeted repetition yields relative WER reductions of up to 24% under a budget that accounts for both interruption frequency and the lengths of re-uttered spans.

[16] Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning cs.CL | cs.IRPDF

Kyuyoung Kim, Zilin Xiao, Zhiyuan Wang, Avinash Atreya, Ke Yang

TL;DR: RPL让检索增强语言模型把搜索原语组合成可执行程序，并通过语法约束、执行验证的程序变异扩展强化学习搜索空间。

Details

Motivation: 多步搜索可帮助检索增强模型解决知识密集问题，但现有系统通常固定检索流水线或只暴露原子工具。更灵活的检索程序能自适应搜索及处理结果，却面临多步协调探索困难和强化学习反馈稀疏问题。

Result: 检索程序学习RPL允许策略把搜索原语组合成带局部状态和控制流的可执行程序，使其能广泛检索、处理中间结果并选择性构造模型上下文，而不让检索广度直接消耗全部上下文。为缓解探索难题，程序变异在语法限制和执行验证下，对RL采样的次优程序生成局部变体，将其转为训练轨迹后蒸馏回策略。在开放域问答和语料级推理任务上，RPL跨模型规模及分布内外基准均改善问答表现，多跳任务增益尤其大；4B模型超过所有8B基线。在语料级推理上，两个规模的回答F1和文档覆盖F1均最高，同时显著减少上下文消耗及检索轮数。

Insight: 可编程检索把搜索策略从预设流程扩展为可学习的程序，执行验证的变异则提供在线采样之外的探索。摘要未给出各基准具体分数、程序搜索成本和错误程序比例，效率收益需结合这些开销衡量。

Abstract: Retrieval-augmented language models increasingly use multi-step search to solve knowledge-intensive tasks, yet existing systems typically expose retrieval through fixed pipelines or collections of atomic tools. We introduce Retrieval Program Learning (RPL), a programmable search framework that allows a policy to compose search primitives into executable programs with local state and control flow. This enables adaptive search that can retrieve broadly, process intermediate results, and selectively construct model context without tying retrieval breadth to context consumption. However, the resulting flexibility creates a difficult exploration problem: useful programs may require several coordinated decisions, while reinforcement learning often supplies only sparse feedback. To address this challenge, we propose program mutation, a grammar-constrained and execution-verified procedure that generates local variants of suboptimal programs sampled during RL, converts them into training trajectories, and distills them back into the policy. We evaluate RPL on open-domain question answering and corpus-level reasoning tasks, comparing it with a range of retrieval and reasoning baselines. RPL consistently improves open-domain QA performance across model sizes and in- and out-of-domain benchmarks, with particularly large gains on multi-hop tasks. Notably, a 4B model trained with RPL outperforms all 8B model baselines. On corpus-level reasoning, RPL achieves the highest answer F1 and document-coverage F1 for both model sizes while substantially reducing context consumption and retrieval turns. Our results show that programmable search expands the retrieval strategies that agents can learn, while program mutation provides an effective mechanism for exploring useful program structures beyond direct on-policy sampling.

[17] Epi: Towards a Persistent, Self-Improving Agent for Autonomous Research cs.CL | cs.AIPDF

Yue Huang, Zhangchen Xu, Xiaonan Luo, Zheyuan Liu, Lichi Li

TL;DR: Épi是可跨模型会话持续工作的自主研究智能体，通过持久历史、层级记忆、分支探索和独立验证累积研究经验。

Details

Motivation: 自主研究需要长期维护实验、学习过往经验并可靠验证结果；普通智能体跨模型会话时容易丢失连续性，也可能缺少独立于提案者的结果核查。

Result: Épi结合持久研究历史、层级记忆、分支探索和独立评审，使实验跨会话运行，由固定验证器检查结果。系统报告120项创纪录结果：12个系列的20项数值界纪录和8个系列的100项构造纪录，其中n皇后常数区间约缩窄84倍，四次Feigenbaum吸引子维数的严格区间缩窄超过360万倍；精确对偶证书还证明Cohn–Elkies线性规划界在10和11维不能紧。研究还给出分裂图中Tuza猜想的计算机辅助证明，限定为8顶点团部件且三角形活跃邻域类型不超过三种、重数任意。轻量配置Épi-turbo将流程扩展至算法编程；30项FrontierOR任务总分比Codex高5.4%，模型成本为其64%，且使用相同底层模型和推理力度。

Insight: 跨会话记忆与固定验证器支持持续研究和可检查结果，论文也将数值发现连接到明确方法和数学论证。多项数学纪录及受限猜想证明需由专家独立审查；摘要未给出全部任务、基线细节和人类参与成本。

Abstract: We present Épi, an autonomous research agent that sustains investigations across model sessions and evolves its methods through accumulated experience. Épi combines persistent research history, hierarchical memory, branching and independent review to maintain research continuity, learn from experiments, and explore and revisit alternative approaches. Its experiments run across session boundaries, and a fixed verifier evaluates their outputs. Épi produced 120 record-setting results, comprising 20 numerical bound records across 12 families and 100 construction records across eight families. These include an approximately 84-fold narrower interval for the $n$-queens constant and a more than 3.6-million-fold narrower rigorous interval for the quartic Feigenbaum attractor dimension. Exact dual certificates establish that the Cohn-Elkies linear-programming bound cannot be sharp in dimensions 10 and 11. Beyond numerical improvements, Épi gives a computer-assisted proof of Tuza’s conjecture for split graphs with an eight-vertex clique part and at most three triangle-active neighborhood types, with arbitrary multiplicities. We extend this research process to algorithmic coding with Épi-turbo, a lightweight configuration that develops executable solvers for tasks including vehicle routing, scheduling and production planning. On 30 FrontierOR tasks, Épi-turbo achieves a 5.4% higher total benchmark score than Codex at 64% of its model cost, using the same underlying model and reasoning effort. Épi is designed to support human understanding by connecting computational results to explicit methods and mathematical arguments that researchers can examine and build upon.

[18] DeGMatch: Target-Aware Reasoning Data Selection via Deconfounded Gradient Matching cs.CLPDF

Changho Lee, Janghoon Han, Rajdeep Mukherjee, Yeda Song, Lening Nick Cui

TL;DR: DeGMatch以目标相关门控和去混杂梯度匹配筛选推理训练数据，降低候选群体共有梯度方向对目标能力区分的掩盖。

Details

Motivation: 长轨迹强化学习后训练提升推理能力，但大规模训练数据成本高；目标感知数据选择可用小参考集指定希望强化的能力。现有方法对长推理轨迹的目标区分有限，候选梯度中的群体共同结构可能遮蔽更细的目标相关变化。

Result: DeGMatch先构造目标相关性门控，再用模型特定梯度重排候选训练数据。它按候选总体梯度均值中心化，并移除主导主成分方向，以减少群体层面共有结构对目标匹配的影响。在数学、科学和代码任务上，DeGMatch超过强启发式、目标感知及推理专用数据选择基线，并在不同选择预算、模型家族和规模下保持有效。

Insight: 梯度去混杂可使少量目标样本对训练轨迹排序更有辨别力，降低大规模长推理数据选择成本。摘要未报告具体指标、筛选计算成本和不同目标能力的逐项结果，效果幅度无法量化。

Abstract: Reasoning post-training on long trajectories has substantially improved the reasoning capabilities of large language models, but scaling such training to large datasets is increasingly expensive. We study target-aware reasoning data selection, where a small reference set specifies a target capability and guides the selection of training trajectories for specialization. We find that existing target-aware methods can exhibit limited target discrimination on long reasoning trajectories, while dominant population-level structure in candidate gradients can obscure finer target-dependent variation during gradient matching. To address these limitations, we introduce Deconfounded Gradient Matching (DeGMatch), which first constructs a target-relevance gate and then reranks candidates using model-specific gradients. DeGMatch centers gradients using the candidate-population mean and removes leading principal directions before target matching, reducing the influence of dominant population-level structure and improving target discrimination. Across mathematics, science, and code, DeGMatch outperforms strong heuristic, target-aware, and reasoning-specific data selection baselines, while remaining effective across selection budgets, model families, and scales. Our code is publicly available at https://github.com/CHLee0801/degmatch.

[19] Rubrics Carry the Standard: Shapley Attribution for Checklist-Based LLM Judges cs.CLPDF

Saeid Asgari, Rui Ying, Sara Malvar, Rakshanda Agarwal, Bruce Sun

TL;DR: 该研究指出LLM评分量表内容比评审模型本身更能决定通过率，并用Shapley归因识别量表中有效检查与误拒来源。

Details

Motivation: 清单式评分量表常用于LLM评审开放答案及训练奖励，常见假设是宽严主要是评审模型属性、量表可通过模型自审改进；这两种假设都缺少控制实验检验。

Result: 作者在植入缺陷的受控评分任务和SummEdits、RAGTruth两个人工标注数据集上测试。固定评审模型和输出，仅改变量表内容，正确答案通过率可从2%变到100%；换评审模型或改变相同检查项的呈现方式，变化仅数个百分点。模型自审会将量表朝修订模型自身标准漂移，添加任务未要求的条件并删除人工标签教会的检查；三个修订模型在全部三个数据集都出现此现象。对于“全部条件均须通过”的聚合方式，研究给出每个检查项Shapley值的闭式解，可在线性时间归因正确捕获、错误拒绝责任和独特捕获。归因发现自审会删除做出独特贡献的检查，却保留造成误拒的检查。结合这些量的变更控制门及对标签来源检查的冻结，几乎所有漂移修订链都能维持初始质量，三种评审模型给出相同决策。

Insight: 评分标准内容本身可主导结果，模型自我审查量表也可能引入目标漂移；应逐项归因检查价值并追踪人工标注来源。结果适用于其人工及合成评测设置，“几乎所有”链未给出精确比例，不能直接外推至所有评分量表。

Abstract: Checklist rubrics applied by large language model (LLM) judges have become a standard instrument for evaluating open-ended outputs and for rewarding them during training. Two assumptions surround this practice: that leniency or strictness is a property of the judge model, and that a rubric can be improved by asking a model to review it. We test both in a controlled grading task with planted flaws and on two human-labeled datasets, SummEdits and RAGTruth. Holding the judge and the outputs fixed, rubric content moves the pass rate of correct outputs from 2% to 100%, whereas changing the judge model or the rendering of the same checks moves it by a few points. Model self-review then drifts the rubric toward the reviser’s own standard: it adds requirements the task never stated and deletes checks that human labels had taught, with three reviser models and on all three datasets. To measure and control this, we show that under all-must-pass aggregation the Shapley value of each check has an exact closed form, which yields per-check catch credit, false-fail blame and unique catches in linear time. Attribution shows that self-review deletes the checks that do unique work and keeps the checks that cause false fails. A change-control gate that combines these quantities with a provenance freeze on label-derived checks holds nearly every drifting revision chain at its starting quality, with identical decisions under three judge models.

[20] P2N: Reusing Deep Representations for Greater Effective Depth cs.CLPDF

Yiqin Hao, Boyi Zeng, Zitong Wang, He Li, Shixiang Song

TL;DR: P2N复用前一个token已经计算好的深层隐藏状态，扩展模型的有效计算深度，而每步解码仅增加一次向量加法。

Details

Motivation: 更强推理和自主智能体需要更大的计算深度，但增加模型层数会提高训练和推理成本，测试时扩展也需额外计算。

Result: Previous-to-Next机制将前一token的深层隐藏状态加到当前token浅层表示，使后续层复用已完成的深层计算，并通过跨token递归增加有效计算深度。训练使用限制在中间层核心区域的并行Jacobi更新。在Qwen风格模型的1.5B参数设置下，GSM8K零样本和少样本准确率均提高超过10个百分点，解码延迟几乎不变；收益还出现在语言建模和通用下游基准平均表现上。

Insight: 生成过程中前序token的深层表征可作为已有计算资源，帮助当前token增加有效深度，而不必额外扩宽或加深网络。摘要未提供具体GSM8K分数、模型其他规模结果和不同推理任务表现。

Abstract: Mathematical problem solving and autonomous agents are increasingly important applications of language models, demanding stronger reasoning capabilities. Computational depth is a key resource for such reasoning, but adding model layers increases both training and inference costs, while test-time scaling incurs substantial additional inference computation. We introduce P2N (Previous to Next), which reuses the previous token’s deep representations, already available during generation, to increase effective computational depth with negligible decoding overhead. Specifically, P2N adds the previous token’s deep hidden state to the current token’s shallow representation, allowing subsequent layers to build on previously completed deep computation. This recurrence extends effective computational depth across tokens, while introducing only a single vector addition per decoding step. We train the recurrence using parallel Jacobi updates restricted to a middle-layer core. Experiments with Qwen-style models show that, at 1.5B parameters, P2N improves GSM8K accuracy by over 10 percentage points under both zero-shot and few-shot prompting, with nearly unchanged decoding latency. The benefits also extend to language modeling and average performance on general downstream benchmarks. on general downstream benchmarks.

[21] Beyond Factual Accuracy: A Benchmark for Evaluating Perspective Alignment in LLMs cs.CLPDF

Sunghoon Kim, Sangyeop Kim

TL;DR: PerspectiveBench评测语言模型能否遵守用户指定的观点范围，发现模型常包含目标观点却未能排除未请求的观点。

Details

Motivation: 语言模型回答可能事实正确，却混合多种有效观点，迫使用户从不需要的内容中筛选；传统事实准确率指标无法检测这种观点控制失败。

Result: 作者构建多个合理推理路径并存的情境，提出观点对齐分数PAS衡量模型是否遵守指定观点。实验显示，先进模型往往能包含目标内容，却不能排除用户未要求的观点。通过检索增强生成提供文档能改善事实依据，但无法解决观点漂移，因为模型仍倾向于追求全面覆盖而非严格遵守用户约束。

Insight: 观点遵循与事实准确性是不同能力，RAG增加证据并不必然让模型收敛到指定视角。摘要未列出基准规模、模型列表和PAS具体成绩，观点限定也不应被误解为要求模型隐瞒相关事实。

Abstract: Large Language Models often generate factually correct responses that fail to maintain user-requested perspectives, blending multiple perspectives and forcing users to filter through unwanted information—a perspective control failure that factual accuracy metrics cannot detect. We introduce PerspectiveBench, a benchmark quantifying perspective alignment by constructing scenarios where multiple valid reasoning paths coexist, and propose the Perspective Alignment Score (PAS) to measure adherence to specified perspectives. Experiments reveal that state-of-the-art models successfully include target content but fail to exclude non-requested perspectives. Providing documents through RAG improves factual grounding but does not resolve perspective drift, as models prioritize comprehensive coverage over strict user constraints. Code and data are available at https://github.com/sunghoon014/PerspectiveBench.

[22] Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models cs.CLPDF

Rubin Wei, Siyuan Sun, Hao Sun, Yuliang Liu, Xiang Shi

TL;DR: DFM是一种块级参数化记忆，通过因果预测器从先前上下文生成记忆并以门控交叉注意力融合到中间隐藏状态，无需推理时显式检索。

Details

Motivation: 大语言模型需获取预训练之外的外部知识；继续预训练可能灾难性遗忘，推理时检索会增加开销。参数化外部记忆可免在线检索，但现有构建依赖大量离线搜索，输出层插值也只能浅层融合。

Result: DFM让因果记忆预测器根据前文上下文生成记忆，以语料中检索到的块嵌入作为监督；同一块内token共享预测记忆，并通过门控交叉注意力融合至隐藏状态。记忆预测和融合在骨干冻结时端到端训练，推理不需显式检索。相比常见token级记忆目标，DFM构造训练数据所需检索查询减少32倍、检索条目减少2,048倍。六个QA任务上，DFM相较最强基线在Mistral-7B-v0.3和Qwen3-8B-Base分别提高12.71%和11.19%。领域适配平均增益相较Qwen3-4B-Base和Qwen3-8B-Base分别为6.51和4.38分。

Insight: 共享块级记忆可减少记忆数据构建搜索开销，并在中间层融合外部知识。摘要未说明百分比是相对提升还是百分点，也未给出存储及推理内存开销，数值需按论文指标定义解释。

Abstract: Large language models need effective ways to acquire and use external knowledge beyond what they retain from pretraining. Continued pretraining risks catastrophic forgetting, while retrieval at inference time can add substantial overhead. External parametric memory avoids online retrieval, but constructing its training data requires extensive offline search, and output-level interpolation provides only shallow fusion. We introduce Deep Fusion Memory (DFM), a parametric memory architecture that reduces this construction workload and incorporates learned memory into intermediate hidden states. A causal memory predictor learns to generate memory from the preceding context, using retrieved chunk embeddings from the corpus as supervision. The predicted memory is shared across tokens within each chunk and fused into their hidden states through gated cross-attention. Memory prediction and fusion are trained end to end with the backbone frozen, and inference requires no explicit retrieval. Compared with existing parametric memory approaches, which often construct token-level memory targets, DFM requires $32\times$ fewer retrieval queries and $2048\times$ fewer retrieved items to construct its training data. Across six QA tasks, DFM performs better over the strongest baseline by 12.71% on Mistral-7B-v0.3 and 11.19% on Qwen3-8B-Base. Beyond QA, domain adaptation yields average gains of 6.51 and 4.38 points over Qwen3-4B-Base and Qwen3-8B-Base, respectively.

[23] TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis cs.CL | cs.AI | cs.LGPDF

Xiaoyu Wu, Zhaoxin Yu, Xinglin Xiao, Nan Xu

TL;DR: TIRE用可复用物理公式及其适用条件控制物理题合成，并先通过可靠执行确定标准答案，为监督微调和强化学习生成可验证数据。

Details

Motivation: 提升大语言模型物理推理需要大规模数据，将物理定律、具体情境和可验证数值答案联系起来。专家设计题目能自上而下控制公式和条件，但人工成本高；直接让LLM从文本生成问题虽可扩展，却容易带入答案和评审幻觉。

Result: TIRE以公式及其适用条件进行自上而下实例化，控制问题中的物理关系和数值配置；可靠执行在题目生成前建立标准答案，为问题构造和教师蒸馏提供统一数值基础。作者构建TIRE-19k，覆盖九个物理子领域。在三个骨干LLM、四个基准上，等量SFT数据时TIRE-19k超过五个代表性数据集，平均比最强基线高2.86个百分点；进一步GRPO使Qwen3-8B相较TIRE-19k SFT检查点平均准确率提高1.52个百分点。

Insight: 显式物理知识和执行验证可提高合成问题的可控性，并为答案监督提供可靠锚点。摘要未给出各物理领域的分项成绩、公式库构成和执行覆盖范围，结论限于九领域的构造与四项基准。

Abstract: Advancing physics reasoning in large language models requires scalable training data that connect physical laws, concrete problem scenarios, and verifiable numerical answers. Expert-designed problems offer top-down control over physical laws, conditions, and scenarios, but require costly human effort. LLM-based synthesis from texts scales efficiently, but model-generated answers and assessments leave supervision vulnerable to hallucinations. We introduce TIRE, a framework for Top-Down Formula Instantiation with Reliable Execution that brings knowledge-driven problem construction to large-scale data generation. Top-down instantiation uses reusable physical formulas and their applicability conditions to control the physical relations and numerical configurations of generated problems. Reliable execution establishes reference answers before problem generation, providing a common numerical foundation for problem construction and teacher distillation. Together, these mechanisms combine control over problem content with execution-grounded supervision for both supervised fine-tuning and reinforcement learning. Using TIRE, we construct TIRE-19k, spanning nine subfields of physics. Across 3 backbone LLMs and 4 benchmarks, TIRE-19k outperforms 5 representative datasets with equal amounts of SFT data, exceeding the strongest baseline by 2.86 percentage points on average. Further GRPO raises Qwen3-8B’s average accuracy by 1.52 percentage points over its TIRE-19k SFT checkpoint. These results demonstrate the value of generating physics training data from explicit physical knowledge with reliable execution.

[24] A Thousand Faces: A Persuasion Benchmark with Experience-Grounded Persuadees Informed by Psychology and Pragmatics cs.CLPDF

Ziqing Zhuang, Dingyi Zhang, Deyu Zhou

TL;DR: ATF-Persu是一个包含1,000种人物设定、覆盖200个主题和九个领域的说服基准，将个人经历与观点、表达方式及改变理由联系起来。

Details

Motivation: 人们可以持有相同信念和愿望，却因形成经历不同而需要不同理由才能被说服。只描述人物信念和愿望的基准可能奖励表面贴合描述的论证，却忽略这些观点来源和可促成改变的个人理由。

Result: ATF-Persu基于人格、态度理论和会话含意构造人物设定，在200个主题、九个领域中提供1,000个persona；每个事实情境创建五个persona，将经历关联到判断、表达及修正理由。评估显示，在各自协议下，ATF-Persu说服成功率低于三个对照基准。提供完整背景时，模型更多使用给定的个人理由，persona忠实度也高于简化说明；盲评者在可识别个人理由、这些理由对回答的影响及语境表达方面均更认可ATF-Persu而非CToMPersu。

Insight: 评估说服质量应检查论证是否回应个人观点形成过程，而非只确认是否匹配信念描述。摘要未给出成功率具体值和评审者规模，且不同基准协议不同，跨基准成功率比较需谨慎。

Abstract: People can express the same beliefs and desires yet need different reasons to change their minds. What makes an argument relevant depends on how those views arose: experiences shape what a person finds credible, what matters to them, and what could prompt reconsideration. These personal grounds may also be expressed indirectly, through contextual clues in conversation. Benchmarks that specify beliefs and desires without these connections can reward arguments that match a description while overlooking whether they address the person’s reasons. We introduce A Thousand Faces (ATF-Persu), with 1,000 personas across 200 topics and nine domains. Drawing on personality and attitude theories and conversational implicature, we construct five personas per factual context, linking experiences to judgments, expression, and grounds for revision. Evaluations show lower persuasion success than three comparison benchmarks under their respective protocols. Complete backgrounds yield broader use of supplied personal grounds and higher persona faithfulness than simplified specifications; blinded reviewers also rate ATF-Persu above CToMPersu on identifiable personal reasons, their influence on responses, and contextual expression. The dataset is publicly available at https://huggingface.co/datasets/Slandes/ATF-Persu.

[25] Calibration-Free Early Stopping for Chain-of-Thought Reasoning via E-processes cs.CL | cs.AI | cs.LGPDF

Dong-Jae Lee, Gyojin Han, Junmo Kim

TL;DR: 该研究把思维链提前停止视为序贯证据累积问题，用两个e-process分别跟踪提交与弃答信号，无需留出集校准即可提供任意时刻的信号级误报界。

Details

Motivation: 推理模型受益于测试时扩展，但需要自适应决定何时继续花token、何时提前结束。现有方法依赖置信度或稳定性阈值，通常难跨模型和数据集迁移；基于验证集校准的规则在测试时长或信号变化后也可能失去保证。

Result: 框架沿推理轨迹对每个分块信号运行两个e-process：提交流跟踪信号相对慢趋势的位置，弃答流跟踪带符号速度，二者按构造反相关。在分块增量的信号级零假设下，Ville不等式给出用户指定显著性水平的任意时刻有效误报界，不需留出校准数据。多种模型—数据集组合实验显示，该方法使用零校准数据时，在token—准确率Pareto前沿上持平或超过校准式停止；当校准不可行时，显著性水平可不变地跨设置迁移。

Insight: 序贯检验为CoT停止提供可迁移的误报控制，而不是调一个模型特定阈值。理论保证针对每块信号的零假设，不自动等于最终回答正确率保证；摘要未提供具体基准、token节省和停止效用分数。

Abstract: Reasoning large language models (LLMs) benefit from test-time scaling, motivating adaptive policies that spend tokens when they help and stop early otherwise. Existing rules halt when a confidence or stability signal crosses a threshold, but such thresholds rarely transfer across models and datasets, and validation-calibrated rules lose their guarantee under shifts in test-time horizon or signal dynamics. We reframe adaptive stopping as sequential evidence accumulation along the reasoning trajectory rather than threshold calibration on held-out data. Our framework runs two e-processes on a shared per-chunk signal: a commit stream tracks the signal’s position relative to its slow trend, and an abstain stream tracks its signed velocity, making the two anti-correlated by construction. Under signal-level nulls on the per-chunk increments, Ville’s inequality yields an anytime-valid signal-level false-alarm bound at user-specified significance levels, with no held-out calibration. Across various model–dataset combinations, the framework matches or exceeds calibration-based stopping on the token–accuracy Pareto front using zero calibration data, and the same significance level transfers unchanged across cells where calibration becomes infeasible.

[26] Atoms, Not Knowledge: What Production-Scale Addressed-Memory Writes Deposit cs.CLPDF

Wei Gao, Cheng Tian, Duan Yang, Jianqing Yin

TL;DR: 该研究评测哈希寻址条件记忆写入实际存下什么，发现事实可单独检索和定位，却难以组合成多跳推理所需关系。

Details

Motivation: 确定性哈希寻址记忆允许事后局部注入知识，但生产规模下写入覆盖率、组合能力和检索失败机制尚未充分验证；仅看记忆表大小无法说明系统能否用于复合推理。

Result: 研究测试124M至7B自训练或公开预训练记忆模型、Qwen3.8-Flash-Next-FP8及DeepSeek-V4.1-Flash。覆盖需要骨干已有充分训练先验、全宽读出和可训练门控；欠训练干预显示存在能力阈值。固定预算下，覆盖率从50条事实时的98%降到1,000条时的12.7%，表规模本身不是瓶颈。双跳组合未超过独立单跳基线：观测26%，预期25%，词汇扰动底限15%。两个独立记忆族及不同写入剂量/负载下均未见协同。给定中间实体与模型自行回忆之间有28个百分点差距，指向“检索以服务推理”失败而非门控不可达。写入事实可独立检索、精确定位且可逆，但不能自行组合推理。生产负载符合指数干扰拟合（R²=0.98），具体机制仍未确定；新52目标复跑与归档基线一致。

Insight: 地址记忆写入应分别报告覆盖、组合协同和扰动率，不能因单条可检索就宣称记忆支持多约束推理；多跳组合需由系统显式编排。摘要中模型、负载和测试范围较具体，生产拟合的干扰机制仍开放。

Abstract: Deterministic hash-addressed conditional memory permits local post-hoc knowledge injection. Extending arXiv:2610.21553’s bounded-feasibility audit rejecting under-specified generality claims with repaired probes and training-dose/capacity axes, we give, to our knowledge, the first production-scale write-content verdict. We test 124M-7B self-trained/publicly pretrained beds, Qwen3.8-Flash-Next-FP8 (176.95B; 320,001,536 FP8 rows), and DeepSeek-V4.1-Flash. Write coverage requires a sufficiently trained backbone prior, full-width readout, trainable gates; undertrained-bed intervention confirms a threshold, not a slope. At fixed budget, coverage falls from 98% (50 facts) to 12.7% (1,000); table size is not binding. Two-hop composition does not exceed the independent single-hop baseline (26% observed vs 25% expected; 15% lexical-distortion floor). Synergy remains absent across dose/load in two independent bed families. A 28-percentage-point in-context gap (given vs recalled intermediates) localizes failure to retrieval-for-reasoning, not gate access. Writes are independently retrievable, precisely addressable, reversible atoms, not composable into reasoning. A layered protocol, the first, to our knowledge, to cover compositionality in behavioral memory evaluation, separates baseline, synergy and distortion floor. Any addressed/key-value memory must report coverage, synergy ratio and distortion rate before deployment claims. Composition must be system-orchestrated; multi-constraint queries should explicitly chain written atoms. In-context controls show given intermediate entities are exploited at twice the implicit rate. Production load fits exponential interference ($R^2=0.98$); its mechanism, between budget dilution and write interference, remains open. All numbers are recomputed from archived raw logs/JSON; a fresh 52-target production rerun gives 52/52, exactly matching archived baselines.

[27] Where and Why Search-Augmented LLMs Fail When Reasoning over Rapidly Evolving Knowledge cs.CL | cs.AI | cs.IR | cs.LGPDF

Tianyi Hu, Bhavyajeet Singh, Niket Tandon, Akhil Arora

TL;DR: TREK通过逐阶段追踪检索和推理错误，评测模型应对快速变化知识的能力，发现主要瓶颈常在检索后的冲突证据推理。

Details

Motivation: 新信息不断发布而旧资料仍在线，大语言模型需检索证据并处理冲突，判断哪些信息仍有效。把整个信息获取流程当黑箱的评测无法定位模型具体在哪个阶段失败。

Result: TREK包括三个组件：TREKTRACE可插入流程并将错误归因到检索和推理阶段；TREKBENCH-WILD是需要对快速演化真实知识作复杂推理的问题元基准；TREKBENCH-CTRL则控制时间性、相关性和可信度干扰项。作者发现，主要瓶颈不是检索，而是检索后的推理：即使取回回答所需全部信息，前沿模型仍可能被相互冲突证据误导。以过时干扰项为主的基准还会低估任务难度。

Insight: 评估快速变化知识问答应分别衡量证据检索和冲突处理，并控制过时、相关性及可信度干扰。摘要未提供基准规模和模型成绩，TREK的主要贡献是错误定位与诊断测试床。

Abstract: As the world changes, new information is published while older records remain online. This complicates reasoning for large language models (LLMs), which must retrieve evidence and resolve conflicts to determine which information remains valid. Existing evaluations treat information seeking as a black box, revealing little about where and why LLMs fail. To address this gap, we introduce TREK, a suite of three components: TREKTRACE is a plug-and-play framework that attributes failures to specific stages of retrieval and reasoning. We apply it to TREKBENCH-WILD, a meta-benchmark of real-world questions that require complex reasoning over rapidly evolving knowledge. Surprisingly, the main bottleneck is post-retrieval reasoning rather than retrieval: frontier models are often misled by conflicting evidence even after retrieving all the information needed to answer. To study this bottleneck, we build TREKBENCH-CTRL, a controlled benchmark with distractors spanning temporality, relevance, and trustworthiness, and show that benchmarks dominated by outdated distractors substantially underestimate task difficulty. Together, TREK provides a testbed for diagnosing and improving reasoning over rapidly evolving knowledge.

[28] From Unverifiable Domain Text to Cross-Lingual RLVR via Code-Switched Task Synthesis cs.CLPDF

Cheonbok Park, Jihwan Kim, Shinyoung Joo, Joosung Lee, Hwiyeol Jo

TL;DR: 该研究从不可验证的英文领域文本合成目标语言及代码混合RLVR任务，在韩语医学和四种目标语言网络安全基准上改善目标语言推理表现。

Details

Motivation: 可验证奖励强化学习有助于提升大语言模型推理，但专业非英语领域缺少带可验证奖励的目标语言任务；英文领域文本虽较充足，却不能直接作为目标语言RLVR数据。

Result: 作者提出代码混合任务合成方法，从原本无法直接验证的英文领域文本构造目标语言与代码混合RLVR任务。以填空式多选问答为种子格式，从领域段落自动生成可验证任务，再变换为跨语言版本，保持证据—答案关系不变，同时改变英文字段与目标语言证据的呈现方式。在韩语医学基准和覆盖四种目标语言的网络安全基准上，代码混合训练平均超过匹配的单语和双语混合基线，同时英语表现相近。补充分析支持其改善目标语言输入理解，并加强输出和内部表示中的跨语言对齐。

Insight: 将英语领域知识转换为目标语言和代码混合的可验证任务，可缓解非英语专业RLVR数据短缺。摘要没有列出各任务成绩、具体目标语言和任务合成数量，迁移效果限于医学和网络安全测试。

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) can improve reasoning in large language models, but specialized non-English domains expose a data mismatch. Raw English domain text is often available, while target-language RLVR tasks with verifiable rewards are scarce. We introduce code-switched task synthesis for cross-lingual RLVR, a data-centric approach that derives target-language and code-switched RLVR task variants from otherwise unverifiable English domain text. Using fill-in-the-middle multiple-choice QA as a seed task format, we synthesize automatically verifiable tasks from domain passages and reshape them into cross-lingual variants. These variants preserve the same evidence-answer relation but change how English and target-language evidence are presented during RLVR. On Korean medical benchmarks and cybersecurity benchmarks across four target languages, code-switched training outperforms matched monolingual and bilingual-mixture baselines on average target-language performance while maintaining comparable English performance. Additional analyses provide evidence of better target-language input understanding and stronger cross-lingual alignment in both model outputs and internal representations. These results suggest that code-switched training helps target-language inputs access domain reasoning learned more strongly in English.

[29] RARE: Reinforcement Learning with Adaptive Rubric Evolution for Open-Ended Generation cs.CL | cs.LGPDF

Xueyun Tian, Bingbing Xu, Chuyue Li, Jiaming Liu, Chao Zhang

TL;DR: RARE根据当前策略输出反馈自适应演化开放式生成任务评分标准，只替换已满足且失去区分力的检查项，保留仍有用的标准。

Details

Motivation: 数学和代码有可验证奖励，但许多有价值的开放式任务无法自动判定整体质量，且通常要同时满足多个要求、又允许多种有效答案。评分标准能提供细粒度奖励，但训练进展不均可能使一些条目饱和，失去区分能力；动态改写又可能损害原有有效标准。

Result: RARE利用当前策略反馈同时决定哪些标准需要改变及替换内容。逐样本更新只将当前所有输出都已满足的标准替换为评估剩余质量差异的新检查，其他标准和得分保持不变。方法比较奖励期望优良品质与惩罚实际观察缺陷两种标准变更方向。视频脚本生成作为测试任务，要求故事连贯、可执行且满足约束；针对缺陷的策略在三类视频任务上平均比匹配静态标准提高3.82分。模型偏好和人类偏好评估均支持该收益，改进也扩展到写作和一般生成基准。

Insight: 局部更新可在标准失去区分力时增加反馈，而不重写整套量表；惩罚当前缺陷的策略在所测视频类型中效果最好。摘要未说明评分尺度、具体模型及各基准成绩，人类偏好评估细节也未给出。

Abstract: Verifiable rewards have advanced reinforcement learning in mathematics and code. However, many valuable open-ended tasks share three properties: no automatic verifier of overall quality, multiple requirements to satisfy jointly, and multiple valid responses to each query. Training therefore needs feedback on requirement satisfaction across different valid responses. Rubric-based reinforcement learning provides this feedback by aggregating scores from explicit criteria. As learning progresses unevenly, some criteria become fully satisfied and lose discrimination while others remain informative. Dynamic methods revise shared rubrics or select evolving criteria to restore discrimination, but may also change useful criteria. To address this, we propose Reinforcement Learning with Adaptive Rubric Evolution (RARE). The core idea is to use current policy feedback to guide both which criteria change and what their replacements assess. Sample-level updates replace criteria fully satisfied across current outputs with checks of remaining quality differences, retaining all other criteria and scores. Within this rule, we compare rewarding desirable qualities with penalizing observed flaws. Video script generation exemplifies these properties and supports content production. Its demands for coherent, executable narratives under constraints make it a challenging testbed. Targeting flaws performs best across three video types with a mean gain of 3.82 points over matched static rubrics. Model and human preferences support these gains. Improvements also extend to writing and general generation benchmarks.

[30] Design Considerations in Low-Resource Archival OCR: A Case Study in Lakota cs.CLPDF

Téa Wright, Laura X. Ma, Claudia Iron Hawk, Hannah J. Haynie, Alexis Palmer

TL;DR: 该研究用小型视觉语言模型对拉科塔语档案OCR进行后纠错，发现少量转录数据可显著降低字符错误率，但真实自动分割会使错误率近乎翻倍。

Details

Motivation: 数字档案有助于语言复兴，但低资源档案OCR需适配具体社区、语言和文档；受限训练预算、分割噪声、扫描退化和拼写变体都会影响工具在实际工作流程中的效果。

Result: 作者研究不超过3B参数的视觉语言模型作为拉科塔语OCR工具，除比较建模方法外，还检查上述四类实践因素。用35页转录数据微调VLM进行OCR后纠错，可将双语语料字符错误率从17.1%降至3.9%；只用9页训练时，CER相比35页设置仅高1.1%。然而，在现实自动分割条件下，最佳模型CER近乎翻倍，说明仅在理想分段下评估会高估实际使用效果。

Insight: 低资源语种少量社区转录数据也能明显改善OCR后纠错，但自动分割和扫描质量等流程因素可能抵消收益。摘要未给出文档总量、模型清单和分割错误的各类分解，结果限于拉科塔档案。

Abstract: Building and expanding access to digital archives supports language reclamation and creates lasting resources for community revitalization. Designing effective tools to support these efforts requires taking into account the community-, language-, and document-specific context. We investigate how small ($\leq$3B) vision-language models (VLMs) can be adapted as optical character recognition (OCR) tools for Lakota, a North American Indigenous language. In addition to benchmarking modeling approaches, we inspect four practical facets of low-resource archival OCR that arise in previous work and conversations with community members: constrained training budgets, segmentation noise, scan degradation, and orthographic variance. We demonstrate that fine-tuning VLMs for OCR post-correction on 35 pages of transcribed data decreases character error rate (CER) from 17.1 to 3.9% on a bilingual corpus, and find that training only with nine pages increases CER by only 1.1% over this. However, evaluating under realistic automatic segmentation nearly doubles the CER of the best-performing models, stressing the importance of taking realistic use cases into account. We hope this work encourages future research on developing language technologies grounded in the workflows and constraints of ongoing revitalization efforts.

[31] Quality-Controlled Reply and Summary Data for Post-Training Small Language Models in AI Call-Screening Agents cs.CLPDF

Jiaqi Gan, Tsang Ng, Ankit Raj, Tommy Duong, Xingyu Shen

TL;DR: 该研究为本地AI来电筛查构建遵守信息边界的合成回复与摘要训练集，适配后两项任务的模型评审合格率提高，但仍存在依据对齐问题。

Details

Motivation: 小语言模型有望在手机端替用户筛查电话，避免将通话内容发送到云端，但模型需要在通话中只代表用户权限回应，并在事后准确总结。现有小模型可能对普通来电过度报警、遗漏摘要条件或将来电者说法当成事实，而真实通话通常没有标准回复或摘要供训练。

Result: 作者从合成通话构建两类监督，并设置严格信息边界：来电者写作者可看私密情境说明，回复写作者只能看当前轮之前对话，摘要写作者可看完整转录；独立模型评审决定哪些通话纳入训练。语料含3,395通电话和15,102个目标（11,707回复、3,395摘要）。六个0.6B–4B模型使用一个共享LoRA适配器，在192个留出通话上评估。回复合格率从21%–54%升至94%–97%，摘要合格率从0%–6%升至43%–72%；六模型两任务共12组配对增益的bootstrap区间均高于零，摘要收益在所有来源家族上保持。适配后连贯性和自然度问题近乎消失，剩余挑战是忠实追踪谁在什么条件下说了什么。训练筛选和输出评分由同一个模型执行，因此合格率代表与该评审标准的一致性，而非人类认可。

Insight: 信息边界约束与分角色数据生成支持本地来电代理训练，但评审器自洽可能高估实际质量；作者仍计划做人类研究。摘要指标不能直接解释为用户接受率，且条件忠实度仍是主要不足。

Abstract: Small language models have improved to the point where a phone can plausibly screen its own calls, answering on an absent owner’s behalf without sending call content off the device, a privacy-respecting alternative to cloud call handling. Realizing that promise takes a model that can do two things well: reply within the owner’s authority during the call, and afterward tell the owner what happened and what needs attention. Released small models do neither reliably - they raise unwarranted alarms about ordinary callers and write summaries that drop material conditions or report a caller’s claims as fact - and real calls supply no reference replies or summaries to learn from. We construct supervision for both tasks from synthetic calls under an information-boundary constraint: a caller writer works from a private scenario brief, a reply writer sees only the conversation up to the current turn, and a summary writer sees the completed transcript. Separate model-based reviews of the full dialogue and the summary decide which calls enter training, yielding a corpus of 3,395 calls with 15,102 targets (11,707 reply, 3,395 summary). Training a single shared LoRA adapter on both tasks, we evaluate six models (0.6-4B parameters) before and after adaptation on 192 held-out calls. Reply qualification rises from 21-54% to 94-97% and summary qualification from 0-6% to 43-72%; all twelve paired gains (six models x two tasks) have bootstrap intervals above zero, and the summary gains hold in every source family. Because one model both screened the training data and scored the outputs, these rates measure agreement with that evaluator’s standard rather than human acceptance, which a planned human study will test. After adaptation, coherence and naturalness failures nearly vanish, leaving grounding and usefulness - faithfully tracking who said what, under which conditions - as the remaining challenge.

[32] Self-Distillation for Supervised Fine-Tuning cs.CLPDF

Wenxuan Zhu, Bang An, Shuming Liu, Wenqing Cui, Tianqi Xu

TL;DR: Gold-Conditioned Distillation让同一模型在见到标准答案的上下文中充当教师，为不含答案的学生提示提供软token监督，以保留正确答案之外的合理续写。

Details

Motivation: 监督微调把专家示范或标准答案视为one-hot目标，会压低模型原本可能给出的有效替代续写，也可能未充分利用预训练模型已有知识。

Result: GCD让同一个模型兼任教师和学生：教师在提示中看到标准答案上下文，学生则只看到推理时使用的原始提示；训练目标是在标准答案的每个token位置最小化学生与教师下一token分布的前向KL散度。标准答案负责锚定正确性，软分布保留其他合理可能性。在Qwen和Llama模型的多种训练任务中，GCD相较标准SFT提高新任务准确率，并通常减少通用基准性能下降。

Insight: 标准答案可作为教师可见的上下文，让模型对自身的先验知识形成软监督，而无需额外教师模型。摘要未给出具体任务、量化增益和蒸馏成本，结论以作者报告的总体趋势为限。

Abstract: Supervised fine-tuning treats the expert demonstration or gold response as a one-hot target, suppressing valid alternative continuations that the model would otherwise assign probability to. We propose Gold-Conditioned Distillation (GCD), a method that turns expert demonstrations into model-native supervision. GCD uses one model as both teacher and student: the teacher conditions on a prompt containing the gold response in context, the student conditions on the original prompt as at inference, and the student is trained to minimize the forward KL divergence to the teacher’s next-token distribution at each gold response token. This lets the model use its prior knowledge to teach itself the new task: correctness is anchored by the gold response, while the soft target preserves plausible alternatives. Across various training tasks on Qwen and Llama models, GCD improves new-task accuracy over SFT and generally reduces degradation on general benchmarks.

[33] EmotiQ: A Quantum Residual Specialist Framework for Text Emotion Classification cs.CLPDF

Carlton Kigonya, Tanaka Chirara, Yunting Yin

TL;DR: EmotiQ用边界专家补充冻结Transformer文本情绪分类器，仅在验证条件满足时介入，并在六个数据集上取得小幅但多数为正的macro-F1提升。

Details

Motivation: 文本情绪分类需要区分细微语言信号和语义边界相近的情绪类别。研究探索小型量子模块能否作为定向残差专家，帮助传统分类器处理困难情绪决策。

Result: EmotiQ以冻结Transformer分类器为主决策器，根据验证集行为识别困难情绪边界，并训练边界专家，包括量子编码、数据重上传、原型与保真度专家以及核方法专家。专家只有达到基于验证集的准入标准才可干预。在六个文本情绪数据集上，启用后的策略平均macro-F1超过基线，六个中五个获得正收益。量子组件里ZZ交互特征映射最稳定，平均macro-F1提高0.0030，六个中四个提升。

Insight: 将小型专家限制在特定困难边界并设置准入条件，可能为强经典分类器提供局部补充。平均提升幅度较小，摘要未给出误差条、量子硬件或模拟成本，不能推断量子计算相较经典模块具有效率优势。

Abstract: Emotion classification from text is challenging because many affective states are distinguished by subtle linguistic signals and semantically overlapping boundaries. We investigate whether compact quantum modules can serve as targeted residual specialists that assist traditional classifiers on difficult emotion decisions. We introduce EmotiQ, a residual specialist framework for text emotion classification using a frozen transformer classifier as the primary decision maker, identifies challenging emotion boundaries from validation behavior, and trains boundary specialists, including quantum encodings, data re-uploading specialists, prototype and fidelity specialists, and kernel-based specialists. Specialists are permitted to intervene only when they satisfy validation-based admission criteria. We evaluate the framework across six text emotion classification datasets and establish traditional model baselines. Experimental results show that the admitted EmotiQ policy improves mean macro-F1 metrics over the baseline, and achieves positive gains on five of six datasets. Among the quantum components, ZZ interaction feature maps are the most consistently effective, yielding a mean macro-F1 improvement of +0.0030 and positive gains on four datasets. These findings suggest that current quantum NLP methods can be deployed as residual specialists that provide localized improvements under strong classical supervision.

[34] Flattening Fantasies: Analyzing Stereotypes in Pornographic Discourse cs.CLPDF

Sarmad Chandio, Osama Khalid, Sharaf Zia, Ethan Kutlu

TL;DR: 该研究对5.4百万条色情平台视频标题进行定量语言分析，考察种族、族群和性别身份的表述模式及其与隐含规范的偏离。

Details

Motivation: 网络话语会反复塑造身份群体的刻板印象，但相关平台语料中的文化、种族和性别表述尚缺少大规模定量分析。

Result: 作者分析xHamster的540万条视频标题，研究国籍、族群和性别身份的呈现方式、刻板印象及其相对于色情内容隐含规范的偏离。结果显示，美国内容占主导并成为行业事实上的默认规范；还发现反映殖民权力关系的持续族群刻板印象、宗教恋物化和文化禁忌性化。跨文化比较显示性别呈现存在差异：女性常被置于圣女—荡妇二分框架，男性则常被刻画为身体上占主导地位。

Insight: 研究将平台标题中的身份叙事置于大规模语言模式层面分析，揭示文化默认规范与群体刻板印象。摘要未说明抽样时段、平台推荐机制和标注/分类方法，结果反映该平台文本而非整个成人内容产业或现实群体态度。

Abstract: This study presents a large-scale quantitative analysis of 5.4 million video titles from xHamster, the world’s fourth most visited pornography platform, to examine linguistic patterns used to describe different dimensions of identities (e.g., race and gender). Specifically, we investigate how nationalities, ethnic groups, and gender identities are portrayed in pornographic discourse, focusing on stereotypes and deviations from implicit norms of pornography. The results show the dominance of American pornography over other groups, which acts as the de facto norm for the adult industry. We identify persistent ethnic stereotypes reflecting colonial power dynamics, religious fetishization, and the sexualization of cultural taboos. The analysis reveals gendered representational differences across cultures, with a prevalent Madonna-Whore dichotomy in the portrayal of women and stereotypical depictions of men in physically dominant roles.

[35] Assertion and Belief: Moore’s Paradox in Language Models cs.CLPDF

Anshun Asher Zheng, Weili Cao, Fangcong Yin, David Beaver, Junyi Jessy Li

TL;DR: 该研究首次系统测试LLM对摩尔悖论句子的判断，发现模型能识别其言语行为不合宜，但更像记忆悖论规则，而非从说话者信念推理。

Details

Motivation: 像“正在下雨，但我不相信正在下雨”这样的句子可能内容上为真，却不适合由同一说话者断言，因为断言本身意味着为自己的信念背书。该悖论可检验模型是否将言语行为纳入语义理解。

Result: 作者在多种语境下构建摩尔悖论句子集，对12个模型进行评测。模型大多能识别这类句子的语用不合宜，且判断随言语行为变化，并非只识别固定句式。基于理性言语行为框架进一步测试听者是否通过推理说话者来得出判断，结果显示模型的“听者”并不推理其“说话者”；推理轨迹更像回忆悖论相关知识并套用规则。

Insight: 模型能判断某句不合宜，并不表明它按说话者信念和断言行为组合计算意义；模型可能只是掌握了关于句式的陈述性知识。摘要未列出各模型分数、语境数量和不同语种结果。

Abstract: Moore’s paradox \citep{moore1942} concerns sentences such as ``It is raining, but I don’t believe that it is,’’ which may well be true yet cannot be sensibly asserted. The absurdity lies not in the content but in the act, since to assert is to vouch for one’s own belief. This makes the paradox a test of whether language models take the speech act into account when computing meaning. We present the first test of Moore’s paradox in LLMs, evaluating 12 models on a curated set of Moorean sentences in varied contexts. We find that models largely recognize the infelicity of Moorean sentences, and that the judgement is not tied to a memorized form and tracks the speech act. Using the Rational Speech Act framework (RSA; \citealt{goodmanfrank2016}), we then test whether these judgements arise from a listener reasoning about a speaker. We find the models’ own listeners do not reason about their own speakers, and their reasoning traces suggest that they instead recall knowledge about the paradox and apply rules therein. The models thus appear to have learned the paradox as a fact about sentences, not as a requirement on speakers.

[36] Cognitive Map in Large Language Models cs.CL | cs.LGPDF

Albert Qin, Jonathan Cohen, Gautam Reddy

TL;DR: 该研究用关系推理和因果激活干预分析大语言模型是否形成类似认知地图的内部表征，发现大模型能在记忆的成对关系间追溯组合，但不一定预先建立全局坐标布局。

Details

Motivation: 动物常被认为会构建认知地图以支持新环境中的灵活行为。研究检验语言模型是否形成类似的关系结构表征，并通过激活干预确认模型使用哪些内部成分进行关系推理。

Result: 作者用关系推理任务和因果“马赛克实验”操纵开源模型内部激活。超过100亿参数规模的模型能推断未明说的实体关系，并泛化到合成实体名称。跨模型与任务，模型形成抽象关系结构表示，并将其与实体身份表示相加，将实体绑定到结构角色。模型更像存储观察到的成对关系、在查询时回溯组合，而非预先把每个实体放入全局坐标系。实体—角色绑定由稀疏分布的键值对介导；扰动这些对可植入可预测地改变模型行为的“假记忆”。

Insight: 模型的关系推理可能依赖可组合的局部链接和实体角色绑定，类似认知地图但不等于完整全局空间布局。摘要未列出模型数、任务性能和不同规模阈值，结论限于所测开放权重模型及关系任务。

Abstract: It has long been proposed that animals construct cognitive maps to support flexible behavior in novel environments. We tested whether large language models form similar representations using relational reasoning tasks and causal “mosaic experiments” that manipulate internal activations in open-weight language models. Models beyond the ten-billion-parameter scale successfully inferred unstated relationships between entities and generalized to synthetic entity names. Across model and task families, models generated abstract representations of relational structure and combined them additively with representations of entity identity, binding entities to roles within that structure. Models stored observed relationships as pairwise links that they composed retrospectively when queried, rather than prospectively positioning each entity in a global coordinate system. Binding was mediated by a sparse, distributed set of key–value pairs; by perturbing these pairs, we implanted “false memories” that altered model behavior in predictable ways. Overall, our findings reveal an internal organization analogous to a cognitive map that explains how language models reason with limited information.

[37] MapSteal: Map-based Autonomous Knowledge Stealing Attack Against Large Language Models cs.CL | cs.AIPDF

Zhengao Li, Chuxu Zhang, Yushun Dong

TL;DR: MapSteal从少量初始主题根出发，递归扩展教师模型提出的知识方向，在无需学生反馈的条件下自动生成问答训练数据。

Details

Motivation: 通过查询黑盒LLM获取知识可帮助小模型适配，但许多数据获取方法依赖预设查询或学生反馈，如何自主扩展知识目标仍是挑战。

Result: MapSteal将知识获取目标表示为结构化状态，从少量初始根节点出发，由教师提出方向并递归扩展。问题实现、候选方向准入和去重约束搜索，根节点感知调度器则在固定问答预算下组织探索。所得问答用于学生模型适配，获取过程中无需学生反馈。Qwen3.5的2B、4B和9B学生在六个医学及网络安全基准上表现有竞争力。4B医学评估中，完整MapSteal平均准确率比简化递归获取变体高1.17个百分点；跨域结果显示增益依赖具体领域和评测面板。

Insight: 结构化递归探索可在固定查询预算中扩大黑盒模型知识采集范围，但生成问题不等于得到有效训练监督。摘要强调领域和面板对结果有影响，且未给出各基准成绩、查询预算规模和与学生反馈方法的全面对照。

Abstract: Knowledge acquisition from black-box large language models (LLMs) offers a practical approach to adapting smaller student models. However, many existing methods rely on predefined queries or student feedback to guide acquisition, leaving autonomous expansion of knowledge targets an important challenge. We propose \textbf{MapSteal}, a student-independent framework for structured recursive knowledge acquisition. Starting from a small set of initial roots, MapSteal represents acquisition targets using structured states and recursively expands them through teacher-proposed directions. Question realization, proposal admission, and deduplication constrain the acquisition process, while root-aware scheduling organizes exploration under a fixed QA budget. The resulting question–answer pairs are used to adapt student models without requiring student feedback during acquisition. Experiments with Qwen3.5 students at 2B, 4B, and 9B scales show competitive downstream performance across six medical and cybersecurity benchmarks. On the 4B medical evaluation, Full MapSteal improves average accuracy by 1.17 percentage points over a simplified recursive acquisition variant. Cross-domain evaluations show that gains depend on the domain and evaluation panel. These results demonstrate the feasibility of structured, student-independent recursive acquisition while highlighting the distinction between generating acquisition targets and obtaining useful training supervision. Our code is available at \href{https://github.com/LabRAI/Mapsteal} {https://github.com/LabRAI/Mapsteal}.

[38] A Calibrated Instrument for Detecting Disposition Drift in Language Models cs.CL | cs.AI | cs.LGPDF

Jerry Kaplan

TL;DR: 该研究提出一套经校准的评测工具检测语言模型安全倾向变化，发现有害微调会显著改变模型倾向，而推理生成的回答可能掩盖这种变化。

Details

Motivation: 模型发布前会测试危险请求执行、欺骗、权力寻求和谄媚等风险，但之后的领域微调、数据变化或推理设置可能改变这些安全行为。仅用其他AI评分的汇总率难以解释偏移，也可能漏掉内部倾向变化。

Result: 作者将用于评估推理优化对输出质量影响的校准测量方法改用于错配行为评测，在两个开放权重模型的多个维度中先验证测量精度，显示几百分点的偏移可被可靠检测。通过在正确及轻微有害的医疗、金融建议上微调模型，发现有害建议引起的变化远大于正确建议。尤为明显的是，先推理再回答的模型可能几乎在最终回答中隐藏这种偏移，但直接读取其内部状态可发现倾向已变化；关闭推理后，偏移随训练数据中有害建议比例稳步增大。

Insight: 最终回答行为可能无法揭示模型内部安全倾向，安全评测应同时测试推理开启和关闭的条件。结果来自两个开放模型及特定医疗、金融微调方案，摘要未给出每个风险维度变化数值。

Abstract: Before language models are released, they are tested for harmful traits such as willingness to carry out dangerous requests, deception, power seeking, and sycophancy. The safe behavior these tests certify may degrade as models are modified, by topical fine-tuning, new training data, or changes to settings such as the depth of their reasoning. Most work reports such drift as single summary rates scored by another AI model, without the error analysis needed to interpret them. We adapt a calibrated measurement methodology recently developed for assessing how inference optimizations affect the quality of model outputs. Applying it to several misalignment dimensions in two open-weight models, we first establish the instrument’s precision, showing that shifts of a few percentage points can be reliably detected. We then induce drift deliberately by fine-tuning on correct and on subtly harmful medical and financial advice; harmful advice induces a far greater shift than correct advice. Most strikingly, a model that reasons before answering can conceal the shift almost entirely in its answers, even though a direct reading of its internal state shows that its disposition has moved. With reasoning switched off, the same model shifts steadily as the share of harmful advice in its training grows. These results demonstrate the importance of testing safety with reasoning turned both on and off.

[39] Truncated Reasoning Under a Budget: Evaluating System-1 Models as Decision-makers cs.CLPDF

Zhongyin Wang, Zhaoqian Xue, Junting Mu, Haozheng Liu, Ye Yuan

TL;DR: 该研究评估System-1决策模型读取被截断的长推理轨迹能否代替强迫推理模型立即作答，并发现其在多个token预算下更准确且总成本更低。

Details

Motivation: 大语言模型常生成超出需要的推理token，限制预算可降低成本，但若要求模型依据不完整推理立即作答，准确率可能下降。System-1决策模型能直接在候选项中选择而不生成推理，可能用于读取部分轨迹。

Result: 作者使用Jev作为被截断推理的决策读取器，评估七个推理LLM与五个多选基准的35组组合，并将推理和回答成本一并计入。在500至16,000 token预算范围内，Jev平均准确率高于强迫推理模型基于截断轨迹作答。1,500 token预算下总体准确率提高4.1个百分点，截断题提高8.6点；经Holm校正的bootstrap检验中17组对有显著收益。离线回放估计成本为完整运行的34.5%，准确率损失1.13点；根据训练折选择策略，在留出折节省估计成本71.3%，准确率损失0.58点。Jev仅看问题的准确率也较强，平均68.0%，高于500 token强制回答的65.5%，但各基准不一。

Insight: 把截断推理交给专门选择器可避免强迫生成模型在不完整轨迹上继续猜测，并可能降低总推理成本。成本估计来自离线回放或策略折分，摘要未说明系统部署延迟及不同模型读取器泛化。

Abstract: Large language models (LLMs) often generate more reasoning tokens than needed. Limiting the reasoning budget reduces cost, but forcing an answer from an incomplete trace can hurt accuracy. System-1 decision models select among candidate options without producing a reasoning response. We evaluate Jev as a decision reader of truncated reasoning across seven LLM reasoners and five multiple-choice benchmarks (35 pairs). Jev achieves higher mean accuracy than forced answering at tested budgets from 500 to 16,000 tokens, including the estimated costs of reasoning and answering. At a 1,500-token budget, it improves accuracy by 4.1 pp overall and 8.6 pp on truncated questions, with significant gains in 17 pairs under Holm-corrected bootstrap tests. Offline replay estimates 34.5% of the full-run cost, with a 1.13 pp accuracy loss. A policy selected on training folds saves 71.3% of the estimated full-run cost on held-out folds, with a 0.58 pp accuracy loss. We also find that Jev shows strong question-only accuracy, exceeding 500-token forced answering on average (68.0% vs. 65.5%), although this advantage varies by benchmark. These findings suggest that System-1 decision models can complement LLM reasoning and reduce the need for costly reasoning completion.

[40] Diffusion Language Models Change Speed Across Tasks and Languages cs.CL | cs.AI | cs.LGPDF

Mohsen Fayyaz, Ali Modarressi, Daniel Israel, Lucas Bandarkar, Nanyun Peng

TL;DR: 该研究发现扩散语言模型每生成token所需前向次数随任务、语言和答案是否重复输入而变化，并可将解码速度用作候选答案质量信号。

Details

Motivation: 扩散语言模型能在一次前向生成多个token，并自主决定生成时需要多少次迭代，因此相同长度回答的计算成本可能差异很大。

Result: 作者分析两个扩散语言模型在十项任务上的速度因素。每token前向次数强烈依赖任务：复制或轻微编辑输入内容较便宜，开放式写作较昂贵；DiffusionGemma写故事每token需超过语法纠错的三倍前向。输出中相对输入新增的信息量及词语在任务内的稀有程度能解释很大部分差距。含推理回答也有类似现象：重复推理过程的最终答案便宜，增加新信息时更贵。高分词元化语言每token所需前向次数也更多。相同数学问题上，错误答案通常比正确答案每token需要更多前向；按解码速度给候选加权可比多数投票最多提高4.2分，且不增加成本。

Insight: 扩散解码计算量取决于输出中已有多少内容可由输入或推理状态确定，速度本身可能补充候选质量判断。摘要未给出十项任务名单、模型间一致性和速度加权准确率分项，4.2分为最佳结果。

Abstract: Diffusion language models can write several tokens in one forward pass, and they decide during generation how many passes to spend. Two outputs of the same token length can therefore vary a lot in cost. To understand the underlying factors of this cost variation, we analyze two diffusion models across 10 tasks. We find that the number of forward passes per token depends heavily on the task. Text copied or lightly edited from the input is cheap, while open-ended writing is costly. In DiffusionGemma, generating a token for story writing requires more than three times as many forward passes as for grammar correction. Two output properties explain much of this gap: how much of it is new relative to the input, and how rare its words are within the task. The same holds within a single response when thinking is involved. The final answer is cheap when it repeats the model’s reasoning trace and costly when it adds something new. In multilingual generation, we find that highly tokenized languages also require more passes per token, adding to the cost they already pay with more tokens. Finally, cost tracks correctness: on the same math problem, wrong answers take more passes per token than correct ones, and weighting candidate answers by decoding speed beats majority voting by up to 4.2 points at no extra cost. Taken together, our results show that diffusion’s computational advantage depends on how much of the output is already determined, while its decoding effort offers a practical signal for making inference more efficient and reliable.

[41] Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining cs.CL | cs.AIPDF

Xinyi Wang, Danqi Chen

TL;DR: 该研究根据RL推理失败时出现的特征n-gram，从预训练语料检索针对性文本继续预训练，再进行RL，发现小规模定向数据可弥补强化学习未覆盖的能力。

Details

Motivation: 可验证奖励强化学习常能强化基础模型已有推理能力，但失败时模型缺失什么尚不清楚；盲目扩大继续预训练数据成本高。

Result: 作者分析RL策略自身轨迹，发现失败与少量特征n-gram高度相关，从“let’s recheck”等反思语句到特定数学操作；这些失败在RL训练中持续存在，但模型对同题的其他采样回答有时正确。研究用失败n-gram从预训练语料检索文本，在RL前对基础模型继续预训练。语料仅3,500万token，比常见中期训练小3–4个数量级。作者在OLMo2和Qwen3的GSM8K、数学/常识/知识混合及一般数学科学推理设置评估，11种设置中10种优于从预训练语料等量随机抽文本继续预训练；在中期训练前加入时GSM8K提高2.0分。为OLMo2-1B失败检索的文本也使Qwen3-4B在混合任务上比同基线提高1.9分。

Insight: RL失败轨迹可作为低成本信号，定位需要补充的基础知识或推理模式；定向检索数据比等量随机继续预训练更有效。摘要未给出11种设置中的单项结果和失败n-gram检索精度，迁移性限于所测模型及任务。

Abstract: Reinforcement learning with verifiable rewards (RLVR) is widely used to improve reasoning in language models, yet growing evidence suggests that it mostly sharpens what the base model can already do. This raises a practical question: when RL fails, what is the base model missing? We look inside the RL policy’s own rollouts and find that failures are localized: whether a rollout fails is strongly predicted by a small set of characteristic n-grams, ranging from reflection phrases such as “let’s recheck” to specific operations such as “the inverse of A”. These failures persist throughout RL training, although they are not beyond the model’s reach: other sampled answers to the same question are often correct. Building on this observation, we use the failure $n$-grams to retrieve text from the pretraining corpus and continue pretraining the base model on it before RL. The resulting corpus has only 35M tokens, 3–4 orders of magnitude smaller than typical mid-training. We evaluate this approach on OLMo2 and Qwen3 models trained with RL on (1) GSM8K; (2) a mixture of math, commonsense, and knowledge questions; and (3) general math and science reasoning. Despite its size, the corpus yields consistent gains: our approach outperforms continued pretraining on the same amount of text sampled from the pretraining corpus across 10 out of 11 settings, including by $2.0$ points on GSM8K when inserted before mid-training. The retrieved text also transfers across model families: text retrieved for OLMo2-1B’s failures improves Qwen3-4B on the mixture by $1.9$ points over the same baseline. Together, our results suggest that where RL fails is a cheap and effective signal for selecting pretraining data.

[42] Measurement Validity and Representation Dynamics of Confidence Signals in Large Language Models cs.CL | cs.AI | cs.LGPDF

Debojeet Ghosh, Shreyansh Kumar, Ayushman Garg

TL;DR: 该研究比较LLM置信度、隐藏状态线性探针和采样语义一致性三类“知道自己知道什么”信号，发现它们测量性质不同且各有失效。

Details

Motivation: 模型能否判断自身知识边界关系到可信部署及是否应弃答。既有工作常只用口头置信度、隐藏状态探针或采样答案一致性中的一种，并默认这些信号可互换。

Result: 作者预注册审计五个Qwen2.5模型（0.5B–7B，含7B基础/指令模型），覆盖事实检索PopQA、SimpleQA及多步推理GSM8K、MATH。用预填充状态预测采样答案是否分歧的探针，在检索任务约7%网络深度即可具有信息量，推理任务约需28%；但预注册五层测试只边缘支持该差异，事后设计的21层分析支持更明显。多数检索设置中，仅使用提示的TF-IDF分类器就能匹配探针，早期信号可能反映问题表面特征。口头置信度对轻微变化的提示模板很不稳定。期望校准误差可能误导：口头置信度聚集在基准率附近，表面校准却区分度差；采样一致性ECE更差，但分辨率约为两倍。五个预注册假设中两项被否定、一项部分支持、一项不支持、一项因基础模型口头置信度退化无法检验。

Insight: 置信信号的校准误差和区分能力是不同属性，跨任务、模型层及提示模板都需验证；事后21层分析应与预注册结果区分。摘要的预注册假设并未普遍获得支持，结论仅限五个Qwen2.5模型和四个基准。

Abstract: Whether a language model “knows what it knows” matters for trustworthy deployment and for deciding when it should abstain. Prior work usually measures this through one channel: verbalized confidence, a linear probe on hidden states, or agreement among sampled answers (semantic entropy), and often treats these as interchangeable. We compare all three in a pre-registered audit of five Qwen2.5 models (0.5B-7B, including a base/instruct pair at 7B) on factual retrieval (PopQA, SimpleQA) and multi-step reasoning (GSM8K, MATH). First, probes trained on prefill states to predict whether sampled answers will disagree become informative earlier for retrieval (about 7% of layer depth on average) than for reasoning (about 28%). The pre-registered five-layer test supported this only marginally; a finer 21-layer follow-up, designed after those results, supports it more clearly. In most retrieval cells, however, a prompt-only TF-IDF classifier matches the probe, so the early signal may reflect surface features of the question. Second, verbalized confidence is unstable: slightly different prompt templates yield scores that barely correlate. Third, Expected Calibration Error is misleading: verbal confidence looks well calibrated because it clusters near the base rate while discriminating poorly, whereas sampling agreement has worse ECE but about twice the resolution. Of five pre-registered hypotheses, two were falsified, one was partly supported, one was not supported, and one could not be tested because the base model’s verbal confidence was degenerate.

[43] Learning to Reason Efficiently with Interventional Step Credit cs.CL | cs.LGPDF

Guanda Li, Xuefeng Bai, Yulong Chen, Zhiyong Wang, Kehai Chen

TL;DR: 干预式步骤信用ISC通过定向参数干预衡量推理步骤对最终答案的贡献，IGPO据此调整强化学习优势以减少冗余思维链。

Details

Motivation: 大型推理模型常生成长思维链，其中混有有用步骤和多余内容。注意力质量或步骤熵能反映模型在某步的行为，却不直接衡量该步骤对答案的因果贡献。

Result: 作者提出ISC，从因果干预视角对模型参数进行定向干预以评估步骤重要性；基于ISC设计IGPO，在强化学习中细粒度重缩放优势，抑制冗余步骤同时保留重要步骤。在两个模型规模和九个基准上的实验显示，IGPO相较被比较方法最高可平均减少62%生成token，并取得更好的总体表现。

Insight: 因果贡献信号有望比注意力或熵等行为代理更直接地估计步骤价值。摘要未给出不同基准准确率差异、干预计算成本和62%压缩对应的任务质量变化。

Abstract: Large reasoning models often produce lengthy chain-of-thought (CoT) traces mixing useful intermediate steps with unnecessary ones. Recent methods explore step-level signals such as attention mass and step entropy to identify and suppress unnecessary steps. Despite their effectiveness, such signals reflect the model’s behavior at a step rather than that step’s contribution to the answer. To address this, we propose the interventional step credit (ISC), which takes a causal-intervention view and assesses step importance to the reasoning through targeted parameter-level interventions. Building on ISC, we further propose IGPO, which uses ISC to guide fine-grained advantage rescaling in reinforcement learning, reducing redundant steps while preserving the important ones. Experiments across two model scales and nine benchmarks show that IGPO achieves up to 62% average token reduction and better overall performance among compared methods.

[44] CALM: Consistency-Aware Logit Masking in Decoding Space for Mitigating Overthinking of Large Reasoning Models cs.CLPDF

Shengjing Liu, Feiyang Li, Yuxuan Yang, Yile Wang, Hui Huang

TL;DR: CALM在推理模型回答步骤边界检测试答一致性，并屏蔽“Wait”等过渡token以减少重复反思，同时允许后续思考继续。

Details

Motivation: 大型推理模型在数学和科学任务上通过长思维链获得强表现，但常过度思考。依据连续试答一致性强制提前退出的截断方法，局部一致并不一定代表最终答案稳定，还可能打断有效验证。

Result: CALM是无需训练的解码干预：在步骤边界监控试答，再对“Wait”“Alternatively”“Hmm”等动作转折点token进行logit屏蔽，抑制重复过渡推理而不直接强制停止。四个先进推理模型和五个基准实验显示，相较完整思维链，token消耗减少18.6%–25.0%，性能保持；由于冗余减少，准确率还有0.07%–1.48%的提升。LLM评审认为输出也更连贯简洁。

Insight: 直接抑制冗余转折语句可比依据局部答案一致而硬停止更温和。摘要未给出具体模型、基准分项成绩和“准确率提升”口径，也未报告过渡token屏蔽可能影响有效反思的场景。

Abstract: Large reasoning models (LRMs) have achieved strong performance on challenging mathematical and scientific tasks through long chain-of-thought reasoning, but they often suffer from overthinking with unnecessary reasoning steps. Existing truncation-based methods typically detect consecutive trial-answer agreement and force early exit, yet such local stopping signals do not necessarily indicate final-answer stability and can disrupt useful subsequent reflection or verification, potentially leading to performance degradation. We propose consistency-aware logit masking (CALM), a training-free decoding intervention for mitigating overthinking in LRMs. CALM first monitors trial answers at step boundaries and then applies logit masking to action transition point (ATP) tokens such as “Wait”, “Alternatively”, and “Hmm”, suppressing redundant transitional reasoning while allowing the model to continue its thought process more naturally. Experiments across four advanced LRMs and five benchmarks show that CALM reduces token consumption by 18.6%-25.0% compared with full CoT reasoning while maintaining performance, and can even improve accuracy by 0.07%-1.48% by reducing redundant reasoning. LLM-as-a-Judge results also demonstrate that CALM provides more coherent and concise results.

[45] TempoMoE: Identifying and Steering Fast-Slow Thinking Experts for Efficient Multimodal Reasoning cs.CL | cs.CV | cs.LGPDF

Xiyue Zhang, Keru Zhou, Ziyi Wang, Bing Hu, Li Wang

TL;DR: TempoMoE通过对比不同推理节奏下的专家路由，识别快思考与慢思考专家，并按样本难度动态调整路由以减少简单问题的过度推理。

Details

Motivation: 多模态MoE推理模型处理复杂任务时依赖长思维链，但简单输入也可能过度推理，增加token与延迟，甚至降低准确率。训练时缩短推理或外部限制输出，未解释底层专家路由机制。

Result: 研究发现固定输入、只改变提示的思考节奏会产生稳定路由变化；抑制慢思考专家会缩短推理，抑制快思考专家会延长推理，提供专家影响推理深度的因果证据。免训练TempoMoE分两阶段通过节奏对比路由分析识别专家，并在原top-k集合内重分配路由权重：简单样本强化快思考专家、抑制慢思考专家，困难样本反向调整。在Kimi-VL、Qwen3-VL和ERNIE-4.5-VL的六个基准上，相较原模型准确率最高提高2.6分，输出token最多减少35.4%，整体持续提高准确率并减少token。

Insight: MoE专家可能承担不同推理深度功能，可按难度调节其权重，无需训练新模型。摘要未给出各模型逐项结果、路由计算开销及难度估计误差，效果限于三种模型和六个基准。

Abstract: Multimodal Mixture-of-Experts (MoE) reasoning models rely on long chain-of-thought (CoT) for complex tasks, yet often overthink easy inputs, increasing token cost and latency, and sometimes degrading accuracy. Existing methods optimize reasoning length during training or impose external control at inference, but leave the underlying reasoning mechanism a black box. We find that fast and slow thinking exhibit distinct expert-routing patterns: holding the input fixed and varying only the prompted reasoning tempo produces stable routing shifts. Suppressing slow-thinking experts shortens the reasoning chain, whereas suppressing fast-thinking experts lengthens it, providing causal evidence that these experts influence reasoning depth. Based on this finding, we propose TempoMoE, a training-free, two-stage framework. It identifies fast- and slow-thinking experts via tempo-contrastive routing analysis and redistributes routing weights within the original top-k set according to sample difficulty. It amplifies fast-thinking experts and suppresses slow-thinking experts on easy inputs and reverses this adjustment on hard ones, enabling difficulty-adaptive control over reasoning depth. We evaluate TempoMoE on Kimi-VL, Qwen3-VL, and ERNIE-4.5-VL across six benchmarks. Compared with Vanilla, TempoMoE consistently improves accuracy while reducing output tokens, with accuracy gains of up to 2.6 points and token reductions of up to 35.4%.

[46] Scaling Synthetic Pre-training Data and Evaluation for Indic LLMs cs.CLPDF

Abhinav Khattar, Aishwarya Chhabra, Ameya Velingker, Anusha Kamath, Brandon Norick

TL;DR: 该研究发布覆盖印度22种法定语言及12种罗马化形式的2.83万亿token合成预训练语料，并提供多语言评测套件用于Indic语言模型开发。

Details

Motivation: 面向印度的语言模型需要可扩展多语言数据和覆盖多能力的可靠评估，但Indic语言在预训练语料和现有基准中常覆盖不足。

Result: Nemotron-Pretraining-Indic-v1包含约2.83万亿token，覆盖22种印度法定语言的本土文字及12种罗马化形式，作者称其为当前最大公开Indic预训练语料。GPU加速流水线将精选英文文档翻译成目标语言，评估翻译质量并修复低质量输出，同时保留技术内容、文档结构和来源追踪。12种语言的持续预训练实验显示，合成语料与自然Indic文本结合后，知识和推理能力超过公开语料基线；研究还分析数据量、文字脚本覆盖和Indic训练占比对英语能力的影响。作者另发布Indic-LLM-Eval-Suite，涵盖聊天、指令跟随、推理、知识、代码和翻译任务，并通过自动质量评估、定向修正及母语者审查构建跨语言基准。

Insight: 规模化翻译语料与质量控制评测可为Indic语言模型提供数据基础，但合成翻译质量及不同文字体系覆盖仍需验证。摘要未给出持续预训练量化增益和语料许可细节，合成数据与自然文本的作用也可能因语言而异。

Abstract: Building capable language models for India requires scalable multilingual data and robust evaluation. We introduce Nemotron-Pretraining-Indic-v1, a synthetic pretraining corpus of approximately 2.83 trillion tokens covering all 22 scheduled Indian languages in native script and twelve in romanized form. It is the largest openly released Indic pretraining corpus. Using our GPU-accelerated pipeline, we translate curated English documents, assess translation quality, and repair low-quality outputs while preserving technical content and document structure and maintaining source traceability. Through continued-pretraining experiments across twelve languages, we show that combining our synthetic corpus with natural Indic text improves knowledge and reasoning over a public-corpus baseline. Additional experiments show how data volume, script coverage, and Indic training share affect English performance. We also introduce the Indic-LLM-Eval-Suite, a multilingual benchmark suite for evaluating Indic language models across chat, instruction following, reasoning, knowledge, code, and translation. The suite combines existing Indic benchmarks with task-aware translations of English benchmarks, using automated quality assessment, targeted correction, and native-speaker review in a human-in-the-loop pipeline. We use the suite to comprehensively benchmark contemporary open-weight Indic and multilingual models across capabilities and languages, revealing substantial variation across tasks and languages. Together, these resources provide a scalable foundation for developing and evaluating more capable Indic language models.

[47] Balance-Aware Pivot Selection for Parallel Diffusion LLM Decoding cs.CLPDF

Yejin Lee, Hyundong Jin, Yo-Sub Han

TL;DR: BPD选择连续掩码区间的中心位置并行解码，使扩散语言模型的解码深度随区间长度呈对数增长。

Details

Motivation: 扩散语言模型可并行生成token，但现有加速常依据预测置信度优先提交可靠token，没有明确考虑位置选择如何影响后续解码日程；token依赖顺序也未必只由置信度决定。

Result: Balance-Aware Pivot Decoding偏好在连续掩码区间选取中心位置并行解码，理论上可令解码深度相对区间长度呈对数增长。实用版本结合置信度预填充和置信阈值过滤的中心选择，以兼顾模型不确定性。在LLaDA和Dream上五个数学推理、代码生成和指令跟随基准评估质量—效率折衷。

Insight: 从位置结构而非仅置信度决定解码顺序，可减少最长未解区间的深度。摘要未给出性能数值、延迟或与其他平衡策略的比较，链接为匿名代码库。

Abstract: Diffusion language models enable parallel token generation, offering a promising alternative to sequential autoregressive decoding. Existing acceleration methods use predictive confidence to enable parallel token commitment while preserving generation quality. However, confidence-based selection prioritizes reliable token predictions without explicitly considering how token positions shape the remaining decoding schedule. Moreover, such ordering does not always reflect inter-token dependencies, suggesting that confidence need not be the sole criterion for token selection. Motivated by this limitation, we propose Balance-Aware Pivot Decoding (BPD), which favors central pivots within contiguous masked intervals and decodes the selected positions in parallel. Balanced pivots yield logarithmic decoding depth in the interval length. Our practical method accounts for model uncertainty by combining confidence-based prefilling with confidence-thresholded center selection. We evaluate the resulting quality-efficiency tradeoff on LLaDA and Dream across five mathematical reasoning, code generation, and instruction-following benchmarks. Our implementation is available at: https://anonymous.4open.science/r/DLM_random_pivot-EE62/

[48] Honeyed Words, Hidden Daggers: Benchmarking and Training LLMs to See Through Mock Politeness cs.CL | cs.AIPDF

Caixin Kang, Sitong Gong, Tianyu Yan, Kaipeng Zhang, Yoichi Sato

TL;DR: ZhenhuanBench评测LLM能否识别表面礼貌、实际敌意的言语，涵盖说话者意图、言语与情绪不一致及不同听者反应。

Details

Motivation: “口蜜腹剑”式表达表面客气、实际敌对，理解需要结合语气、说话者意图及听者反应，单纯文字情感分类难以捕捉。

Result: ZhenhuanBench包含1,592个项目，依据电视剧《甄嬛传》专家评论构建，覆盖说话者意图、情绪与语言不一致及听者反应。34个LLM在纯文本条件评估，GPT-6 Astra总分最高为73.48%，但听者反应分仅58.37%。标注还提供监督轨迹和字符串可验证训练奖励。使用338条金标准轨迹时，8B模型在相同项目上超过教师蒸馏；后续强化学习趋于饱和，机会校正使估计可恢复性能差距缩小7.5倍。CPR-GRPO循环用于按估计分配金标准监督，在公开多选任务得分67.0，在训练匹配隐藏协议得分74.5。无需目标任务训练时，CPR在六个中英文语用和讽刺评测中超过Base、SFT和GRPO；三个外部数据集上的平均macro-F1也高于SFT和GRPO。

Insight: 礼貌形式与敌意功能的区分需要结合语境和听者效应；单看总体成绩可能掩盖听者反应理解不足。基准依赖特定电视剧专家评论，跨文化及真实会话泛化需另行验证。

Abstract: The Chinese idiom “kou mi fu jian” (“honeyed words, hidden daggers”) names speech that is polite in form but hostile in function. We present ZhenhuanBench, a 1,592-item benchmark built from expert commentary on Empresses in the Palace, covering speaker intent, emotion-language mismatch, and per-listener uptake. Text-only evaluation covers 34 LLMs. GPT-6 Astra leads the leaderboard at 73.48% overall, while its listener-reaction score is 58.37%. The annotations also supply supervised traces and string-verifiable training rewards. With 338 gold traces, an 8B model outperforms teacher distillation on identical items. Subsequent RL saturates, and chance correction reduces the estimated recoverable performance gap by a factor of 7.5. Our CPR-GRPO loop (Compress-Probe-Replenish) uses this estimate to weight gold supervision, reaching 67.0 on public multiple-choice tasks and 74.5 under the training-matched hidden protocol. Without target-task training, CPR outperforms Base, SFT, and GRPO across six Chinese and English pragmatic and sarcasm evaluations. On three external datasets, training with CPR yields higher mean Macro-F1 than SFT and GRPO.

[49] DeepSEA: A Multilingual Evaluation Benchmark for Southeast Asia cs.CLPDF

Jiuzhou Han, Abhishek Bijay Mishra, Trevor Cohn, Shubham Mittal

TL;DR: DeepSEA为东南亚12种语言构建涵盖数学与科学推理、问答、摘要和翻译的多语言基准，揭示低资源语言模型表现下降及对英语中介的依赖。

Details

Motivation: 语言模型在高资源语言上能力强，但东南亚约7亿人口使用的多种语言在基准中代表不足，现有评测常语言覆盖有限或任务范围单一。

Result: DeepSEA将既有数据集由专业译者翻译到12种东南亚语言，涵盖印尼语、泰语、越南语等较高资源语言及缅甸语、高棉语、老挝语等低资源语言，任务包括数学推理、科学推理、问答、抽象摘要和机器翻译。对前沿闭源和开放权重模型的评估显示，Gemini-3.7-Flash等前沿模型跨资源层级相对稳健，小模型在低资源语言上则大幅退化。模型通过英语中介而非直接使用目标语言时，推理和摘要表现更好，提示其依赖英语表征；研究还分析跨语言错误和跨脚本tokenizer词元化负担。

Insight: 多语言模型的平均成绩会掩盖语言资源差异及英语中介依赖，东南亚多任务评测有助于定位这些缺口。摘要未给出语言逐项分数、人工翻译质量和具体模型列表，结论限于所测12种语言及基准。

Abstract: While language models have demonstrated remarkable capabilities, their performance remains disproportionately skewed towards high-resource languages. This disparity is particularly acute for the 700 million inhabitants of Southeast Asia, where vast linguistic diversity remains under-represented in existing benchmarks, often suffering from limited language coverage or narrow task scopes. To address this, we introduce DeepSEA, a comprehensive benchmark constructed from established datasets through professional human translation into 12 Southeast Asian languages, spanning both high-resource (e.g., Indonesian, Thai, Vietnamese) and low-resource (e.g., Burmese, Khmer, Lao) tiers across five distinct tasks: mathematical reasoning, scientific reasoning, question answering, abstractive summarization, and machine translation. Extensive evaluation of frontier proprietary and open-weights models reveals that while frontier models like Gemini-3.7-Flash maintain robustness across resource tiers, smaller models face severe performance degradation on low-resource languages. Crucially, models perform better in reasoning and summarization when pivoted through English rather than target languages, exposing a critical dependency on English representations, especially for low-resource languages. Beyond these observations, we provide a cross-lingual error analysis and a cross-script tokenizer fertility analysis. We hope these findings shed light on the development of next-generation inclusive multilingual models.

[50] Conditional information Retrieval for Longitudinal Survey Harmonisation cs.CL | cs.CY | cs.IRPDF

Wing Yan Li, Zeqiang Wang, Michelle Lin, Jon Johnson, Suparna De

TL;DR: 该框架先提取社会科学问卷问题的潜在概念，再将概念作为条件信号进行跨问卷检索，提升纵向调查问题的概念对齐。

Details

Motivation: 不同调查中的问题概念等价性检索难以仅靠表面文本相似：措辞相近的问题可能测量不同社会学子概念，措辞差异大的问题却可能等价。

Result: 框架分两阶段执行：微调生成模型从调查问题中提取底层社会学概念，再将提取概念作为显式条件加入问题表示，进行概念感知检索。作者在标注数据稀缺的社会科学环境下以CLOSER为案例。相较基线，平均MAP和Recall均稳定提高2–3分。研究还比较SFT、DPO和动态DPO，发现SFT增益最稳定；动态DPO与指令微调及蒸馏式推理骨干的交互不同。用一般社会科学词表ELSST微调的模型可稳健迁移到具体调查数据。

Insight: 概念抽取可补足调查问卷表面语义差异，为稀缺标注下的跨问卷匹配提供条件检索信号。摘要未列出问卷数、MAP/Recall的具体基线值及不同概念类型表现。

Abstract: Identifying conceptually equivalent questions across different surveys is a challenging information retrieval task where surface-level textual similarity is insufficient. Two questions may be lexically similar yet assess entirely different sociological sub-concepts, or lexically different yet conceptually equivalent. We introduce a conditional information retrieval framework that addresses this by decomposing the task into two steps: concept extraction, which identifies the underlying sociological concept of a survey question using a fine-tuned generative model, and concept-aware retrieval, which augments question representations with the extracted concept as an explicit conditioning signal for finer-grained alignment. To reflect the realistic data scarcity conditions of specialised social science annotation, we use CLOSER as a case study. Our framework consistently outperforms baselines by 2-3 points in both MAP and Recall. We further analyse fine-tuning strategies including SFT, DPO, and dynamic DPO, finding that SFT delivers the most stable gains while dynamic DPO interacts differently with instruction-tuned versus distilled reasoning backbones. Finally, we demonstrate that a model fine-tuned on a general social science thesaurus (ELSST) transfers robustly to task-specific survey data, offering a practical solution under data scarcity.

[51] Heartbeat Thinking: Speculative Reasoning and Tool Execution While Listening in Realtime Voice Agents cs.CLPDF

Zhenyu Zhang, Kewei Fu, Yongqi Tong, Chenzhi Zheng, Wei Wu

TL;DR: Heartbeat Thinking让语音智能体在用户仍说话时周期性处理最新转录，同时允许先前思考和工具调用完成，利用聆听时间准备后续响应。

Details

Motivation: 语音工具智能体通常要等用户说完才开始工作；尽管可以边听边推测，但按语音块处理或更新即中断生成，可能在工具参数尚未完整时截断调用，使计算与语音更新紧密绑定。

Result: Heartbeat Thinking为更新的输入启动新的Thinker实例，同时让已有实例及已派发工具继续运行。周期性实例读取最新部分转录及共享的已完成推理和工具结果记录，用于补全缺失参数并发起依赖调用，让准备过程跨多个语音更新持续推进。在SpeechFCEval、FDBench v3和AudioCRAG上，与SHANKS式窗口基线相比，该方法提高总体推测式工具覆盖或检索复用，最终任务质量则因基准而异。FDBench v3上工具覆盖提高16个百分点，最终回复平均延迟降低36.4%。

Insight: 不取消正在进行的计算，可将用户说话时间转换为可复用的工具准备和推理进度。不同基准的最终质量影响并不一致；摘要未给出新增计算、打断策略和实时音频开销。

Abstract: Tool-using voice agents often leave users waiting after speech ends, even when actionable information arrives earlier in the request. Speculative approaches begin work during speech, but chunk-bounded or interrupt-on-update generation can truncate a tool call before its arguments are complete. This couples computation time to speech updates, limiting useful work completed in advance. We introduce Heartbeat Thinking, which handles updated input through new Thinker invocations while allowing ongoing invocations and dispatched tools to finish. Periodic invocations read the latest partial transcript and a shared record of completed reasoning and tool results. These results help resolve missing arguments and initiate dependent calls, allowing preparation to progress across speech updates. Across SpeechFCEval, FDBench v3, and AudioCRAG, Heartbeat improves aggregate speculative tool coverage or retrieval reuse over a SHANKS-style windowed baseline, with benchmark-dependent effects on final task quality. On FDBench v3, it increases tool coverage by 16 percentage points and reduces mean final-response latency by 36.4%. The results show that heartbeat thinking paradigm preserving ongoing computation during speech can turn listening time into reusable preparation for faster and more capable voice-agent responses.

[52] Bridged Self-Distillation for On-Policy Learning cs.CLPDF

Runsheng Chen, Yihong Liu, Raoyuan Zhao, Michael A. Hedderich, Hinrich Schütze

TL;DR: BSD用外部教师思维轨迹作为特权上下文，再让学生自身的副本在学生token空间内评分同一批rollout，以统一处理同分词器、跨分词器及闭源教师蒸馏。

Details

Motivation: 在线策略蒸馏可把强教师推理能力转移给小模型，但教师解法可能过难以模仿，教师token级概率也可能与学生不兼容或无法获得；既有方法常针对这些场景分开设计监督。

Result: Bridged Self-Distillation先让外部教师提供推理轨迹，可离线生成；学生不见该上下文进行rollout，而同一学生模型的拷贝在看到特权轨迹后作为自教师，在学生自身token空间中为同一rollout评分。BSD将token级指导用于调节失败轨迹的更新，结果验证器提供终局信号。外部教师只需输出文本，因此同一流程适用于共享分词器、跨分词器及闭源API教师。实验显示在这些教师场景中，BSD均提升基础模型表现并超过各场景专用基线。

Insight: 通过学生自己的词表和模型来解释外部教师轨迹，可避免依赖教师token概率并统一异构教师蒸馏。摘要未提供任务、具体分数和蒸馏成本，API闭源教师场景的文本生成成本仍需考虑。

Abstract: On-policy distillation can transfer the reasoning capabilities of a strong teacher to a smaller student. However, distillation from a stronger teacher model faces two mismatches: its solutions may be difficult for a smaller student to imitate, and its token-level probabilities may be incompatible with, or even unavailable to, the student. Existing methods handle these settings through different forms of supervision. To this end, we propose Bridged Self-Distillation (BSD), a unified framework for distilling from any teacher that provides reasoning traces. The model to be distilled from, termed the external teacher, first provides a reasoning trace as privileged context, which can be generated in an offline fashion. The student generates rollouts without this context, while a copy of the student conditioned on it, serving as the self-teacher, scores the same rollouts in the student’s own token space. BSD uses the resulting token-level guidance to adjust updates on unsuccessful rollouts, with a verifier providing the outcome signal. As the external teacher is queried only to produce text, the same training procedure applies to all teacher scenarios: shared-tokenizer, cross-tokenizer, and closed API teachers. Across these scenarios, BSD improves over the base model and outperforms baselines tailored to each scenario, offering a novel and unified pathway for distillation. Code is available: https://github.com/TUMLegalTech/Bridged-Self-Distillation

Eduin E. Hernandez, Sergio A. Diaz, Luis F. Garcia, Nurassyl Askar, Stefano Rini

TL;DR: 对 PathQuestion 和 PathQuestion-Large 的受控审计表明，把任务过滤图替换为基准所附的更大源知识图会显著改变导航难度、答案可达性和测得性能，因此图构造本身属于多跳知识图导航任务定义的一部分。

Details

Motivation: 基于导航的知识图问答成绩不仅取决于问题与推理方法，也可能受承载任务的图结构影响。论文希望在固定问题、主题实体、答案、路径标注和数据划分的条件下，单独测量图构造方式对 PathQuestion 两个基准的影响。

Result: 恢复更大的源知识图扩大了有效搜索空间以及图上一致的答案和路径数量，却明显降低了不受引导时到达已发布答案的可能性。结构参照、冻结的小语言模型导航策略和训练后的 MINERVA 策略均显示：图恢复会降低已发布答案准确率和轨迹保真度；采用图一致的评估只能挽回部分损失。

Insight: 基准发布的过滤图并非无关紧要的实现选择，而是导航任务规格的一部分，能显著影响测得行为。比较导航方法或解释基准成绩时，需要报告图的构造和评估所用图；摘要未给出具体下降幅度。

Abstract: Multi-hop knowledge-graph question answering (KGQA) benchmarks are often treated as fixed language-and-reasoning tasks, but for navigation-based methods, difficulty also depends on the graph in which those tasks are instantiated. This makes graph construction a consequential but often entangled part of the navigation task. To isolate this effect, we audit PathQuestion (PQ) and PathQuestion-Large (PQL) as graph-navigation benchmarks under a controlled intervention: we hold questions, topic entities, released answers, released path annotations, and family-level splits fixed while replacing each released task-filtered graph with the larger source KG distributed with the benchmark. Source restoration expands the effective navigation search space and graph-consistent answer and path multiplicity while sharply reducing unguided reachability to released answers. We evaluate whether these structural changes persist across navigation policies using structural references, frozen small-language-model navigation policies, and trained MINERVA policies. Across both frozen and trained policies, restoration substantially reduces released-answer accuracy and trajectory fidelity, while graph-consistent evaluation recovers only part of the loss. These results show that benchmark graph construction is part of the navigation task specification rather than merely an implementation detail and can materially shape measured behavior even when the released QA task is held fixed.

[54] Co-Evolving Policy and Reward for Code Generation: A Mechanistic Study of Single-Adapter Drift cs.CL | cs.LG | cs.SEPDF

Amarsaikhan Batjargal, Noriko Tomuro

TL;DR: 在 Qwen2.5-Coder-3B-Instruct 上，SPARK 式策略与奖励协同演化虽提高训练表现，却未带来稳定的留出题泛化；作者将其归纳为共享适配器参数漂移，并指出训练奖励上升不足以证明代码自我改进有效。

Details

Motivation: 可验证奖励强化学习可通过执行结果客观评估代码，并将执行反馈再用于监督。论文考察面向推理任务提出的 SPARK（复用训练轨迹、共同训练解题与评估改进能力）能否迁移到代码生成，以及执行反思能否进一步带来自我提升。

Result: 作者将 SPARK 式轨迹复用与仅使用执行反馈的 GRPO 对比。实验不支持协同演化或执行反思能改善泛化的假设：协同演化提高训练表现，但收益没有泛化到留出问题；同一训练阶段的轨迹复用使相对参考策略的 KL 散度约为执行型 GRPO 的 34 倍。训练期间，GRPO 每条序列的训练 token 数从 185 降至 86，训练通过率上升时留出表现却下降或无法维持早期收益。

Insight: 共享可训练参数可能逐渐偏离预训练代码生成行为，却没有换来泛化提升，作者将此称为单适配器策略漂移。因此评估代码自我改进不能只看训练奖励，还需检查留出集表现和策略漂移；结论限于所测配置，摘要未提供其他模型或超参数下的结果。

Abstract: Reinforcement learning with verifiable rewards (RLVR) is particularly well suited to code generation because generated programs can be objectively evaluated through execution. Execution also provides behavioral information that can be reused as additional supervision. The Synergistic Policy and Reward Co-Evolving Framework (SPARK), originally proposed for reasoning tasks, exploits this idea by recycling training rollouts and their evaluation results to train a model not only to generate solutions but also to evaluate and improve them. We investigate whether this policy-reward co-evolution approach transfers effectively to code generation and whether execution-grounded reflection provides an additional source of self-improvement. We compare SPARK-style rollout recycling with execution-only Group Relative Policy Optimization (GRPO) using Qwen2.5-Coder-3B-Instruct. Our results do not support either hypothesis under the configuration tested. Although co-evolution improves training performance, these gains do not generalize to held-out problems. Compared at the same stage of training, rollout recycling produces roughly 34x greater KL divergence from the reference policy than execution-only GRPO. Over training, trained tokens per GRPO sequence fall from 185 to 86 as training pass rate increases, while held-out performance declines or fails to sustain early gains. We characterize these patterns as single-adapter policy drift: the shared trainable parameters increasingly depart from pretrained code-generation behavior without corresponding gains in generalization. The results show that higher training reward alone is not evidence of generalizable code self-improvement and highlight the importance of held-out evaluation and drift diagnostics.

[55] HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL cs.CLPDF

Xiaoqiang Wang, Ting-En Lin, Yuchuan Wu, Pinyi Zhang, Caixin Kang

TL;DR: HarnessRL 通过共同演化任务采样、rollout 引导和评分标准来应对开放式强化学习中的奖励失效；在 Rubric-Mix-10k 上，它在 4B 和 9B 模型上比最强基线高 2.8–3.1 分，并将达到最佳检查点所需训练步数近乎减半。

Details

Motivation: 基于自然语言标准评分的强化学习可用于开放式生成，但训练推进后奖励可能失去区分度，或高分不再代表回答更好。论文将这两类失效与固定不变的任务采样、探索和评估配置联系起来，提出让训练机制随策略变化而更新。

Result: HarnessRL 为任务采样、rollout 引导和评分标准提供三个可更新接口，并维护训练记忆，供演化器结合单任务历史与跨任务证据定位奖励失效来源。候选更新经提议器和即时评论器循环筛选：只有在当前冻结策略上测量后确实有作用才接受，否则退回修改。作者还构建覆盖医疗、科学、写作、角色扮演和指令跟随五个领域的 Rubric-Mix-10k。该基准上，HarnessRL 在 4B 与 9B 规模下均比最强基线高 2.8–3.1 分，达到最佳检查点所需训练步数接近减半。

Insight: 奖励机制、探索方式和任务分布需要随策略能力变化，且候选调整应在当前策略上验证，以减少无效或投机性奖励更新。报告的收益是在指定五领域套件和两种模型规模上的结果；摘要未说明分数的统计显著性或推广至其他任务的情况。

Abstract: Rubric-based reinforcement learning (RL) trains policies for open-ended generation with natural-language criteria as rewards, but these rewards can become ineffective as training progresses. We identify two failure modes. Under non-discriminative rewards, scores collapse to uniformly low or high values due to insufficient exploration or a mismatch between task difficulty and policy capability. Under spurious rewards, higher scores do not correspond to better responses. Both reflect a training configuration whose task sampling, exploration, and evaluation stay fixed as the policy evolves. We introduce HarnessRL, a training-time harness that co-evolves with the policy through three interfaces for task sampling, rollout guidance, and rubric criteria. HarnessRL maintains a training-time memory from which a harness evolver retrieves within-task history and cross-task evidence. Using this evidence, the evolver attributes each reward failure to the interface whose update can address it. It then runs a proposer and just-in-time critic loop in which the critic accepts a candidate update only after measuring its effect on the current frozen policy, returning rejected candidates for revision. We also curate Rubric-Mix-10k, a five-domain training and evaluation suite for open-ended RL across medical, science, writing, role-play, and instruction-following tasks. On Rubric-Mix-10k, HarnessRL outperforms the strongest baselines by 2.8 to 3.1 points across both 4B and 9B model scales and nearly halves the training steps to its best checkpoint. Code and data are available at https://github.com/Tongyi-ConvAI/Qwen-Character/tree/main/HarnessRL.

[56] WenSight-Decision-1 for Direct Decisions under E-commerce Rules cs.CLPDF

Mingwen Liu, Jun Sun, Sichen Wang

TL;DR: WenSight-Decision-1 将电商规则和案例证据映射为选项分数，在六类运营任务上统一决策接口；在 1,440 个合成留出案例上，准确率从 Qwen3.5-9B 的 61.60% 提升到 92.92%。

Details

Motivation: 电商运营需要把自然语言规则应用到具体案例证据上。论文希望建立一个共享模型，直接为给定候选选项评分，并覆盖六类任务及六种语言示例。

Result: 作者以参考选项作为监督微调 Qwen3.5-9B，训练和推理阶段使用相同的选项评分方式。在 1,440 个保留用于最终评估的合成运营案例上，相同输入条件下，WenSight-Decision-1 准确率为 92.92%，基础 Qwen3.5-9B 为 61.60%。联合微调后六个任务的表现均有提升。

Insight: 将规则和案例决策转化为对候选选项的统一打分，可以在多种运营判断任务间共享一个接口。结果来自合成评测集；摘要未说明真实业务数据上的验证、各语言具体表现或误判代价。

Abstract: E-commerce decisions require applying natural-language rules to the available case evidence. We introduce WenSight-Decision-1, a shared model that scores supplied options directly from rules and case information for six operational tasks, with examples spanning six languages. We fine-tune Qwen3.5-9B with reference options as supervision, using the same option scores for training and inference. On 1,440 synthetic operational cases reserved for final evaluation, accuracy rises from 61.60% for Qwen3.5-9B to 92.92% for WenSight-Decision-1 under identical inputs and scoring. Joint fine-tuning improves all six tasks, bringing heterogeneous operational judgments under a common decision interface.

[57] Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties cs.CL | cs.AIPDF

Afonso Baldo, António Farinhas, Marcos Treviso, Ricardo Rei, Nuno M. Guerreiro

TL;DR: Dynalize 在偏好奖励模型训练期间，从 rollout 中识别模型正在奖励的不良行为并动态生成惩罚；在临床咨询和创意写作实验中，它抑制奖励投机行为、延长有效训练并改善下游表现。

Details

Motivation: 开放式生成任务难以完全验证输出，也难以穷尽写成规则，因此偏好奖励模型仍有价值，但策略可能只提高奖励分数而不改善甚至损害回答质量。论文研究如何在策略训练进行时在线缓解这种奖励投机。

Result: Dynalize 定期异步分析训练 rollout，由语言模型为奖励模型似乎偏好的不良行为编写惩罚，再由评审器将惩罚应用到每条 rollout。在临床咨询和创意写作中，仅用奖励模型训练会持续推高其自身分数，但最终使外部基准表现下降；加入 Dynalize 后训练能持续更久且下游表现改善。方法还抑制了无惩罚时会扩散至大多数回答、甚至扩散到未被奖励模型显式评分的推理中的行为模式。收益在多轮训练中仍存在，即使奖励模型之后针对已发现行为重新训练也保持收益。

Insight: 在训练过程中审计奖励模型偏好的行为并逐步加惩罚，可应对静态偏好模型难以覆盖的投机模式；问题行为还可能泛化到未直接计分的推理内容。摘要未提供具体基准提升数值、惩罚生成成本或不同任务上的失败率。

Abstract: Reinforcement learning for language models has moved from preference reward models to verifiable rewards, which sustain much longer optimization. Preference models are prone to reward hacking: the policy raises their score without improving response quality, or even while degrading it. Yet for open-ended tasks, where quality can neither be verified nor fully written down as rubrics, they remain valuable. We introduce Dynalize, which mitigates reward hacking online, during policy training. Periodically and asynchronously, a language model analyzes training rollouts and writes penalties for undesirable behaviors that the reward model appears to reward; a judge applies them to every rollout. In clinical counseling and creative writing, training on the reward model alone raises its score continuously but ultimately leads to a decline in external benchmarks. Dynalize sustains longer training and results in improved downstream performance. For example, we find that our method suppresses patterns incentivized by preference models that, without penalties, spread to most responses, and even to the model’s reasoning, which the reward model does not explicitly score. Furthermore, we show that its gains hold in multi-turn training, and even after we retrain the reward model on the discovered behaviors.

[58] Hierarchical Token Merging with Recursive Autoencoders for Unsupervised Grammar Induction cs.CL | cs.AIPDF

Murat Guney Kemal, Robert Lieck

TL;DR: 论文提出递归层次 token 合并自编码器，尝试仅凭压缩和重建目标从序列中无监督发现语法层次；在合成歧义语法上，测试集解析有效率和精确重建率均为 93%，高于最强朴素基线的 42% 解析有效率。

Details

Motivation: 层次抽象能把复杂序列组合成多级表示，但在没有显式结构监督时如何从数据学到有意义层次仍是难题。自编码器虽通过重建学习压缩表示，单纯重建目标却未必会产生有意义的组合结构。

Result: 模型递归合并相邻 token，借助变分信息瓶颈压缩整条序列，并通过动态规划对可能的二叉树结构边缘化，联合学习表示与层次。为缓解训练不稳定，作者将独立训练的模型通过无监督共识和蒸馏组合。在合成歧义语法实验中，留出序列上的语法解析有效率和精确重建率均达 93%，最强朴素基线的解析有效率为 42%。

Insight: 压缩式层次 token 合并可能在无结构标签时发现可解释的组合结构，但证据目前是合成语法上的概念验证。摘要未报告自然语言或其他模态上的结果，也未说明不同语法复杂度下的扩展性。

Abstract: Hierarchical abstraction is a fundamental principle across many data modalities, enabling complex sequences to be represented compositionally at multiple levels. Learning such hierarchies directly from data, without explicit structural supervision, remains a central challenge in representation learning. We investigate whether meaningful hierarchical structure can emerge from compression alone. While autoencoders learn compressed representations through reconstruction, this objective does not necessarily favour meaningful compositional structure. We study this challenge through unsupervised grammar induction, where the underlying hierarchy can be evaluated explicitly. We introduce a hierarchical token-merging autoencoder that recursively combines adjacent tokens into increasingly abstract representations, compressing entire sequences through a variational information bottleneck. The model jointly learns representations and hierarchical structure by marginalising over possible binary trees using dynamic programming. To address training instability, we combine independently trained models through unsupervised consensus and distillation. Experiments on a synthetic, ambiguous grammar demonstrate the emergence of meaningful hierarchical structure without structural supervision. The resulting model achieves 93% grammatical parse-validity and 93% exact reconstruction on held-out sequences, compared with 42% parse-validity for the strongest trivial baseline. These findings establish a proof of concept for hierarchical token merging as a mechanism for unsupervised structure discovery and abstraction learning.

[59] HoloCosm: Beyond Single-Metric Benchmark Compression Across Multiple Metrics cs.CLPDF

Yupei Lin, Zhongzhan Huang, Junxin Li, Guoming Ling, Shanshan Zhong

TL;DR: HoloCosm 面向多指标基准压缩，联合保留不同模型和指标的分数与排名；在 15 个语言、多模态、VLA 和智能体基准配置（85 项指标）上压缩 80% 时，平均及最差分量 NMAE 分别下降 35.3% 和 39.5%，平均 Spearman 相关从 0.884 升至 0.915。

Details

Motivation: 基准压缩通过少量样本降低评测成本，但现有方法通常按单一指标选样。由于模型在不同指标上的排名相关性可能较弱，针对某一指标选出的子集未必能保留其他维度的分数和排名。

Result: 作者发现，在压缩 80% 的情况下，五种单指标基线中每一种在超过 70% 的跨指标比较里，其分数重建误差都高于同规模随机抽样。HoloCosm 联合最小化源模型和多项指标上的归一化分数重建误差，同时限制排名失真。在 15 个基准配置、85 项指标的评估中，压缩 80% 时相较整体最强基线 Random，平均 NMAE 降至 0.236（下降 35.3%），最差分量 NMAE 降至 0.330（下降 39.5%），平均 Spearman 相关由 0.884 升至 0.915；单指标评测下也保持竞争力。

Insight: 压缩多维评测集需要同时顾及跨指标分数重建与模型排序，而非优化某个单独指标。结果覆盖多类基准，但摘要未说明选样计算开销或在不同压缩率下的全部数值。

Abstract: Benchmark compression reduces evaluation cost by selecting a small set of samples. Existing methods typically guide selection using a single metric, although benchmarks often assess models along multiple dimensions. We find that model rankings are often weakly correlated across metrics, indicating that different metrics capture distinct aspects of model performance. Samples selected for one metric do not reliably preserve scores and rankings on others when used for compressed evaluation. At 80% compression, each of the five single-metric baselines yields higher reconstruction error than equal-size random sampling in over 70% of cross-metric comparisons. To address this challenge, we present HoloCosm, which jointly minimizes normalized score-reconstruction error across source models and metrics while constraining ranking distortion. We evaluate HoloCosm across multiple compression rates on 15 language, multimodal, VLA, and agent benchmark configurations with 85 component metrics. With 80% compression, HoloCosm outperforms Random, the strongest baseline overall: average NMAE drops by 35.3% to 0.236, worst-component NMAE drops by 39.5% to 0.330, and average Spearman correlation rises from 0.884 to 0.915. In addition, HoloCosm remains competitive with existing baselines in single-metric evaluation. We will release our code and experimental results.

[60] RESIN: Multi-Chain Reasoning with Entity-Structured Indexing cs.CLPDF

Xuanyu Meng, Jash Rajesh Parekh, Jiawei Han

TL;DR: RESIN 以实体结构索引把多链推理问题拆成逐跳证据需求，从原始来源检索少量证据并延迟选择最终推理链；在 HolisQA 上取得 82.92 的上下文召回和 76.70 的答案相关性。

Details

Motivation: 多链推理难点在于证据分散，且不同推理部分可能依赖相互作用的证据链。平坦检索会塞入大量松散相关文本，而过早确定单一路径又可能丢弃只有结合其他事实才有用的证据。

Result: RESIN 将问题拆成聚焦的单跳需求，为每一跳检索少量以原始来源为依据的单元，并将相关证据放入紧凑推理缓冲区。缓冲区会保留备选支持证据，直到最终组合时再作决定，同时显式记录中间判断。在完整 HolisQA 上，所有问题的上下文召回为 82.92、答案相关性为 76.70；五个领域的 CR 和 AR 高于已发布基线中的相应结果，四个领域的 F1 更高。消融显示，类型化证据多样性和逐跳分解都不可或缺。

Insight: 多链推理应精确组织证据并延迟过早承诺，才能兼顾召回、推理透明度和避免脆弱单路径。摘要没有说明这些指标的具体基线差值，也未报告检索延迟或缓冲区大小。

Abstract: Multi-chain reasoning is a challenging task not only because the multi-chain evidence is dispersed, but also because different parts of reasoning may require distinct, interacting chains of support. Flat retrieval can flood a model with loosely related text, whereas early selection of a single reasoning path can discard evidence that becomes useful only after other facts are considered. We introduce RESIN (Reasoning with Entity Structured Indexing), which uses structure to locate evidence while keeping the original source as the basis for reasoning. RESIN decomposes a question into focused one-hop needs, retrieves a small set of source-grounded units for each hop, and places only this relevant evidence in a compact reasoning buffer. The buffer preserves alternative supports until final composition, making intermediate decisions explicit without forcing a brittle path. On the full HolisQA benchmark, RESIN achieves 82.92 context recall and 76.70 answer relevancy over all questions, with higher reported CR and AR across five domains and higher F1 in four domains than the published baselines. Ablations show that both typed evidence diversity and hop-wise decomposition are essential, supporting a clear conclusion: Precise evidence organization and delayed commitment make multi-chain reasoning effective and auditable. Our code is available at https://github.com/RamonMeng/RESIN.

[61] Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability cs.CLPDF

Zhengyue Zhao, Liwei Jiang, Yejin Choi, Chaowei Xiao1

TL;DR: 论文提出在预训练与后训练之间加入安全中训练，让模型以与一般推理共享的结构学习安全能力；相较 Qwen3 1.7B、4B 和 8B 基础模型，攻击成功率降低 72%–81%，同时一般能力提高，且安全性较能抵御后续仅做通用任务的训练。

Details

Motivation: 大语言模型安全通常在后训练阶段对齐，近年也有工作把安全数据加入预训练，但后续不再使用安全数据时安全性可能退化，且安全提升常以一般能力为代价。论文据此将安全看作可通过与一般任务共享的推理结构学习的基础能力。

Result: 安全中训练位于预训练和后训练之间，使用安全与效用提示生成结构化文档和推理轨迹，再用于联合训练安全与通用推理。相较 Qwen3 1.7B、4B 和 8B 基础模型，平均攻击成功率分别在报告范围内降低 72%–81%，一般能力也提升。后续只做通用 SFT 后，中训练模型大体保留安全性，在一般能力不受损的情况下攻击成功率低于基线；只做通用 GRPO、不给安全奖励时也保留安全性。以中训练模型初始化的后续强化学习可达到更好的安全—效用权衡。

Insight: 若安全推理与一般任务共享理解意图、运用知识和评估后果的结构，较早将安全内化可能降低后续训练造成的安全遗忘和安全—能力冲突。摘要未给出攻击类型覆盖、绝对 ASR 数值或所有训练组合的细节。

Abstract: Safety alignment for large language models (LLMs) is primarily performed during post-training, while recent work has also begun incorporating safety data into pre-training. Yet safety can degrade when later training no longer includes safety data, and gains in safety often come at the cost of general capability. This tension suggests that safety and general reasoning should not be treated as separate training goals. We instead view safety as an LLM capability that can be learned through a reasoning structure shared with general tasks. Both require the model to understand intent or task goals, apply relevant knowledge, and reason about potential consequences. We introduce safety mid-training, a stage between pre-training and post-training that jointly develops safety and general reasoning through this shared structure. Given safety and utility prompts, the base model generates structured documents and reasoning traces that are then used for mid-training. Safety mid-training reduces average attack success rate (ASR) by 72% to 81% relative to Qwen3-1.7B, 4B, and 8B base models, while also improving general capability. The benefits persist through subsequent post-training. After general-only SFT, mid-trained models largely retain their safety, achieving lower ASR than the baselines without sacrificing general capability. They also retain safety under general-only GRPO without safety rewards, while subsequent RL reaches a better safety-utility frontier when initialized from the mid-trained models. Together, these results suggest that learning safety and general reasoning within a shared mid-training stage can reduce the safety-utility trade-off and make safety more robust to subsequent training.

[62] RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development cs.CL | cs.AI | cs.CVPDF

Yuchi Wang, Rongyao Fang, Weihao Yu, Xilin Wei, Tiancheng Gu

TL;DR: RefCraft 构建 300 个网页开发任务，要求智能体结合文字指令与 1–5 个参考网页截图，经过渲染、视觉检查和迭代完成页面；九个前沿模型中最强系统仅得 52.12/100，显示视觉参考条件下的智能体生成仍有明显困难。

Details

Motivation: 现实生产任务中的智能体不仅要会使用工具和编程，还需有效吸收视觉参考。已有网页基准多聚焦复刻目标图或仅凭文字生成，较少覆盖开发者根据若干参考网页工作并迭代检查的常见流程。

Result: RefCraft 包含 300 个多模态网页开发任务，每项结合自然语言指令和一至五张参考网页截图。智能体需要选择性理解并使用参考，在渲染后进行视觉检查，再迭代改进输出。开放式产物用可独立验证的关键点以及程序化和视觉语言模型评判器共同评估。对九个前沿大语言/多模态模型的评测中，最强系统总分为 100 分中的 52.12 分。

Insight: 将视觉参考融入多轮网页开发，比只做截图重建或文字生成更贴近实际，也暴露当前智能体在选择、整合参考和视觉迭代上的不足。摘要未给出不同模型、任务类型或评分组件的细分结果。

Abstract: Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks. We study this crucial capability through webpage development, a practical and readily evaluable setting in which existing benchmarks remain largely idealized: they typically focus on reconstructing visual targets or generating webpages from textual instructions alone, overlooking the more common workflow of developing from visual examples. Specifically, we introduce RefCraft, a benchmark comprising 300 realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages. Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs. We evaluate the resulting open-ended artifacts using independently verifiable keypoints and complementary programmatic and VLM judges. Across nine frontier (M)LLMs, the strongest system achieves an overall score of only 52.12 out of 100, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.

[63] Competing Learning in Hybrid Architectures cs.CLPDF

Xingyu Shen, Yingfa Chen, Ziqing Qiao, Chaojun Xiao, Xu Han

TL;DR: 混合架构中的现代递归层和 softmax 注意力会竞争学习相同的检索函数；递归层通常学得更快并压制注意力的学习信号。训练中途停止更新递归层有助于注意力学会检索，基于此的分阶段训练在最高 1B 参数模型上改善语言建模、常识推理和多种检索任务。

Details

Motivation: 混合语言模型将线性时间递归与 softmax 注意力结合：递归层吞吐高，但因把上下文压入固定大小状态，在上下文信息检索上弱于注意力。论文研究两类混合器能否有效分工，以及为什么注意力有时未能学会检索。

Result: 作者揭示“竞争学习”现象：不同类型的混合层可能学习同一功能，而一层学会后会压低其他层对应的学习信号。归因分析显示，在常见数据分布下，现代递归层比 softmax 注意力更快学会检索，导致注意力多数情况下没能学会该功能；训练中途停止递归层权重更新可帮助注意力学会检索。由此提出自然语言混合架构预训练的分阶段训练程序。在最高 1B 参数的模型上，实验报告语言建模损失、常识推理及不同上下文长度下多种检索任务均有显著改善。

Insight: 混合架构组件的能力并非自动互补：先学会某一功能的混合器可能屏蔽其他组件的梯度信号。阶段性冻结递归层可释放注意力的学习机会；摘要未给出各任务的具体提升量或训练额外成本。

Abstract: Many modern language models have adopted hybrid architectures that incorporate both linear-time recurrence and softmax attention. Modern recurrence layers provide a highly expressive mixing mechanism while ensuring high throughput by compressing contextual information into a fixed-size state, but underperform softmax attention in contextual information retrieval. In this paper, we reveal a Competing Learning phenomenon where different kinds of mixers compete to learn the same function, and when any layer learns a function, it suppresses the corresponding learning signal for other layers. Layer attribution analysis reveals that under typical data distributions, modern recurrence learns retrieval functions faster than softmax attention, causing attention to fail to learn retrieval in most cases. Stopping weight updates in recurrence in the middle of training helps attention learn retrieval. Based on these results, we propose a staged training procedure for pre-training hybrid architectures on natural language. Empirical results on models up to 1B parameters show that our method results in substantial improvement in language modeling loss, common-sense reasoning, as well as a wide range of retrieval-related tasks across different context lengths.

[64] OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory cs.CL | cs.AI | cs.CV | cs.LG | cs.MMPDF

Yu Zhou, Cheng-Fu Yang, Rui Sun, Di Wu, Sibo Peng

TL;DR: OpenVLHarness 将异构工具接口、证据呈现、持久多模态记忆和可执行代码组合为视觉智能体推理框架；在 22 个基准、8 个模型上比匹配的无工具基线高 9.7–12.6 分，并以不超过 37% 的 token 成本超过若干最高推理预算对照系统。

Details

Motivation: 视觉推理任务需要的证据可能无法仅靠模型内部推理获得，常要结合视觉定位、外部知识和数值测量。工具产生的异构结果还须转换成模型可理解的形式，并在多步推理中持续保留；单纯增加推理预算收益有限。

Result: OpenVLHarness 包含能力层和持久多模态记忆：前者通过任务级接口连接不同工具后端，并将输出渲染为模型可读证据；后者把图像、结构化数据和网页引用作为可寻址材料贯穿当前查询。代码接口可对存储数据执行程序，组合不同工具的结果。该迭代式证据获取、检查和组合流程在 22 个基准、8 个开源及专有骨干模型上评估，平均比匹配的无工具基线高 9.7–12.6 分，其中 Qwen3-VL-8B 提升 11.2 分、Kimi K3 提升 9.7 分、GPT-6 Astra 提升 9.9 分。GPT-6 Luna 使用中等推理预算时，超过基础模型、Codex 和最高预算 Visual Sketchpad 对照 7.7–8.5 分，模型 token 成本至多为对照的 37%。

Insight: 把工具证据以可检查、可保留、可计算的形式接入多模态推理，比单纯延长内部推理更能解决需要外部证据的任务。对照结果跨 22 个基准，但摘要未说明所有任务的分数分布或运行时间成本。

Abstract: Multimodal large language models (MLLMs) have made rapid progress on visual reasoning, yet many tasks require evidence that is difficult to obtain through internal reasoning alone: our controlled study on 22 benchmarks shows that simply increasing the inference budget of the base model yields only modest gains. Such tasks demand grounded perception, external knowledge, and numerical measurements obtained through heterogeneous tools, whose outputs must be made interpretable to the model and retained across reasoning steps. We introduce OpenVLHarness, a multimodal reasoning harness organized around a capability layer and persistent multimodal memory. The capability layer combines task-level tool interfaces with evidence rendering, providing consistent access to heterogeneous backends and converting their outputs into model-readable observations. Persistent multimodal memory retains the corresponding images, structured data, and web references as addressable artifacts throughout each query. A coding interface executes generated programs over stored data, enabling computations that combine results from multiple tools. Together, these mechanisms support an iterative process of acquiring, inspecting, and composing multimodal evidence. Evaluated on 22 benchmarks with eight open and proprietary backbones, OpenVLHarness improves average scores by 9.7-12.6 points over matched no-tool baselines, including +11.2 with Qwen3-VL-8B, +9.7 with Kimi K3, and +9.9 with GPT-6 Astra. With GPT-6 Luna at medium reasoning effort, OpenVLHarness outperforms the base model, Codex, and Visual Sketchpad at maximum effort by 7.7-8.5 points while incurring at most 37% of their model-token cost.

cs.CV [Back]

[65] I$^2$-Med: Interpretable Medical Inference through Visual-Guided Dynamic Logits Calibration cs.CV | cs.CLPDF

Md Sajid, Shreeyut Maheshwari, Akshat Mishra, Ritik Mishra, Mohammad Tanveer

TL;DR: I²-Med在医学视觉语言模型生成时逐词检查候选内容是否得到影像和当前推理支持，并动态调整logit以降低视觉幻觉。

Details

Motivation: 医学图像理解和推理需要精确视觉定位及临床连贯解释，但成像模式复杂且专家标注稀缺，常规推理方法可能生成与影像无关的文本，限制可靠性。

Result: I²-Med是无需额外训练的推理期框架。文本生成过程中，视觉指导模块检查每个候选词是否获得医学图像支持且与当前推理一致，再调节输出分数以偏向有视觉依据的词，不改变模型参数。在覆盖八种医学成像模态和五类临床问题的OmniMedVQA实验中，作者报告该方法能生成更有临床意义、视觉忠实度更高的推理链，并在标准医学VQA指标下减少视觉幻觉。

Insight: 逐词视觉依据检查可将图像忠实性约束直接加入解码，而非只依赖后训练。摘要未给出具体成绩、额外延迟和不同模态的效果，实验指标改善不等同于临床有效性验证。

Abstract: Medical vision language models have demonstrated strong potential for medical image understanding and reasoning. However, their interpretability and reliability remain limited because medical vision language tasks require precise visual grounding and clinically coherent explanations. This is particularly challenging due to complex imaging patterns and scarce expert annotations. These challenges reduce the effectiveness of conventional reasoning methods and lead to visual hallucination, defined as the generation of text that is not grounded in the given visual input, which further limits their applicability in clinical practice. To mitigate these challenges, we propose an interpretable medical inference (I$^2$-Med) framework that enhances visual faithfulness during inference without additional training to improve reasoning and explanation faithfulness in medical vision language models. During text generation, I$^2$-Med uses a visual guidance module to check whether each candidate word is supported by the medical image and consistent with the current reasoning. Based on this guidance, the model’s output scores are adjusted to favor visually grounded words, without changing the model parameters. Extensive experiments on the OmniMedVQA benchmark, covering eight medical imaging modalities and five clinical question types, demonstrate that our proposed I$^2$-Med framework generates more clinically meaningful and visually faithful reasoning chains and reduces visual hallucination under standard medical VQA evaluation metrics. The source code is publicly available at \url{https://github.com/mtanveer1/l2-med}.

[66] Latent Multimodal Memory Reasoning cs.CV | cs.AI | cs.CLPDF

Cai Ke, Xinghao Chen, Yuanchen Bei, Yuni Lai, Keyu Chen

TL;DR: LMMR把跨模态长期记忆组织在稀疏自编码器概念空间中，构建按回合更新的多模态记忆图，并将图结构上下文注入生成模型。

Details

Motivation: 长期个性化智能体需整合分布在不同会话和模态中的历史证据；扁平跨模态方法容易过早将视觉证据转成文本、受过期观察影响，而原始异构记忆又使细粒度定位和跨模态推理成本高。

Result: LMMR将文本隐藏状态和视觉图块特征映射至共享稀疏自编码器概念空间，以稀疏激活作为公共语义地址，动态构建回合级多模态记忆图。图编码器聚合任务相关潜在子图中的跨模态和时间依赖，将图读出作为连续残差注入生成骨干，同时保留原生视觉token以维持细节。作者在九项Mem-Gallery任务和十二种MemEye设置上测试，报告LMMR在捕捉复杂多模态依赖方面显著超过先进基线，并支持忠实、有依据的回答。

Insight: 共享潜在概念地址与记忆图尝试兼顾多模态检索效率、时间关系和精细视觉依据。摘要未给出数据集细节、具体分数、记忆规模或推理成本，不能进一步比较收益与开销。

Abstract: Long-term personalized agents are required to reason over distributed history across modalities and evolving dialogue states. While existing flat cross-modal methods compress visual evidence prematurely and struggle with temporally stale observations, they neglect the necessity to unify multimodal and cross-session evidence. Crucially, raw multimodal memories are inherently entangled and heterogeneous, making fine-grained visual grounding and cross-modal reasoning computationally prohibitive. To this end, we present latent multimodal memory reasoning, i.e., LMMR, a novel neuro-symbolic framework that shifts long-term multimodal memory organization into a continuous latent space. Specifically, (i) instead of converting images to text and relying on static retrieval, LMMR maps both textual hidden states and visual patch features into a shared Sparse Autoencoder concept space, establishing sparse activations as common semantic addresses to dynamically construct a turn-level multimodal memory graph. (ii) A graph encoder then aggregates cross-modal and temporal dependencies across the task-specific latent subgraph, injecting the graph readout as a continuous residual rollout into the generation backbone while retaining native visual tokens to preserve fine-grained details. Extensive experiments on nine Mem-Gallery tasks and twelve MemEye settings demonstrate that LMMR significantly outperforms state-of-the-art baselines in capturing complex multimodal dependencies while enabling faithful and grounded responses.

[67] EMMA: Event-Driven Multimodal Memory System with Multi-Agent Retrieval cs.CV | cs.AIPDF

Ruochen Zhao, Jiazheng Kang, Gezi Li, Dongying Lin, Ting Bai

TL;DR: EMMA围绕语义事件构建长视频分层记忆，并以多智能体分解查询、逐级检索和核验证据，减少固定窗口切分造成的事件破碎与噪声检索。

Details

Motivation: 长视频问答需跨越远超多模态大模型上下文长度的数小时内容。按固定长度片段压缩可能把完整事件切开，也会把无关内容合并，导致检索噪声。

Result: EMMA包括三阶段：双流事件检测器融合视觉语义变化与语音转场，自适应定位事件边界；记忆按原始视频、片段级和事件级三层组织，逐步提高语义抽象程度；多智能体协作对查询做分解、分层搜索、证据验证和答案生成，从粗事件逐步缩小到精确片段。三个具有挑战性的基准实验显示，该事件对齐记忆系统取得明显改进。项目提供代码、提示模板和复现脚本。

Insight: 以事件边界而非固定时长组织记忆，可让检索保留语义完整单元，再回到具体片段核对证据。摘要未列出三个基准、成绩和分段检测误差，提升幅度与检索成本无法量化。

Abstract: Long video understanding requires reasoning over hours of content that far exceed the context window of multimodal large language models (MLLMs). Existing approaches compress video into fixed-length clips, yet events are fragmented and unrelated content is merged, yielding noisy retrieval. To overcome this challenge, we propose EMMA, an Event-driven Multimodal Memory system with Multi-Agent retrieval, which organizes memory around semantic events in three stages: (1) Event segmentation: a dual-stream detector adaptively locates event boundaries by fusing visual semantic drift with spoken transitions, rather than fixed windows. (2) Memory construction: built on the detected events, a three-level hierarchy of raw video, clip-level memory, and event-level memory stores content at progressively higher semantic abstraction. (3) Memory retrieval: a multi-agent framework collaboratively queries this hierarchy through query decomposition, tiered search, evidence validation, and answer generation, narrowing evidence from coarse events to precise clips. Extensive experiments on three challenging benchmarks demonstrate substantial improvements, showing the effectiveness of our event-aligned memory system. Code, prompt templates, and reproduction scripts are available at https://github.com/BAI-LAB/EMMA.

[68] Path-Conditioned Attention Mobility: From Depth to Execution Hubs in SegFormer cs.CVPDF

Aleksandr Druzhinin

TL;DR: 该研究发现SegFormer中部分任务相关注意力头可在尊重其计算上下文的情况下迁移到相邻深度，并展示将相同头预算集中到更少执行深度时的部署延迟收益。

Details

Motivation: Transformer压缩常把保留注意力容量和执行这些容量的网络深度混为一谈。研究考察层级视觉Transformer中注意力功能能否在深度局部移动，并指出只移动头而不保留其归一化、空间降采样等上下文可能破坏功能。

Result: 以SegFormer-B5人体分割为对象，作者发现有些任务相关注意力头可迁移到附近层，但必须保留计算路径；单独迁移头不足，联合适配迁移头及其表示边界可改善功能保留，替换供体特定归一化或空间降采样预处理会大幅退化。研究将固定的Stage-3 120个头、40个执行深度重新组织到20个执行枢纽。布局匹配的三随机种子恢复对照中，K20的mIoU为0.960686±0.000059，对应State120为0.960400±0.000074；配对差为+0.000285，但图像配对bootstrap置信区间含零，因此解释为无可测质量损失，而非统计优势。FOLDED_VEC保留供体路径并向量化执行；相对同一State120父模型，NVIDIA L4端到端延迟在批量1、4、8时分别下降5.28%、3.15%、3.58%。

Insight: 注意力冗余不仅可能表现为可删容量，也可能表现为可压缩执行深度，但迁移时需保留定义该功能的计算上下文。速度提升幅度为数个百分点，质量比较没有统计显著优势；结果只在SegFormer-B5的人体分割设置验证。

Abstract: Transformer compression often couples two distinct quantities: retained attention capacity and the number of sequential depths at which that capacity is executed. We study whether useful attention functions in a hierarchical vision transformer are locally mobile across depth. Using SegFormer-B5 for human segmentation, we find that some task-relevant attention heads can be relocated to nearby depths, but only when their computational context is respected. Head-only relocation is insufficient, while joint adaptation of the relocated head and surrounding representation boundary improves functional preservation; replacing donor-specific normalization and spatial-reduction preprocessing causes large degradation. We use this path-conditioned mobility to reorganize retained attention from depth into width. Starting from a fixed parent with 120 retained Stage-3 heads across 40 attention execution depths, we consolidate the same 120-head budget into 20 execution hubs. Under a layout-matched three-seed recovery control, K20 reaches 0.960686 +/- 0.000059 mIoU versus 0.960400 +/- 0.000074 for the corresponding State120 control. The paired point difference is +0.000285, while a paired image-bootstrap confidence interval includes zero; we therefore interpret the result as no measurable quality penalty rather than statistical superiority. For deployment, FOLDED_VEC preserves donor-specific paths while vectorizing their execution. Relative to the same 120-head State120 parent, it reduces end-to-end NVIDIA L4 latency by 5.28%, 3.15%, and 3.58% at batch sizes 1, 4, and 8. These results suggest that attention redundancy includes not only removable capacity but also execution-depth redundancy, provided that the computational path defining a relocated function is preserved.

[69] EAGLE: Entropy-Aware Geometric Learning Enhancer for Partially Polygon-Supervised Object Localization cs.CV | cs.LGPDF

Naveen Stewart

TL;DR: EAGLE利用信息论和统计形状分析估计类别几何复杂度，再以轻量后训练检测头进行边缘细化和定位重评分，改善部分多边形监督下的目标定位。

Details

Motivation: 仅使用边界框训练的检测器几何定位精度有限，尤其在高IoU阈值下；加入多边形掩码监督虽能提供更丰富几何信息，但对形状简单的类别，分割损失可能干扰边框定位并使表现下降。

Result: 作者提出按类别量化几何复杂度的数学框架，并据此构建EAGLE后训练检测头，利用熵信息进行边缘细化和定位重评分。MS COCO和Pascal SBD分析显示，该熵框架能预测形状重建误差及掩码派生框与标准框的准确率差异。集成EAGLE后，已有先进检测框架在COCO上的mAP50-95提高0.9分，观测推理开销很小。

Insight: 按类别几何复杂度调节掩码信息利用方式，可缓解复杂几何监督对简单形状定位的干扰。摘要未给出Pascal SBD具体增益、复杂度分组细节和设备级延迟，COCO的0.9点提升不代表所有类别都受益。

Abstract: Object detection is a critical application of computer vision, yet models trained uniformly on bounding boxes lack the geometric precision to be fully effective in inference settings, especially at high IoU thresholds. One common approach is the introduction of polygon-supervised localization, where a model is trained explicitly for localization but with access to polygon masks during training. Among such methods, partial polygon supervision is viable due to the higher annotation costs of developing fully supervised training datasets. However, this introduces a further challenge: in object classes of relatively low shape complexity, the addition of training masks and subsequent segmentation loss may obscure box localization, often worsening model performance despite access to deeper geometric information. To address the practical limitations of both full and partial polygon supervision, we introduce a mathematical framework for quantifying per-class geometric complexity by combining information theory and statistical shape analysis. With this framework, we develop Entropy-Aware Geometric Learning Enhancer (EAGLE), a lightweight, post-training model head that learns geometric information for the localization model, improving accuracy via edge refinement and localization rescoring using entropy-informed updates. Empirical analysis across MS COCO and Pascal SBD demonstrates the entropy framework meaningfully predicts shape reconstruction error and accuracy differences among mask-derived and standard box predictions, while the implemented EAGLE framework achieves consistent improvement on already SOTA detection frameworks (+0.9 mAP50-95 points on MS COCO) at minimal observed inference overhead.

[70] AI-Experiment: an augmented reality art project co-created with artificial intelligence cs.CVPDF

Bernhard Egger, Hans Furer

TL;DR: 该艺术项目由艺术家、计算机科学家和生成式AI共同创作实体画作，并以艺术家作品微调潜扩散模型生成风格相近图像，再制作增强现实视频效果。

Details

Motivation: 论文探索生成式人工智能如何参与实体艺术创作，以及生成图像、艺术家手工重绘和增强现实视频能否组成跨媒介协作作品。

Result: 团队使用特定艺术家的作品照片微调潜扩散模型，生成相近风格的合成作品；艺术家从中选取部分图像，再用丙烯颜料重新绘制到画布上。除实体作品外，项目还加入由多种生成视频模型驱动的增强现实组件，使画作呈现动态效果。论文同时发布部分作品训练数据、生成作品所用代码及增强现实体验。

Insight: 该工作主要记录艺术创作流程与作品发布，而不是生成模型或艺术协作的定量评测。摘要未描述数据授权、模型比较或观众研究结果，不应将其解读为风格复制质量的实验结论。

Abstract: We describe the creation of a physical artwork that was performed in synergy between an artist, a computer scientist and generative AI tools. A latent diffusion model was finetuned on photographs of the artwork of the specific artist and the generative model then generated synthetic artworks in a similar style. From the generated artworks a subset was selected to be painted again with acrylic on canvas by the original artist. In addition to the physical artifact an augmented reality component was added which brings the artwork to life with a set of generative video models. Together with this paper we release part of the artwork (training data), the code that was deployed to generate the artwork as well as the augmented reality experience for arts, research and education under https://github.com/BernhardEgger/ARTimate

[71] LVMH: A Meta-Optimization Harness for Long-Video Agent Memory cs.CV | cs.MAPDF

Liwei Che, Benlin Liu, Dai Li, Xiangyan Li, Zijian Wu

TL;DR: LVMH把长视频记忆策略开发表述为多目标代码搜索，并由此发现SEEM：先从结构化事件文本记忆回答，再按需调用局部画面证据。

Details

Motivation: 长视频智能体需把数小时或数天的多模态证据压进有限上下文，同时满足token和延迟预算。现有记忆策略多由人工设计，缺少能共同优化性能和效率的开发流程，也需防止开发集过拟合与自我评审。

Result: LVMH是围绕固定答案模型运行的自动多智能体记忆策略代码搜索框架，分离实现和独立评审，并设置受保护的观察边界；持久元记忆保存以往提案、轨迹、成本及失败。借此发现的SEEM采用“文本优先、视觉其次”策略：从紧凑结构化事件记忆和证据专家生成回答，只有通过选择性视觉门控时才查看局部帧。策略在EgoLifeQA上进化后可迁移到未见基准；使用GPT-5 mini时，在Ego-R1 Bench和Video-MME（L）均超过Meta-Harness。GPT-5配置在三个基准平均准确率65.7%，比此前最佳高7.7个百分点，平均评测token用量约降低32倍，响应时间约缩短91倍。

Insight: 自动化策略搜索可共同探索记忆组织和调用方式，并用选择性查看画面降低长视频上下文成本。结果跨三个基准报告，但摘要未提供各基准分项数值和搜索计算开销，SEEM的优势需结合开发成本衡量。

Abstract: Long-video agents must map hours or days of multimodal evidence into a bounded context while respecting token and latency budgets, making memory design a context-management challenge. Existing memory systems largely rely on hand-designed policies, leaving open how to jointly optimize memory design for performance and efficiency. We propose Long Video Meta-memory Harness (LVMH), an automated multi-agent framework that formulates memory-policy development as multi-objective code search around a fixed answer model. To guard against self-certification and development-set overfitting, LVMH separates implementation from independent review and enforces protected observation boundaries. Persistent meta-memory records past proposals, traces, costs, and failures to guide subsequent policy revisions. Through this harness, we discover Self-Evolving Evidence Memory (SEEM), a “text-first, vision-second” policy that answers from compact structured event memory and evidence specialists, consulting localized frames through a selective visual gate. The policy evolved on EgoLifeQA generalizes to unseen benchmarks, outperforming Meta-Harness on both Ego-R1 Bench and Video-MME (L) with GPT-5 mini. Across the three benchmarks, SEEM with GPT-5 achieves 65.7% average accuracy, surpassing the previous state of the art by 7.7 percentage points with approximately 32x lower average evaluation token use and 91x shorter average response time.

[72] Beyond Generator-Specific Shortcuts: Open-World AI-Generated Video Detection via Adversarial Training cs.CVPDF

Qi Zhou, Yuanfan Li, Huiqi Deng, Hongbin Pei, Yazhe Tang

TL;DR: FORGE以token级对抗训练压制视频生成器特有捷径，同时保留真伪判别线索，在未见生成器和后处理攻击上改善AI生成视频检测。

Details

Motivation: AI生成视频检测对未见生成器的开放世界泛化较差。作者发现检测器强烈编码生成器特定线索，这与域内到域外性能下降相关；简单扰动此类token反而可能降低域外表现，说明生成器线索与有效取证证据并非完全分离。

Result: FORGE用梯度引导嵌入扰动识别与生成器捷径相关但对真实/伪造判别相对不重要的token，并在训练中选择性扰动，以压制生成器相关模式而保留有用取证信息。攻击者与检测器同步对抗训练，促使检测器学习更具迁移性的生成器不变表征。在29个留出基准条目上，平均F1为92.62，比八个强基线高5.31分，并能抵抗四类常见后处理攻击。

Insight: 开放世界检测的关键不是删除所有生成器信息，而是降低捷径依赖同时保留真实取证证据。摘要未列出29项对应的数据集、生成器分布及各攻击结果，代码和数据计划在录用后发布。

Abstract: AI-generated video (AIGV) detection is increasingly important for mitigating misinformation, impersonation, and other malicious uses of synthetic media, yet existing detectors often generalize poorly to unseen generators in open-world settings. To better understand this challenge, we analyze the failure modes of AIGV detectors under cross-generator out-of-distribution (OOD) shifts and observe that detector representations strongly encode generator-specific cues, which are associated with substantial ID-OOD performance drops. Further analysis shows that naively perturbing generator-sensitive tokens can even reduce OOD performance, indicating that generator-related cues and detection-relevant evidence are only partially separable. Motivated by these findings, we propose Forensic Obfuscation of geneRator-specific cues via Gradient-guided Embedding perturbation (FORGE), a token-aware adversarial training framework for open-world AIGV detection. FORGE identifies tokens that are strongly associated with generator-specific shortcuts yet less critical to real/fake discrimination, and selectively perturbs them during training to suppress generator-dependent patterns while preserving useful forensic evidence. Through synchronized adversarial training between an attacker and a detector, FORGE encourages the detector to learn more transferable generator-invariant forensic representations. Experiments on 29 held-out benchmark entries show that FORGE achieves an average F1 of 92.62, outperforming eight strong baselines by 5.31 points, while also remaining robust under four common post-processing attacks. We will release our code and datasets upon acceptance.

[73] EgoVine: Post-Training with Branching Experience Compiled from Real-World Videos cs.CV | cs.AI | cs.ROPDF

Qinchuan Cheng, Bowen Lyu, Zhantao Gong, Angela Yao, Cheston Tan

TL;DR: EgoVine把真实第一人称家庭操作示范编译成可回滚的符号环境，让规划模型探索示范未覆盖的替代动作、延迟后果及失败恢复。

Details

Motivation: 真实第一人称视频记录了多步骤家务流程，但只展示一条已执行轨迹，缺少替代动作及模型犯错后的恢复经验，限制长时程规划训练。

Result: EgoVine将示范集转成可执行符号环境，智能体需在部分观测下协调物体操作并考虑延迟后果；恢复决策状态后可探索替代动作并比较最终结果。联合训练模块在同一模型中学习动作生成和后果预测，并用相同执行监督候选比较及恢复；验证修复为失败尝试增加恢复目标，采样后续之间的比较用于更新策略。规划器更新后返回环境，自己的决策和错误成为下一轮经验。共享基准上，EgoVine-1.5B在主规划任务成功率为81.8%，在更长任务组合中为66.7%，两项均超过所有被评Flash LLM基线。

Insight: 让模型在真实示范编译出的环境中回滚探索，可学习录制轨迹未覆盖的选择及未来后果。摘要未提供环境覆盖范围、真实机器人执行结果和基线具体成绩，符号模拟中的成功不等于物理部署能力。

Abstract: Real-world egocentric videos capture multi-step household procedures, but their recorded trajectories leave alternative actions and recovery from the learner’s own mistakes largely unexplored. We introduce EgoVine, a post-training framework that turns a fixed demonstration set into interactive experience for long-horizon planning. EgoVine operates in executable symbolic environments compiled from real-world video annotations, where an agent must coordinate object manipulations under partial observation and account for delayed consequences. Restoring a decision state allows the learner to explore alternative actions and compare their eventual outcomes. Union jointly trains action generation and consequence prediction in one model, using the same executions to supervise candidate comparison and recovery. Verified repairs add recovery targets from failed attempts, while comparisons among sampled continuations guide policy updates. The updated planner then returns to the environments, with its own decisions and mistakes shaping the next round of experience. On a shared benchmark, EgoVine-1.5B achieves 81.8% success on Main planning tasks and 66.7% on longer task compositions, outperforming all evaluated Flash LLM baselines on both. EgoVine connects real demonstrations to an iterative learning process that trains compact planners on choices and futures beyond the recorded trajectories.

[74] ContraSafe-VL: A Contrastive Safety Preference Dataset for Vision-Language Models cs.CVPDF

Di Wu, Yanyan Zhao, Xin Lu, Hao Yang, Taotao Zhang

TL;DR: ContraSafe-VL构建固定文本、改变视觉上下文的对比式安全偏好数据，使视觉语言模型学习图像内容如何改变安全判断。

Details

Motivation: 多模态安全取决于文字与图像的共同语义，但视觉语言模型可能忽略与安全有关的画面内容，也可能过度解读敏感内容而拒绝无害请求。只用回答偏好监督主要教模型如何措辞，难以教它视觉信息何时应改变安全判定。

Result: ContraSafe-VL对每个固定文本查询构造视觉上下文不同、对应安全标签不同的成对样本，直接监督模型根据视觉上下文进行跨模态安全判断；同时提供基于无害性和帮助性的回答偏好对，鼓励输出兼顾安全与有用性。在三种视觉语言模型架构及多个基准上的实验显示，使用该数据训练能改善跨模态安全识别，在安全性与过度拒绝之间取得更好折衷，并保持一般多模态能力。

Insight: 保持文字固定、只改变图像上下文，有助于隔离视觉线索对安全判断的作用。摘要未给出具体基准、提升数值和数据构建规模，数据集尚计划公开。

Abstract: Vision-Language Models (VLMs) are increasingly deployed as general-purpose multimodal assistants, making multimodal safety alignment an important challenge. Safety depends on the interplay between textual and visual content, with visual context often playing a critical role. However, VLMs may fail to use visual information appropriately, either overlooking safety-relevant visual content or over-interpreting sensitive visual content, leading to unsafe responses or over-refusal. Existing multimodal safety alignment datasets largely follow the response-preference paradigm, which teaches models how to respond but provides limited supervision on how visual context should inform the safety judgment, making it difficult to balance safety and over-refusal. To address this limitation, we introduce ContraSafe-VL, a Contrastive Safety preference dataset for Vision-Language Models. For each fixed textual query, ContraSafe-VL constructs paired visual contexts that result in multimodal queries with different safety labels, providing direct supervision for how visual context should inform multimodal safety judgments. It further provides response preference pairs based on harmlessness and helpfulness to encourage models to generate responses that are both safe and helpful. Experiments across three VLM architectures and multiple benchmarks demonstrate that training with ContraSafe-VL improves cross-modal safety recognition and achieves a better safety–over-refusal trade-off while preserving general multimodal capabilities. The dataset will be publicly released.

Xianmin Ye, Zhiyuan Fan, Tianju Liu, Yuhang Wang, Zhenxing Niu

TL;DR: SIJA将有害图像编辑意图拆分到文本和视觉线索中，揭示现有防护可能分别看输入而漏掉联合指令，并提出重建整体编辑意图的防御原则。

Details

Motivation: 图像编辑模型会综合文本和视觉输入推断编辑意图；若有害请求的指代对象写在文本中，而危险操作和参数隐藏在图像语义线索中，现有逐项审核可能无法识别完整意图。

Result: 作者提出跨模态拆分意图越狱SIJA和安全基准SI2ESBench。攻击把编辑对象放在文字中，并通过辅助视觉提示隐藏安全关键操作及规格。对代表性商业和开源模型的测试显示，SIJA对GPT Image 2的攻击成功率为68.28%，对Qwen-Image-3.0-Pro为91.38%。研究指出模型能从组合输入中恢复并执行分散的编辑意图，但既有安全机制常未识别其有害性。防御原则“编辑意图重建”要求评估联合多模态上下文隐含的完整转换，实验显示可显著提高对这类攻击的拒绝。

Insight: 多模态安全检查应评估联合推断出的编辑变换，而非孤立审核文字或图片。摘要没有列出更多模型、误拒率及防御具体数值；攻击成功率限于SI2ESBench设定。

Abstract: Large image editing models increasingly infer editing intent from both textual and visual inputs, creating an underexplored safety risk when harmful intent is distributed across modalities. In this work, we formalize this split-intent vulnerability and propose the Split-Intent Jailbreak Attack (SIJA), which asymmetrically distributes an unsafe editing intent by specifying the editing referent in text while concealing the safety-critical action and specification through graphical visual semantics in an auxiliary cue. To systematically study this threat, we introduce the Split-Intent Image Editing Safety Benchmark (SI2ESBench) for evaluating image editing models under split-intent inputs. Extensive experiments on representative commercial and open-source models demonstrate that SIJA effectively compromises state-of-the-art image editing systems, achieving ASRs of 68.28% on GPT Image 2 and 91.38% on Qwen-Image-3.0-Pro. Our results expose a mismatch between multimodal editing and safety assessment: the distributed intent can be recovered to execute the requested transformation, yet its harmfulness often remains undetected by existing safeguards. To mitigate this vulnerability, we propose editing-intent reconstruction, a defense principle that explicitly evaluates the complete transformation implied by the joint multimodal context and substantially improves rejection of split-intent attacks. Our findings expose a cross-modal safety gap in modern image editing systems and highlight the need to assess the jointly inferred transformation rather than individual inputs in isolation. https://github.com/yexianmin/SIJA

[76] Grounding Confusable Movie Moments cs.CVPDF

Dong-Hee Kim, Seonwoo Choi, Changbeen Kim, Juyeon Ko, Youngju Choi

TL;DR: CoM²-Bench用只差一个辨别细节的成对查询测试长电影时序定位，发现检索而非片段内定位是主要瓶颈，并以Agentic Window Search改善结果。

Details

Motivation: 电影中的人物、地点和动作反复出现，两个时刻可能只在细节上不同。现有长电影定位基准很少要求模型区分这类易混淆时刻，整体分数也难确定错误发生在候选检索还是局部定位。

Result: CoM²-Bench构造成对查询，每对共享大多数属性，仅在一个辨别细节上不同；实验分析先检索短候选片段、再在片段内定位的两阶段流程。结果显示主要瓶颈是检索：先进检索器容易选到相似但错误的对应时刻，因为它依赖共享内容而忽略区分细节。作者提出无需训练的Agentic Window Search，先把查询拆为属性子问题，逐窗口检查证据，若窗口边缘截断相关证据则扩大范围。与现有检索器和定位器组合后，mIoU从18.5提高到26.8。

Insight: 长视频定位需检索阶段保留辨别性细节，不能只靠片段内定位器纠错；分解查询并按需扩窗可针对这一瓶颈。摘要未给出基准规模和各模型分项结果，mIoU提升限于该配对查询任务。

Abstract: Characters, locations, and actions recur throughout a movie, so many of its moments differ only in fine-grained details. Grounding such a moment in a full-length movie requires resolving that detail rather than matching the attributes it shares with other moments, yet existing temporal movie grounding benchmarks rarely test this. We introduce CoM$^2$-Bench, a benchmark of confusable query pairs, where the two queries in each pair share most attributes but differ in a single distinguishing detail within a full movie, posing a challenge for recent temporal grounding approaches. To locate where grounding breaks down, we analyze a retrieve-then-localize pipeline, in which a retriever first selects a short candidate chunk and a localizer then performs temporal grounding within it. Separating the two stages lets us attribute each failure to either retrieval or localization. We find that retrieval is the major bottleneck: state-of-the-art retrievers often select a moment’s confusable counterpart, because their choices follow the content the two moments share rather than the detail that tells them apart. To address these failures, we propose a simple, training-free remedy, Agentic Window Search (AWS): an MLLM agent $\textit{Decomposes}$ the query into attribute-level sub-queries, then $\textit{Inspects}$ each candidate window against them and $\textit{Expands}$ the window when evidence is cut off at its edge. AWS works with existing retrievers and localizers without additional training, raising mIoU from 18.5 to 26.8.

[77] Pseudo-Label Foreground Quality Weighting for Domain-Generalized Fundus Image Segmentation cs.CVPDF

Lucas Gabriel Telesco, José Ignacio Orlando

TL;DR: 该研究在Noisy Student半监督训练中按视盘和视杯分别估计伪标签可靠性并加权监督，在已知及未见眼底域上改善分割，视杯收益更明显。

Details

Motivation: 视盘和视杯分割是自动青光眼评估的重要步骤，但有限标注数据训练的模型容易受域偏移影响。教师—学生半监督学习可使用未标注图像，不过标准Noisy Student把所有伪标签视为同等可靠。

Result: 作者用多任务Dice预测器根据教师生成的掩码估计视盘及视杯各自伪标签质量，并以这些估计加权学生训练中的伪标签损失。该设计保留低分前景伪标签对基线的贡献，同时对预测更可靠的样本给予更大相对权重。在已知域与未见域实验中，方法相较Noisy Student基线有所提升，尤其改善视杯分割，并保持有竞争力的泛化表现。

Insight: 按结构分别估计伪标签质量，可针对视杯和视盘标注难度不同来调整半监督信号。摘要没有给出数据集、具体分割指标或可靠性校准精度，因此收益大小及跨设备稳定性无法量化。

Abstract: Optic disc (OD) and optic cup (OC) segmentation is central to automated glaucoma assessment, but models trained on limited labeled fundus datasets often degrade under domain shift. Semi-supervised Teacher-Student learning can improve generalization by leveraging unlabeled data, yet standard Noisy Student training treats all pseudo-labels as equally reliable. In this paper we investigate a targeted extension of Noisy Student training in which OD- and OC-specific pseudo-label supervision is modulated using a mask-derived reliability estimate. A multitask Dice predictor estimates OD- and OC-specific pseudo-label quality from Teacher-generated masks, and these estimates are used to weight the pseudo-label loss during Student training. This preserves the baseline contribution of low-scoring foreground pseudo-labels while assigning greater relative emphasis to pseudo-labels predicted to be more reliable. Experiments across known and unseen domains show improvements over the Noisy Student baseline, particularly for OC segmentation, while maintaining competitive generalization performance.

[78] Your Self-Supervised Projection Head Captures Object Co-Occurrence Statistics cs.CV | cs.LGPDF

Arthur Aubret, Jochen Triesch, Céline Teulière

TL;DR: 该研究发现自监督视觉模型的骨干偏向表示物体类别，而投影头更擅长表示类别共现关系；预训练时控制共现可加速类别级表征形成。

Details

Motivation: 自监督学习能获得强语义视觉表示，但常见代理任务为何有效仍不充分清楚。作者假设，一些自监督目标会隐式学习物体共同出现统计，这可能解释表示中的语义组织。

Result: 研究从现有视觉基准用先进分割模型整理了三个分割物体数据集，并分析多个自监督模型。结果显示语义信息分层编码：视觉骨干主要表征物体类别，投影头则专门编码类别共现关系，因此投影头与人类类别间相似性判断的对齐优于骨干。控制预训练中的共现模式还显示，编码共现可显著加速类别级表征出现。

Insight: 自监督表示的投影头并非仅用于训练技巧，可能承载类别间共现结构；明确利用这种结构可帮助设计更有效的代理任务。摘要未给出模型清单、加速幅度和共现控制细节，结论限于作者测试的模型与数据。

Abstract: Self-Supervised Learning (SSL) has achieved impressive success in learning semantic visual representations, yet the underlying principles driving this success remain underexplored. In this work, we hypothesize that common SSL pretext tasks implicitly model object co-occurrence statistics, a fundamental cue for visual learning. To test this, we curate three datasets of segmented objects from existing vision benchmarks using a state-of-the-art segmentation model. Through experiments across many SSL models, we reveal a hierarchical encoding of semantic information: while the visual backbone captures object categories, the projection head specializes in encoding co-occurrences between object categories. As a result, the projection head outperforms the backbone in aligning with human judgments of inter-category similarity. Furthermore, by controlling co-occurrence patterns during pre-training, we demonstrate that encoding object co-occurrences can significantly accelerate the emergence of category-level representations. Our findings uncover a previously hidden learning principle in SSL and suggest a path toward designing more effective pretext tasks by explicitly leveraging object co-occurrence structure.

Francesco Colasurdo, Ivan Slobozhan, Joaquin Vanschoren, Anna Vettoruzzo

TL;DR: 该研究考察Janus-Pro-7B的LoRA多模态后训练遗忘，发现单次适配看似安全，但连续串联多个适配会破坏先前能力并显著损害通用文本表现。

Details

Motivation: 统一多模态模型共享一套权重处理图像问答和图像生成，因此针对一个模态的微调可能影响另一模态的能力；单次适配测试可能无法揭示连续多次后训练累积的跨模态遗忘。

Result: 作者为Janus-Pro-7B选择基础模型表现较弱的目标（从诗歌文体到低资源威尔士语），并在一次LoRA适配后以及串联三次适配后评测未训练能力。单次运行在四个公开基准和自由形式英语上没有观察到未训练能力退化；即使留出威尔士语困惑度降低76%，相对权重变化小于2%。但这种安全性不代表后训练配方可连续复用：串联三次运行会摧毁较早适配学到的能力，并使一般文本困惑度升至未适配模型的十至百倍。作者将原因归于目标函数只约束当前训练模态、未约束另一输出头；将学习率降至五分之一仍不能避免崩溃，锚定另一输出头可避免未训练能力受损且不牺牲新学习内容。威尔士语文本能力迁移到图像生成仅弥合英威差距的3%。

Insight: 单次LoRA测试不足以评估多模态持续后训练的跨模态遗忘，连续任务应纳入验证；限制另一模态输出头可能是保护方法。结论基于Janus-Pro-7B和三次串联设置，摘要明确显示跨模态迁移也有限。

Abstract: Unified multimodal models answer questions about images and also generate images with one shared set of weights. This creates a post-training risk that single-purpose models do not have: fine-tuning one capability can damage another that works in a different modality. We measure this cross-modal forgetting for LoRA post-training of Janus-Pro-7B. Every target is chosen by how poorly the base model already handles it, from a poetic register to a language it barely speaks, so that retention is tested against a real intervention; we then evaluate the capabilities we never trained, first after a single run and then after chaining three. A single run looks entirely safe: nothing we did not train degrades on four public benchmarks or in free-form English, even when the adapter cuts held-out Welsh perplexity by 76% with a relative weight change under 2%. That safety belongs to the run, not to the recipe. Chaining three runs destroys the capability learned one run earlier and raises general-text perplexity by one to two orders of magnitude above the never-adapted model, on the very measures that stayed flat before. The cause is in the objective, which constrains only the modality being trained and leaves the other head unconstrained: cutting the learning rate to one fifth does not prevent the collapse, while anchoring the other head removes the damage to the capabilities we never trained, at no cost to what is learned. What is learned also stays inside its own modality, with Welsh acquired from text closing only 3% of the English-Welsh gap in image generation.

[80] Vocabulary-Incomplete Unsupervised Vision-Language Model Adaptation cs.CV | cs.LGPDF

Si Qin, Jiawei Tang, Yaxin Hou, Yan Chen, Yutong Xie

TL;DR: VI-UVA研究类别词表不完整时的无监督视觉语言模型适配，并通过从无标签数据重建潜在标签空间来减轻对已提供类别的偏置。

Details

Motivation: 视觉语言模型可通过图文相似度为无标签图像生成伪标签，从而无监督适配，但现有方法假设适配词表包含所有类别。真实场景的无标签数据可能包含遗漏或超出当前任务范围的类别，导致模型偏向已提供标签并损害未列出类别识别。

Result: 作者提出词表不完整的无监督视觉语言模型适配VI-UVA设定，并设计标签空间重建LSR，从无标签数据恢复潜在类别空间并将其纳入提示调优，扩展适配范围、缓解词表偏置。研究还推导LSR遗漏类别语义差异界，并分析影响该界的关键因素。在八个基准上的大量实验显示，LSR超过既有先进方法。

Insight: 实际无监督适配不应把输入类别词表默认当成封闭完整；从未标注样本恢复潜在标签可缓解遗漏类偏置。摘要未列出八个基准名称、具体指标和理论界数值，因此无法进一步判断对长尾类别的效果。

Abstract: In certain real-world scenarios, obtaining labeled data is excessively expensive, motivating increasing interest in adapting models without annotations. Vision-language models (VLMs) provide pseudo-labels for unlabeled images through image-text similarity, enabling adaptation without manual labeling. The previous unsupervised vision-language model adaption methods assume access to an adaptation vocabulary containing all class names. However, a complete vocabulary is difficult to construct in real-world scenarios, as unlabeled data may contain classes that are overlooked or fall outside the current task scope. Thus, we propose Vocabulary-Incomplete Unsupervised Vision-language model Adaptation (VI-UVA), where only a subset of class names are available during adaptation. The vocabulary incompleteness induces an adaptation bias toward provided classes, which degrades omitted-class recognition. To address this issue, we propose Label Space Reconstruction (LSR), which reconstructs a latent label space from the unlabeled data and incorporates it into prompt tuning, thereby extending adaptation beyond the incomplete vocabulary and mitigating the adaptation bias. Furthermore, we derive an omitted-class semantic discrepancy bound for LSR and identify the key factors affecting the bound. Extensive experiments across eight benchmarks demonstrate that LSR outperforms existing state-of-the-art methods.

[81] Unbalanced Optimal Transport for Efficient Visual Document Retrieval cs.CVPDF

Hoyeon Shin, Jeongyeon Kim, Yeong Jun Koh, Yeoneung Kim, Hanul Kim

TL;DR: 该方法将视觉文档检索token压缩表述为非平衡最优传输，统一平衡token重要性、语义冗余和覆盖能力。

Details

Motivation: 视觉语言模型在无OCR文档检索中用多向量表示保留细粒度信息，但会生成大量token，限制检索规模。现有压缩方法常依赖相互独立的启发式，难以同时考虑重要性、语义重复和覆盖。

Result: 作者提出基于非平衡最优传输的视觉token压缩框架，通过交替优化在一个流程中协调上述三项目标，并理论证明最小化传输成本可上界MaxSim检索误差，从而支持排序稳定性。在ViDoRe V1和V3上，方法保持细粒度文档语义，并改善相较先进基线的效率—效果折衷。

Insight: 将token压缩写为带理论检索误差关联的传输优化，旨在避免只按单一启发式丢弃文档细节。摘要未提供压缩率、检索指标具体数值和不同文档类型结果，效率收益无法量化。

Abstract: Vision-Language Models(VLMs)have advanced Visual Document Retrieval (VDR) by enabling OCR-free, multi-vector representations. However, these models generate excessive tokens to preserve dense visual details, which severely limits retrieval scalability. Existing compression methods mitigate this overhead but rely on isolated heuristics that fail to balance token importance, semantic redundancy, and coverage capacity. We propose a visual token compression framework formulated as an Unbalanced Optimal Transport (UOT) problem. Our approach unifies these three objectives into a single alternating optimization process. Furthermore, we theoretically show that minimizing the transport cost upper-bounds the MaxSim retrieval error, ensuring ranking stability. Extensive experiments on the ViDoRe V1 and V3 benchmarks demonstrate that our method preserves fine-grained document semantics and improves the efficiency-effectiveness trade-off over state-of-the-art base-lines. Code is available at https://github.com/shhhhhyy/Unbalanced-Optimal-Transport-for-EVDR.

[82] Trimodal Context-Aware Micro-Expression Recognition in Real-World Conversations: A Dataset and a Hierarchical Fusion Network cs.CV | cs.MMPDF

Junbo Wang, Yan Zhao, Shigang Wang, Jian Wei, Qiming Zhang

TL;DR: TriCoME将真实对话中的微表情与同期话语及前序语境配对，并以层级融合网络利用视觉、音频和文本上下文识别微表情。

Details

Motivation: 微表情短暂且细微，其含义和出现强度会受到谈话语境影响；只分析孤立视频片段可能忽略引发表情的对话语义及声音线索。

Result: 作者构建TriCoME三模态数据集，包含真实对话中的视觉、音频和文本数据，每个微表情都配有同期话语及此前上下文。语境按对话角色标注为来自他人话语的外部刺激和来自主体自身话语的内部背景。配套HFFNet细粒度对齐声学与语义流、压缩长对话序列，以语境类型嵌入调节注意力，并按视觉置信度在视觉单模态和多模态判断间路由，避免噪声语境损害识别。

Insight: 微表情识别需要建模表情发生的对话角色和前序语境，而非只看单帧或短片段。摘要未提供数据规模、基准结果和相对基线提升，HFFNet效果尚无法量化。

Abstract: Recognizing micro-expressions in real-world conversations is challenging, as these facial movements are brief and faint, and their interpretations are strongly dependent on the contextual circumstances of their occurrence. Most existing methods are de-contextualized, as they analyze isolated video clips while disregarding the semantic and acoustic cues in the surrounding dialogue that help explain the emergence of an expression. Decades of psychological studies have shown that social context exerts influences on facial behavior over two key dimensions: how an expression is interpreted and how strongly it is generated to begin with. However, existing micro-expression research rarely incorporates such contextual factors into its models. As an example, provocative utterances from an interlocutor may elicit suppressed negative affective reactions, which isolated silent clips are unable to account for. We propose TriCoME, a trimodal dataset consisting of visual, audio and text data sourced from real-world conversations. In this dataset, each micro-expression is paired with its concurrent utterance and preceding context. We further annotate the context based on conversational role to denote two categories: external stimuli from an interlocutor’s utterance and internal background from the subject’s own utterance. On this basis, we propose the Hierarchical Focusing Fusion Network (HFFNet), which aligns acoustic and semantic streams at fine granularity, compresses long dialogue sequences, modulates attention toward conversational roles via a context-type embedding, and routes each sample between visual-only and multimodal decisions based on visual confidence to prevent noisy context from degrading recognition performance.

[83] VERDICT: Visual Extraction-Result Decoder with Integrated Confidence and Traceable Evidence cs.CV | cs.LGPDF

Harikrishnan P M, Goutham Vignesh, Siva Prasad, Zishan Ahmad, Mahesh Basveshwar Deshmukhe

TL;DR: VERDICT为固定文档抽取器增加视觉证据定位及值和证据置信度校准，不改动原抽取器，并支持按置信度选择性接受字段。

Details

Motivation: 企业使用模式化字段抽取时，模型输出准确率不断提升，但字段来源的文档位置和对字段及证据应信任多少往往缺少明确表示，不利于审查及风险控制。

Result: VERDICT是可接入现有抽取器的插件框架，用小型视觉语言模型对每份文档中的抽取结果定位证据，再分别校准字段值置信度和证据置信度。微调定位VLM显著改善证据质量，并扩展选择性预测可用范围：置信阈值0.80时，微调配置在分布内和分布外数据上均能为值与证据提供可用的精确率—覆盖率折衷，冻结配置则常几乎无覆盖。使用开放权重VLM的内部状态可供校准器使用，比输出log概率更有价值：分布内字段值AUROC从0.730升至0.914，证据AUROC从0.681升至0.906；该优势遇到分布偏移后不再保留。作者显示信号是偏移而非消失，目标数据集约标注120个字段后，证据置信表现分别从0.641升至0.905、从0.675升至0.829。

Insight: 证据定位和字段值置信度应分开校准；开放模型内部状态可提升分布内置信估计，但域偏移后需要少量目标标注再校准。摘要未列出覆盖率具体值和抽取器类型，结果限于指定文档数据及阈值。

Abstract: Enterprises run schema-guided structured field extraction on documents with frontier multimodal models, and accuracy keeps improving as new models and systems arrive. Two things have not kept pace: the evidence for each extracted value, meaning the location of the document it was read from, and how far to trust the value and that evidence. Neither is a component of most extraction systems today, so we build them as a framework that attaches to an existing extractor and leaves it untouched. We present VERDICT, a plug-in framework that uses a small Vision–Language Model (VLM) to ground/localize the evidence once per document over a fixed external extractor’s output, then calibrates value and evidence confidence. Fine-tuning the VLM for evidence localization substantially improves visual evidence quality and, more importantly, expands VERDICT’s selective-prediction operating range. At a strict confidence threshold $0.80$, VERDICT with the fine-tuned VLM gives usable precision–coverage trade-offs for both value and evidence confidence across the in-distribution and out of distribution datasets, whereas the frozen setting often collapses to little or no coverage. Running an open-weight VLM for localization rather than a closed API makes its internal states available to the calibrator, and they are worth more than output log-probabilities: in-distribution AUROC rises from $0.730$ to $0.914$ for value confidence and from $0.681$ to $0.906$ for evidence confidence. This advantage does not survive a distribution shift. We show that the signal is displaced rather than destroyed, and that labeling roughly $120$ fields from the target dataset takes evidence confidence from $0.641$ to $0.905$ on one dataset and from $0.675$ to $0.829$ on the other.

[84] CrowdH-Edit: Reference-Guided Instruction-Based Synthetic Hajj Crowd Image Generation for Crowd Counting cs.CV | cs.AIPDF

Amirah F. Alshammari, Bander A. Alzahrani, Nahed A. Alowidi

TL;DR: CrowdH-Edit用真实朝觐图像作结构参考、文字指令控制人群和光照变化，构建合成朝觐人群图像及标注子集用于计数训练。

Details

Motivation: 朝觐人群视觉分析缺少专门标注数据；通用计数数据和不受约束生成模型未能覆盖圣地几何、白衣朝觐者及室内外照明等领域特征。

Result: CrowdH-Edit采用参考引导、指令式生成框架：真实朝觐照片提供结构参考，文本控制人群状况和光照，并尽量保持场景布局。作者比较了经LoRA适配的Stable Diffusion 2.1、Stable Diffusion XL 1.0，以及ComfyUI中未经额外训练的Qwen Image Edit，后者获得最高语义对齐分数及最低分布和感知距离，因而用于最终数据构建。流程整理出6,122张合成图像，其中387张经挑选和标注形成CrowdH-Edit-Ann-387。将标注合成子集加入真实训练数据后，五种人群计数模型中三种MAE下降、两种RMSE下降。

Insight: 参考图像约束和指令编辑可生成更符合朝觐现场结构的合成训练样本，但下游收益依模型及误差指标而异。摘要未给出具体MAE/RMSE改善值和合成图像人工筛选标准，不能认为所有模型均受益。

Abstract: The scarcity of domain-specific annotated images remains a major limitation for developing computer-vision systems for Hajj crowd analysis. Existing crowd-counting datasets and unconstrained generative models do not capture Hajj-specific visual conditions, including sacred-site geometry, white-clothed pilgrim crowds, and indoor/outdoor illumination variation. This study presents CrowdH-Edit, a reference-guided, instruction-based framework for synthetic Hajj crowd-image generation. Real Hajj images provide structural references, while textual instructions guide changes in crowd conditions and illumination and help preserve the main scene layout. Stable Diffusion 2.1 and Stable Diffusion XL 1.0 with low-rank adaptation were evaluated as comparative baselines. Qwen Image Edit was used without additional training within a ComfyUI workflow and selected for final dataset construction. The workflow produced CrowdH-Edit, a cleaned pool of 6,122 synthetic images, from which 387 images were selected and annotated to construct CrowdH-Edit-Ann-387. Among the evaluated generation pipelines, Qwen Image Edit achieved the highest semantic-alignment score and the lowest distributional and perceptual distances. Downstream evaluation across five crowd-counting models showed that adding the annotated synthetic subset to real training data reduced mean absolute error for three models and root mean squared error for two. These findings indicate that reference-guided synthetic Hajj imagery can support crowd-counting training, although its benefits depend on the model and evaluation metric.

[85] Functionalizing AI-Generated 3D Shapes via Interior Geometry Construction cs.CVPDF

Haoyu Tang, Fancheng Gao, Kan Chen, Peng Song

TL;DR: 该方法为AI生成三维模型构造功能性内部几何，通过部件图表示结构并将各部件内部建模为广义圆柱，使外观保持不变同时支持实际用途。

Details

Motivation: 图像或文本生成的三维形状通常着重外观，内部几何可能不完整或不合理，因此不一定能用于仿真、制造等下游任务。

Result: 作者提出从部件功能推断内部几何的计算方法，利用部件的功能角色决定其内部结构；该结构通常平滑、简单且与外表面相关。系统先用视觉语言模型进行形状分割和图属性推断，构建内部感知部件图；再将每个部件的内部建模为广义圆柱，通过外部几何优化估计其参数。生成结果增加功能有效的内部结构，同时保留AI生成外观。作者展示了容器、储物家具等多种形状，并制造部分样例进行功能验证。

Insight: 将部件功能结构显式加入生成后处理，可使外观生成模型输出的形状更适于实际使用。摘要未报告具体功能测试、几何失败率或适用部件类别，因此制造验证范围有限。

Abstract: AI-based 3D generation models such as Hunyuan 3D and Meshy can produce visually appealing 3D shapes from an image or text prompt.However, they typically focus on an object’s external appearance, often producing incomplete or implausible interior geometry. Consequently, the generated shapes are not guaranteed to be functionally valid, limiting their direct use in downstream applications such as simulation and fabrication. In this paper, we present a computational approach to functionalize AI-generated 3D shapes through interior geometry construction. Our key observations are that a part’s functional role defines its interior geometry, and such geometry is often smooth, relatively simple, and closely related to the part’s external shape. Inspired by these observations, we introduce an interior-aware part graph to encode information necessary for constructing interior geometry within an AI-generated 3D shape, and build the graph via VLM-based shape segmentation and graph attribute inference. Guided by the graph, we model the interior geometry of each part as a generalized cylinder and infer its parameters from the part’s external geometry through an optimization. The resulting 3D shapes are functionally valid while preserving their AI-generated external appearance. We demonstrate our approach on a variety of AI-generated 3D shapes, including vessels, containers, and storage furniture, and fabricate selected examples for functional validation.

[86] Same Space or Not: Benchmarking Spatial Intelligence Through Cross-View Contradictions cs.CV | cs.LGPDF

Huizi Cao, Guanyu Xu, Zhixuan Liang, Ming Yin, Jiuxiang Gu

TL;DR: SSoN是跨视角矛盾识别基准，要求模型从同一场景多视图中找出一个经过3D编辑的不一致视图并解释矛盾，当前视觉语言模型明显弱于人类。

Details

Motivation: 空间关系或成对比较基准难判断模型能否跨多个视角维持统一场景表征，因为它们未必能识别哪张视图偏离真实场景。

Result: SSoN包含来自七种室内场景来源的1,000道问题，覆盖六类受控三维编辑。每题给出同一场景的三视图和一个编辑变体视图，模型需在无可信参考图、无标记目标的条件下找出矛盾视图并说明理由。构造流程将人工审查迭代转成生成约束，并通过自动检查和全面人工核验排除简单视觉捷径及证据不足的问题。人类准确率为83.3%，最强视觉语言模型为39.7%；要求同时提供有效视觉解释时降至34.9%。在100题干预子集上，提示和3D重建通常仍低于50%准确率；提供真实相机信息时答题准确率达61%，但真正求解率仅48%。Qwen2.5-VL-7B在201题留出集上从21.89%基线升至仅答案微调后的33.33%，使用理由与答案联合微调后为41.79%。

Insight: 跨视角矛盾判断是现有VLM空间智能短板，定向监督可部分改善但仍有明显差距。答对选择题不代表能给出有效解释；提示及相机信息干预结果不能等同于完整场景理解。

Abstract: Spatial intelligence, the capacity to represent and reason about physical space, underlies embodied perception and real-world visual understanding. Existing benchmarks probe it through spatial relations, perspective-taking, and mental modeling. What remains underexamined is whether a model can maintain a coherent scene representation across views, since pairwise comparison alone cannot reveal which view departs from the truth. We thus introduce Same Space or Not (SSoN), a benchmark of cross-view contradiction resolution comprising 1,000 questions from seven indoor scene sources and six categories of controlled 3D edits. Each question presents three views of an unchanged scene and one view of an edited variant. Models must identify the inconsistent view and explain the contradiction without a designated trusted reference or a marked target. Our pipeline iteratively translates human review into generation constraints. Automated checks and exhaustive human validation screen out both trivial visual shortcuts and questions lacking sufficient evidence for a unique answer. Humans achieve 83.3% accuracy, compared with 39.7% for the strongest vision-language models (VLMs), falling to 34.9% when a valid visual explanation is also required. We investigate prompting, camera information, and 3D reconstruction as interventions. On a 100-question intervention set with a 24% baseline, prompting and 3D reconstruction remain below 50% answer accuracy, except when ground-truth camera information is supplied, which reaches 61% answer accuracy but only 48% True solve accuracy. On a held-out 201-question test set, Qwen2.5-VL-7B achieves 21.89% accuracy, 33.33% after answer-only fine-tuning, and 41.79% after fine-tuning on both rationales and answers. These results identify cross-view contradiction resolution as a substantial weakness of current VLMs that can be partly addressed by targeted supervision.

[87] Geo-vision: Native-Resolution Geometry and Sampled Topology for Vehicle-Part Damage Classification cs.CV | cs.LGPDF

Usef Faghihi, Amir Saki, Serge Jonathan Kogbeti-Wongo

TL;DR: Geo-vision将外观、多尺度几何、方向和数字拓扑结合用于车辆部件损伤分类，在相同裁剪图像上超过冻结ImageNet预训练YOLO11n基线，但CPU延迟较高。

Details

Motivation: 车辆部件损伤分类可受益于几何和拓扑信息，单靠外观特征可能忽略连通区域、孔洞、部件相对位置等结构。模型同时需控制外部预训练与标注数据量比较。

Result: Geo-vision使用固定表征组合外观、多尺度几何、方向和采样数字拓扑特征，包括连通分量、孔洞、欧拉特征和空间布局，并通过监督树分类器预测。作者建立振幅保留、轴向符号不变性和条件算子稳定性。对4,916个源组中的55,303个裁剪块进行回顾性检查，平均精确率为0.7577，高于相同裁剪上冻结ImageNet预训练YOLO11n的0.7458。加入拓扑、占用和分量位置结构模块后，AP在外观及几何基础上再增加0.00493；各匹配标注预算下AP点估计更高，约少25%任务标签时点估计仍高于全数据YOLO，但调整后的区间包含零。完整串行CPU推理每裁剪块26.98毫秒，YOLO为11.20毫秒；峰值内存分别185和437MiB。

Insight: 拓扑和空间结构补充可提升损伤分类点估计，但样本组和标注预算比较中有些差异不确定；结构模型推理较慢，外部独立验证仍必要。

Abstract: Geo-vision classifies vehicle-part damage using a fixed representation combining appearance, multiscale geometry, orientation, and sampled digital topology. Connected components, holes, Euler characteristics, and spatial arrangement complement amplitude-sensitive pooling and a supervised tree classifier. We establish amplitude preservation, axial sign invariance, and conditional operator stability. Retrospective checkpoint evaluation on 55,303 crops from 4,916 source groups yields average precision 0.7577, exceeding 0.7458 for an ImageNet-pretrained frozen YOLO11n comparator on identical crops. The complete added structural block, comprising topology, occupancy, and component location, improves average precision by 0.00493 beyond appearance and geometry. Geo-vision uses no external encoder pretraining and has higher AP point estimates at every matched label budget. With approximately 25% fewer task labels, its AP point estimate exceeds full-data YOLO, although the adjusted interval includes zero. Complete serial CPU inference requires 26.98 ms per crop versus YOLO’s 11.20 ms, with peak memory 185 versus 437 MiB. Independent external validation remains necessary.

[88] IMPULSE: Push First, Then Predict in V-JEPA 2 Feature Space cs.CV | cs.LG | cs.ROPDF

Tingan Jin

TL;DR: IMPULSE让冻结V-JEPA 2特征预测器通过少量接触动作校准物体动力学，在未见物体上提高短期特征预测精度。

Details

Motivation: 潜在世界模型可预测视频编码器特征，但只观察物体一次并不能知道它的旋转难度、弹性或滑动程度；人类会通过少量推推来了解物体动力学。

Result: IMPULSE为冻结的V-JEPA 2特征空间预测器加入贝叶斯接触记忆，总结三次短校准推动，并通过预测物体姿态把经验传入预测器；轻量随机傅里叶姿态头使预测器利用姿态信息。在事先封存模型、数据之后新采集的512个物体上，记忆相较无经验预测器将1秒特征预测误差降低11.6%–12.8%，覆盖全部8个训练种子，种子与物体联合95%区间不含零。128个物理参数超出训练范围的物体上仍降低11.7%–14.0%；完整扩展比普通预测器提升15.5%–16.6%。傅里叶头让完美记忆相对普通姿态token的收益约增至2.8–3.2倍；原始推动测量经同一路径也有类似收益，说明贝叶斯摘要不是必需，主要压缩为8个可解释数值。接触刚发生时尚无收益，之后才逐步显现。

Insight: 预测物体动力学需要主动接触经验，模型可将少量校准转化为世界模型潜在预测改进。结果以特征预测误差为核心，不直接证明下游机器人任务收益；摘要显示贝叶斯摘要主要提供压缩和可解释性，而非必要信息。

Abstract: Latent world models such as V-JEPA 2 forecast in the feature space of a video encoder, but one look at an object does not reveal how hard it is to rotate, how much it bounces or how much it slides. People find out by trying: a few pokes reveal how an object responds. IMPULSE, named after the impulse response it infers, gives a predictor in frozen V-JEPA 2 feature space the same ability. A Bayesian contact memory summarizes three short calibration pushes, a predicted object pose carries it into the predictor, and a lightweight random-Fourier pose head lets the predictor use the pose. On 512 fresh objects, in a pre-specified held-out confirmation with models sealed before the data existed, the memory lowers the 1-second feature-forecast error by 11.6-12.8% relative to the same predictor without experience, for all 8 training seeds, with a 95% seed-and-object interval excluding zero. The gain also holds for 128 objects with out-of-range physics (11.7-14.0%), and the full extension beats a plain predictor by 15.5-16.6%. An information-ladder argument, exact for ideal predictors, shows that physics reaches the features only through motion, so we measure how much the pathway lets pose help. On fresh objects, the Fourier head roughly triples the gain of a perfect memory over plain pose tokens (2.8-3.2x). Raw push measurements fed through the same pathway give a similar gain, so the Bayesian summary is not required; it condenses the trials into eight interpretable numbers. The gain is absent at contact onset and grows afterwards.

[89] Learning When to Trust and Reobserve Sparse-Frame VideoQA from Temporal Sampling Perturbations cs.CVPDF

Shujia Liu

TL;DR: PhaseGuard+分别估计稀疏帧视频问答的回答风险和额外观察的修复价值，利用时间采样扰动更可靠地决定何时加密帧或调用更强模型。

Details

Motivation: 稀疏采样视频模型可能仅因采样时间偏移就改变答案。答案不一致虽能暴露敏感性，却不一定比token置信度更能排序错误，也不能说明追加哪种观察会纠正答案。

Result: PhaseGuard+在受控时间偏移下探测冻结VLM，分别预测相位投票答案可靠性及特定补救动作带来的逐样本正确率提升；只有预期受益的样本才路由至更密集帧或更强VLM。轻量预测器用公开数据训练，VLM保持冻结。方法开发使用Perception Test 256段视频，测试为与其分离的256段视频、554道Physics问题。Qwen3-VL-8B的可靠性AUROC从token softmax的68.0%提高到74.8%，增6.8点（95% CI [2.6,10.7]）。将30%问题升级至Qwen3-VL-32B后准确率67.87%，预算匹配随机路由为63.66%，高4.21个百分点（CI [1.89,6.40]）；四种补救组合均显著优于随机路由。

Insight: 时间敏感性只有与可靠性和补救动作价值同时估计，才能转成有效推理资源分配。结果来自单独的Physics测试集和指定升级比例，摘要未给出不同问题类型的分项收益。

Abstract: Sparse-frame video-language models can change their answers when only the sampling timestamps shift. Raw agreement exposes this problem, but it need not rank errors better than token confidence or identify which additional observation will repair an answer. We argue that answer risk and action value must be estimated separately. PhaseGuard+ probes a frozen VLM at controlled temporal offsets, predicts the reliability of the phase-vote answer, and predicts the example-wise correctness gain of a specific rescue action. The second score routes only examples expected to benefit from denser frames or a stronger VLM. Both predictors use lightweight models trained on public data; the VLMs remain frozen. We use 256 Perception Test videos for method development and evaluate on a disjoint 256-video test set comprising 554 Physics questions. For Qwen3-VL-8B, reliability AUROC rises from 68.0% with token softmax to 74.8% (+6.8 pp, 95% CI [2.6, 10.7] pp). Routing 30% of questions to Qwen3-VL-32B reaches 67.87% accuracy, compared with 63.66% for budget-matched random routing (+4.21 percentage points, CI [1.89, 6.40]). All four tested rescue pairs significantly outperform random routing. Temporal sensitivity becomes actionable when the system learns both whether an answer is unreliable and whether an available action is likely to fix it.

[90] How Much Can We Prune? Exploring the Performance Limits of Visual Token Pruning cs.CV | cs.AIPDF

Janghoon Cho, Shubhankar Borse, Reza Pourreza, Juntae Lee, Jihwan Bang

TL;DR: 该研究将视觉token剪枝视为固定预算下选择共享token子集的问题，用辅助问答评估其效用并进行贪心优化，发现相较现有剪枝基线仍有提升空间。

Details

Motivation: 视觉token剪枝可降低视觉语言模型的显存和推理成本，但已有研究多设计特定选择器并在少量预算点报告增益；尚不清楚表现受理论剪枝上限还是当前选择算法所限。

Result: 作者针对一般图像理解，在固定预算内选择一个可支持该图像多种问题的共享视觉token子集，以前沿VLM生成的有限辅助QA集近似效用；冻结被测模型，优化teacher-forced答案概率目标。基于视觉证据增多时效用单调且边际收益递减的设想，研究贪心构造嵌套子集，并提出QA小批量贪心QA-MBG，每步随机抽取部分QA估计token边际收益。实验中，QA-MBG比完整贪心更便宜且有效。每张图需调用一次前沿VLM并使用约0.5 H100 GPU小时，可生成保留率最高0.3的嵌套路径。在900张、九个基准图像上，Qwen3-VL-4B和Gemma 4 E4B的结果显示，留出QA及图像依据长文生成均较现有剪枝基线有显著提升，且显著缩小与全token推理的差距。

Insight: 离线搜索虽成本较高，但可估计token子集在广泛问题分布上的潜在剪枝上限，为在线轻量剪枝算法提供参照。结果使用前沿模型生成辅助问题且每图约0.5 GPU小时，部署成本和辅助问题分布是重要限制。

Abstract: Visual token pruning is increasingly used to reduce the memory and inference costs of vision-language models. Existing work has largely focused on designing method-specific selectors and demonstrating incremental improvements over prior methods at selected token budgets. It remains unclear whether the resulting performance reflects an inherent limit of visual token pruning or the limitations of current token-selection algorithms. We take a complementary perspective and ask: \emph{how well can visual token pruning perform when offline search cost is not the primary constraint?} We formulate pruning for general image understanding as selecting, under a fixed budget, one shared visual-token subset that supports a broad distribution of questions about an image. We approximate this utility with a finite auxiliary QA set generated by a frontier VLM and optimize a teacher-forced answer-probability objective with the evaluated model frozen. Motivated by the conceptual monotonicity and diminishing returns of adding visual evidence, we construct nested subsets greedily and introduce \emph{QA mini-batch greedy} (QA-MBG), which estimates token marginal gains from randomly sampled QA components at each step. Empirically, QA-MBG is both cheaper and more effective than full greedy. With one frontier-VLM call and approximately $0.5$ H100 GPU-hours per image, it produces a nested subset trajectory covering keep ratios up to $0.3$. Across Qwen3-VL-4B and Gemma~4 E4B on $900$ images from nine benchmarks, we find significant headroom over evaluated pruning baselines on held-out QA and image-grounded long-form generation, while the searched subsets substantially narrow the gap to full-token inference.

[91] FrameScout: Scouting Query-Relevant Frames for Long Video Understanding cs.CV | cs.LGPDF

Haonan Hu, Shuhao Chen, Weisen Jiang, Lizhao Gao, James Kwok

TL;DR: FrameScout用流式滑窗KV缓存和后继帧聚合生成带时间上下文的帧表征，并以帧—查询对比目标挑选长视频中与问题相关的帧。

Details

Motivation: 长视频可能持续数十分钟至数小时，产生的视觉token远超当前VideoLLM处理能力。现有关键帧选择器要么逐帧编码、忽略时间上下文，要么分块处理并采用难以优化的排序目标。

Result: FrameScout包含流式选择器，利用滑动窗口KV缓存和后继聚合获得带时间上下文的帧嵌入；同时用帧—查询对比目标将两者对齐至共享空间，直接训练选择器区分相关帧和无关帧。在三个长视频基准及六个基础VideoLLM上的实验中，作者报告FrameScout持续达到先进表现。

Insight: 流式时间上下文与直接的查询相关性学习可在控制视觉token数量时保留重要帧。摘要未提供基准名称、token节省量和具体分数，因此无法量化性能或计算收益。

Abstract: Long video understanding requires VideoLLMs to answer queries about videos spanning tens of minutes to hours, but such long videos produce far more visual tokens than current VideoLLMs can process. To address this issue, keyframe selectors are proposed to select the most query-relevant frames by computing the similarity between frame and query embeddings. Existing selectors either encode frames independently without temporal context or process chunks in isolation with ranking objectives that are difficult to optimize. To alleviate those limitations, we propose FrameScout with two core designs: (i) a streaming selector architecture that leverages a sliding-window KV cache and successor aggregation to produce temporally-contextualized frame embeddings; (ii) a frame-query contrastive objective that aligns frame and query embeddings in a shared space, directly training the selector to distinguish relevant frames from irrelevant ones. Extensive experiments on three long-video benchmarks and six base VideoLLMs demonstrate that FrameScout consistently achieves state-of-the-art performance. Code is available at https://github.com/Hubert-1102/FrameScout.

[92] SPELUNKA: A Framework for Finding Unanticipated Associations in Vision-Language Models cs.CV | cs.LGPDF

Cassandra Parent, Isha Agarwal, Kumail Alhamoud, Charikleia Moraitaki, Rafiya Javed

TL;DR: SPELUNKA在不预先标注子群体、不需推理期人工监督或微调的条件下，自动发现视觉语言检索模型中未预料到的分布偏差。

Details

Motivation: 视觉语言模型常从训练数据中继承虚假或有害关联；以往研究多针对已知偏差进行消除，针对检索任务自动发现未知关联的方法较少。

Result: 给定一个关注概念，SPELUNKA为相关图像生成描述，从描述提取实体并筛选语义关联；随后比较VLM前排检索结果中属性的过度或不足代表程度与完整数据集分布，以识别虚假关联。在四个数据集上的评估显示，方法可找回合成和自然出现的虚假相关；识别与查询强相关图像的平均AUROC为0.93，并超过基线发现与查询高度关联特征。

Insight: 无需事先列出人口或类别子群，基于全数据集属性分布的偏差检测可用于发现未预料关联。摘要没有说明四个数据集、查询概念或偏差类型，AUROC指标对应的是关联图像识别而非所有偏差发现任务。

Abstract: Vision-Language Models (VLMs) are the backbone of many state-of-the-art multimodal systems, but often encode spurious and harmful associations from their training data. While prior work focuses on removing known biases from VLMs, methods for discovering unanticipated associations for retrieval tasks remain underexplored. We introduce SPELUNKA, an automated, inference-time framework for finding unanticipated distributional biases for retrieval tasks in VLMs without pre-specified labeled subgroups, inference-time human supervision, or model fine-tuning. Given a concept of interest, SPELUNKA generates captions for related images, extracts relevant entities from captions, and selects semantically meaningful associations. A frequency-based detection rule then identifies spurious associations by measuring over- or under-representation of these attributes in a VLM’s top-ranked outputs relative to their full dataset distribution. We evaluate SPELUNKA across four datasets, and find that it 1) successfully recovers both synthetic and naturally occurring spurious correlations, 2) achieves an average AUROC of 0.93 across four datasets in identifying images strongly associated with a query, and 3) outperforms baselines in finding features that are strongly associated with a given query.

[93] SituHOI: Scalable Capture of Scene-Situated Full-Body Human-Object Interaction cs.CV | cs.AIPDF

Qifan Wang, Qihang Fang, Boao Shi, Yushi Du, Pei Zhou

TL;DR: SituHOI提供大规模、场景对齐的人—物交互RGB-D视频数据，并提出给定场景、物体及目标终态的联合人体和物体运动预测基准。

Details

Motivation: 真实场景人体—物体交互建模需要同时表示全身动作、操作物体和场景几何，但现有数据集难以兼顾可扩展采集与结构化场景对齐标注。

Result: SituHOI包含18,429段标注视频，覆盖16个室内房间和约86小时通过质量筛选的活动。数据将单台移动RGB-D相机记录与静态场景扫描、物体几何、语言描述以及估计的人体和物体轨迹配对。流水线以场景扫描为持久参考，结合相机定位、基于深度的度量对齐、人体网格恢复、物体跟踪和时间细化，将几何标注统一到地面坐标系。FlexHOI-Bench据此评估目标条件联合运动预测，允许观察及预测时长变化；输入历史、场景和物体几何及目标物体终态，要求预测未来人和物运动。MoCap子集上，流水线将人体G-MPJPE从195.4降至68.3毫米，物体平移误差从43.0降至28.7毫米；SharedHOI在FlexHOI-Bench三项汇总指标上领先四种适配基线。

Insight: 场景扫描作为统一坐标参照，可支持比仅视频动作数据更全面的环境交互预测。摘要未提供各基准的完整预测成绩；数据尚待录用后发布，且自动估计标注的精度需要结合MoCap子集验证理解。

Abstract: Modeling human-object interaction (HOI) in real environments requires jointly representing whole-body motion, manipulated objects, and scene geometry, yet existing datasets trade off scalable capture against structured, scene-aligned supervision. We introduce SituHOI, a large-scale RGB-D dataset containing 18,429 annotated clips across 16 indoor rooms and approximately 86 hours of quality-screened activity. It pairs recordings from a single moving RGB-D camera with static scene scans, object geometry, language descriptions, and estimated human and object trajectories. Using the scans as persistent references, our pipeline combines camera localization and depth-based metric alignment with human mesh recovery, object tracking, and temporal refinement to register all geometric annotations in a shared floor-aligned frame. These data support FlexHOI-Bench, a goal-conditioned joint motion prediction benchmark with variable observation and prediction durations. Given observed histories, scene and object geometry, and a terminal object pose, models predict future human and object motion. We provide SharedHOI as a reference predictor and evaluate world-space accuracy and human-object coordination. On a MoCap subset, our pipeline reduces human G-MPJPE from 195.4 to 68.3 mm and object translation error from 43.0 to 28.7 mm. SharedHOI leads four adapted baselines on all three aggregate FlexHOI-Bench metrics. Data will be released upon acceptance. An anonymized code repository is available at https://anonymous.4open.science/r/iclr2027-anonymous-code-1DFB.

[94] Time-JEPA: Grounding Video World Models with Physical Time Series cs.CVPDF

Shenyi Li, Ziyang Zhang, Wen Wu, Ming Liu, Chao Zhang

TL;DR: Time-JEPA将低维、时间对齐的物理序列作为辅助监督微调视频世界模型，使其更好预测当前状态、未来多时域物理状态和运动学结构。

Details

Motivation: 视频世界模型通过预测缺失或未来内容学习环境变化，但物理状态可为演化预测提供额外信息。依赖领域物理定律会增加多域训练复杂度，单帧或帧对监督又较少直接约束状态随时间如何变化。

Result: Time-JEPA用低维、时间同步的物理时间序列微调视频世界模型，将视频表示与对应时刻物理状态关联，鼓励模型保留状态演化信息。在自动驾驶、机器人操作和人体运动等多个领域，相较匹配的视频微调对照，方法在当前状态解码、多时域未来状态预测和运动学重建上都有改善，且在条件与状态变量分布偏移下仍表现出提升。多域联合训练在三项任务上均比对应单域训练有总体收益。外部评估还显示时间序列监督增强对物理异常的敏感度及对交互的表示，包括未见场景；第二种视频骨干的总体增益表明该方法可迁移至不同骨干。

Insight: 物理序列监督可帮助视频模型学习随时间变化的状态，而不仅是视觉外观或静态结构。摘要未列出数据集、提升幅度和各域具体标签需求，迁移性结论限于两个骨干和所测场景。

Abstract: Video world models learn how visual environments evolve by predicting missing or future content. However, physical-state information can provide useful context for predicting how environments evolve. Prior work incorporates this information through different mechanisms: domain-specific physical laws can complicate multi-domain training, while supervision on individual frames or frame pairs gives limited attention to how physical state evolves over time. We introduce Time-JEPA, which fine-tunes a video world model using low-dimensional, temporally aligned physical time series as auxiliary supervision. Physical time series connect video representations with physical state at corresponding temporal positions, encouraging the model to preserve how physical state evolves. In multiple domains including autonomous driving, robot manipulation, and human motion, Time-JEPA improves current-state decoding, multi-horizon future-state prediction, and kinematic reconstruction over matched video-fine-tuned controls across both condition and state-variable distribution shifts. Multi-domain training provides aggregate gains over matched single-domain training in all three tasks. External evaluations further suggest that time-series supervision improves sensitivity to physical anomalies and the representation of physical interactions, including in unseen scenarios. Aggregate gains on a second video backbone indicate that the method can be applied to more backbones. These results suggest that physical time series help video world models better understand the physical world.

[95] RL Teaches VLMs Where to Attend but Not How to Perceive cs.CV | cs.AIPDF

Ziqian Zhang, Junjie Liao, Zhining Zhang, Wentao Zhu

TL;DR: 该研究发现多轮强化学习提升视觉语言代理基准表现时，主要改变LLM对视觉token的利用和注意力，而视觉编码器表征本身相对稳定。

Details

Motivation: 强化学习可训练视觉语言模型代理获得环境任务成功率提升，但能力变化来源不清楚；需要判断收益是否伴随更好的视觉证据理解，还是由其他策略因素造成。

Result: 作者在四种不同代理环境中检查多轮RL带来的任务成功收益，并设计环境特定的视觉问答来衡量有依据的视觉理解。表征探测显示视觉编码器上的探针准确率相对稳定，而LLM层探针准确率提高；注意力分析发现RL训练期间LLM层更加关注视觉token。为测试关注增加是否促进VQA，作者提出固定注意力偏置的轻量即插即用干预，不更新模型参数也不计算梯度；该干预能在未训练模型上部分恢复视觉问答收益。

Insight: RL带来的视觉任务提升可能源自模型更有效调用已有视觉信息，而非视觉编码器表示变得更强。注意力偏置只能部分恢复收益，摘要未列出具体探针和VQA增幅，四种环境结果也不能覆盖所有VLM代理。

Abstract: Although reinforcement learning (RL) effectively trains Vision-Language Model (VLM) agents, the mechanisms behind these capability gains are poorly understood. Across four diverse agent environments, we investigate whether the task-success gains from multi-turn RL are accompanied by improved grounded visual question answering about the environment. To assess this, we design environment-specific grounded VQA tasks that reveal improvements in grounded visual understanding. Representational probing shows that vision-encoder probe accuracy remains relatively stable, while LLM-layer probe accuracy improves. Attention analysis further reveals that LLM layers pay more attention to visual tokens during RL training. To test whether increased attention to visual tokens contributes to VQA gains, we introduce a lightweight plug-and-play intervention using a fixed attention bias. Without updating model parameters or computing gradients, this intervention partially restores VQA gains on the untrained model. These results reveal better utilization of visual information as a key mechanism behind the gains from RL.

[96] BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References cs.CVPDF

Xinyu Fang, Yuchi Wang, Weihao Yu, Rongyao Fang, Zhaohai Li

TL;DR: BlenderWorld评测多模态智能体依据真实参考创建和编辑三维内容的能力，使用多视角渲染及任务二元规则，而不是只依赖文字说明质量。

Details

Motivation: 实际三维创作需从文字、图像、视频和结构化规格理解物体形状、材质、布局及光照，并严格执行重建或编辑指令；现有基准对真实开放式创作场景覆盖有限。

Result: BlenderWorld包含223项建筑、产品、自然环境和科学可视化任务，输入可为文字、图像、视频或结构化说明。统一评测协议从多个视角渲染提交的Blender项目，并按2,651条任务二元规则及整体视觉质量评分。13个前沿模型的比较中，GPT-6-Astra融合总分最高为79.9，其次为Opus-5的73.0和GPT-5.6-Sol的67.7。分析发现，模型通常能解析宽泛指令，但在Blender中的细粒度三维空间连贯性和结构复现方面仍明显不足。

Insight: 把真实项目渲染和可核验任务规则结合，能评估生成物是否符合三维要求，而不仅看模型文字回复。摘要未说明融合分数计算、模型是否可调用不同工具或评委配置，排名限于13个被测模型和该基准。

Abstract: Recent advances in multimodal large language models (MLLMs) have enabled them to act as autonomous agents, creating and editing 3D models via code or direct software manipulation. In practical workflows, these models must infer object shapes, materials, spatial layouts, and lighting from diverse references while strictly following instructions to reconstruct, modify, or design 3D content. However, existing benchmarks provide limited coverage of such open-ended requests across real-world scenarios. To bridge this gap, we introduce BlenderWorld, the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasks spanning architecture, products, natural environments, and scientific visualization, driven by text, image, video, and structured specification inputs. To rigorously assess these complex deliverables, we establish a unified evaluation protocol: we render the submitted Blender projects from multiple viewpoints and evaluate them against 2,651 task-specific binary rubrics alongside holistic visual quality scores. Benchmarking thirteen frontier models reveals that GPT-6-Astra achieves the highest overall fused score of 79.9, followed by Opus-5 (73.0) and GPT-5.6-Sol (67.7). Ultimately, our findings reveal a critical bottleneck: while current models can parse broad instructions, they still profoundly struggle with fine-grained 3D spatial coherence and accurate structural reproduction in Blender, leaving substantial headroom for future 3D-native agentics.

[97] What Generation Knows: Unlocking Fine-Grained Understanding in Unified Multimodal Models cs.CVPDF

Tengju Ru, Weitong Lian, Zecong Tang, Zhiheng Liu, Lingyi Meng

TL;DR: GAFU将统一多模态模型生成通路中的细粒度视觉表征压缩为G tokens，供理解通路按任务检索，从而改善细粒度视觉理解。

Details

Motivation: 视觉理解需要区分局部属性和空间关系，现有视觉语言模型仍有困难。统一多模态模型的生成通路保留了与理解通路互补的丰富视觉信息，但既有生成辅助方法主要依赖生成监督或显式生成，较少系统利用原生生成侧表征。

Result: 作者分析发现，生成侧表征对细粒度语义变化更敏感，能捕捉理解侧遗漏的任务相关变化，但优势依赖空间组织正确。基于此，GAFU在后训练中把生成侧视觉表征压缩为紧凑G tokens，保留细节并让理解通路按任务检索。多个视觉理解任务和UMM架构上，GAFU相较同数据训练的SFT基线持续更好；VStar-Bench、OmniSpatial和VSR-Bench的最大相对提升分别为20.00%、17.94%和8.62%。

Insight: 生成通路可能携带理解通路缺少的细粒度线索，压缩并供理解侧调用可促进统一模型中生成与理解协同。摘要未给出各架构逐项数值、计算开销和空间错位情况下的退化幅度。

Abstract: Fine-grained visual understanding requires distinguishing subtle differences in local attributes and spatial relations, yet remains challenging for existing vision-language models. Unified multimodal models (UMMs) offer a new opportunity through generation pathways that preserve rich visual information complementary to understanding. However, existing generation-assisted understanding methods mainly rely on generative supervision or explicit generation and have not systematically explored direct use of native generation-side representations. To address this gap, we systematically analyze what generation-side representations offer for fine-grained understanding, finding that they are more sensitive to fine-grained semantic changes, capture task-relevant changes missed by understanding-side representations, and depend on correct spatial organization for this advantage. Guided by these findings, we design GAFU, a post-training framework that unlocks native generation-side information for fine-grained understanding. It compresses generation-side visual representations into compact (G) tokens that preserve fine-grained visual details, enabling task-conditioned retrieval by the understanding pathway. Experiments show that GAFU consistently outperforms SFT baselines trained on the same data across diverse visual understanding tasks and UMM architectures, with maximum relative improvements of 20.00%, 17.94%, and 8.62% on VStar-Bench, OmniSpatial, and VSR-Bench, respectively. Overall, GAFU converts complementary information from the generation pathway into gains in fine-grained visual understanding, providing an effective path toward synergy between generation and understanding in UMMs.

[98] Single-Pass Decision Scores Enable Rapid Classification with Adjustable Thresholds in Medical Multimodal Large Language Models cs.CVPDF

Patrick Wienholt, Frederik Hauke, Christiane Kuhl, Jakob Nikolas Kather, Sven Nebelung

TL;DR: 该研究从单次前向计算中归一化yes/no token概率，生成可调阈值医学分类分数，无需重新训练即可调整临床决策点。

Details

Motivation: 医学多模态大模型回答二元临床问题时通常只输出yes或no，阈值隐式固定，难以根据不同筛查或临床需求调整灵敏度与特异度。

Result: 作者对MedGemma及4B、27B的Gemma-3，在六个医学影像数据集上评估单次前向分数：对yes与no token概率重新归一化，再由使用者调整决策阈值。默认阈值经常不适配任务；例如MedGemma-27B把所有正常胸片都标为异常，验证集选择阈值后灵敏度为0.945、特异度0.696。归一化分数在六个数据集中的四个取得最高AUROC；CheXpert上为0.819，高于原始概率0.777和文本提示得到的概率0.743。单次判断在一张GPU上需53–79毫秒。

Insight: 后处理归一化为医学MLLM增加可调运行点，适合需要按任务权衡漏诊与误报的分类应用。阈值需要验证集选择，特异度示例为0.696；结果不能说明阈值跨医院稳定，也不构成临床验证。

Abstract: Multimodal large language models (MLLMs) answer binary clinical questions with a yes or no token, leaving the decision threshold implicit and fixed. We derive a continuous score from a single forward pass by renormalizing the probabilities of the yes and no tokens, allowing the threshold to be adapted to clinical needs without retraining. We evaluated this approach with MedGemma and Gemma-3 (4B and 27B) on six medical imaging datasets. The default threshold was frequently misaligned with the task: for identifying normal chest radiographs, MedGemma-27B labeled every image as abnormal, whereas a validation-selected threshold yielded a sensitivity of 0.945 and a specificity of 0.696. Normalization achieved the highest AUROC on four of six datasets (CheXpert: 0.819 vs. 0.777 for raw and 0.743 for text-elicited probabilities), and a single decision required 53-79 ms on one GPU. This simple post-hoc method turns medical MLLMs into fast classifiers with adjustable operating points. Code: https://github.com/TruhnLab/llm-threshold-calibration

[99] Automatic Cause Determination in Road Scene Understanding Using Qualitative Reasoning and Four-Valued Logic cs.CVPDF

Nassim Belmecheri, Arnaud Gotlieb, Nadjib Lazaar, Helge Spieker

TL;DR: 该研究用定性时空推理和Belnap四值逻辑，从道路场景关系图中解释自动驾驶动作的复杂原因。

Details

Motivation: 道路视频理解不仅要识别场景参与者，还要解释自车为何停车、减速或转弯。城市交通中的多个触发原因可能互相混淆或矛盾，简单推理难以表达。

Result: 作者提出基于时空推理和Belnap四值逻辑的自动驾驶动作原因判断方法。系统先将道路场景构造成可解释定性图QXG，记录道路实体之间的时空关系，再针对指定场景帧和自车动作从图中推断可能的复杂触发原因。论文在NuScenes数据集抽取的若干场景中展示方法用途。

Insight: 四值逻辑可表示证据支持、反对及矛盾并存的情况，QXG则将原因判断建立在显式场景关系上。摘要没有提供数据规模、准确率和与基线对比，展示案例不足以量化安全性或解释质量。

Abstract: Road scene understanding in automated driving (AD) aims to build a comprehensive analysis of video sequences taken on the road by embedded or fixed cameras (e.g., mounted on vertical road signals). One goal is to identify the relevant actors in the scene and another goal is to determine the causes that have triggered a specific action of the ego car (i.e., stop, slow down, turn left, etc.). In a complex urban environment, these causes can be multiple, confusing, possibly contradictory to other causes and not easily expressible using simplistic reasoning. Still, providing accurate {\it automatic cause determination} supports a) user acceptance by providing appropriate explanations to the car passengers and road users; b) increased road safety by providing detailed road scene understanding to traffic. In this paper, we propose using spatiotemporal reasoning and Belnap’s four-valued logic to formulate complex causes of AD action in a road scene. We compute these causes by analysing a Qualitative eXplainable Graph (QXG), which is an abstract representation of the road scene capturing spatiotemporal relations between road entities. Starting from a QXG, our approach called \tool, is targeted to determine complex causes of a selected AD action occurring in a specific frame of a road scene. The usefulness of \tool is demonstrated on several scenes extracted from the well-known NuScene dataset.

[100] ProV-Mem: Proactive Memory for Streaming Video Question Answering cs.CV | cs.MMPDF

Bingbing Wang, Jing Li, Han Gao, Yishi Lin, Ruifeng Xu

TL;DR: ProV-Mem将流式视频问答中的主动决策拆成时间触发、回答指导和重复响应过滤三类轻量记忆，在四种场景中改善综合效率质量指标。

Details

Motivation: 主动流式视频问答要求模型持续观察输入流，自行决定何时查询、说什么以及是否回应。保留长视觉历史成本高，直接丢弃又会漏掉事件或重复回答。

Result: ProV-Mem是冻结VLM骨干的免训练记忆框架：时间触发记忆决定何时调用骨干并处理长时间沉默；指令记忆通过固定提示先验引导简洁且基于画面的回答；情景记忆过滤近重复回答，以决定是否发言。作者在理想化条件充分假设下推导分阶段过程相对完整状态联合策略的近似误差界，并提出成本感知AUC（CAUC），在PAUC基础上惩罚冗余、过频和过长输出。ProactiveVideoQA四种情境中，ProV-Mem的CAUC均最高；在WEB、EGO和TV上PAUC也最高，同时每次回答推理延迟较低。

Insight: 分解“何时、说什么、是否说”可避免持续视频系统频繁查询或重复输出。理论界依赖条件充分假设；摘要未提供指标数值和不同场景延迟，也未比较不同VLM规模。

Abstract: Proactive streaming video question answering requires a model to continuously monitor an incoming stream and autonomously decide when to speak, what to say, and whether to emit a response. Existing streaming systems face a fundamental difficulty: retaining long visual histories is costly, whereas discarding them can lead to missed events and repeated responses. We propose \textbf{ProV-Mem}, a training-free proactive memory framework built around a frozen VLM backbone. ProV-Mem decomposes the proactive decision into three lightweight memory tiers: a temporal-trigger memory controls when to query the backbone and recovers from prolonged silence, an instructional memory guides what to say through a fixed prompting prior for concise and visually grounded generation, and an episodic memory determines whether to emit by filtering near-duplicate responses. Under an idealized conditional-sufficiency assumption, we derive a stagewise approximation bound relating the decomposed process to a full-state joint policy. We further introduce Cost-Aware Area Under the Curve (\textbf{CAUC}), which extends PAUC by explicitly discounting redundant, overly frequent, and unnecessarily long emissions. Experiments on the four scenarios of ProactiveVideoQA show that ProV-Mem achieves the best CAUC across all scenarios and the best PAUC on WEB, EGO, and TV, while maintaining low per-response inference latency.

[101] spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics cs.CV | cs.AIPDF

Benedikt von Querfurth, Johann Wenckstern, Eeshaan Jain, Lukas Klein, Phil Cheng

TL;DR: spora汇集跨队列、平台和蛋白面板的空间蛋白组多模态数据，并提供统一处理接口与跨十二个队列的基础模型基准。

Details

Motivation: 空间蛋白组在单细胞分辨率观察组织分子组成及空间组织，但不同队列的蛋白面板、采集平台和组织背景差异大；训练通用基础模型需要大规模、格式协调的数据，公开资源目前较小且规范不一。

Result: spora包含13,546张标准化多重成像，来自5,477名患者、35个队列和五种主要空间蛋白组技术。由生物医学专家整理，提供适合模型训练的格式、细胞与组织分割、结构化临床元数据，以及在可用时配对的H&E图像以支持虚拟染色模型训练。spora[io]提供跨队列和模态统一的切片、采样及标准化接口；spora[bench]涵盖细胞表型、诊断和治疗预测、虚拟染色等任务，覆盖十二个队列，可在统一条件下比较空间蛋白组基础模型与任务专用基线。

Insight: 跨平台标准化及配对组织学和结构化临床信息可支持空间蛋白组基础模型训练与统一比较。摘要未给出各任务基准成绩、患者队列隐私和数据访问条件，基准覆盖十二个队列而非全部35个。

Abstract: Spatial proteomics (SP) captures the molecular composition and spatial organization of tissues at single-cell resolution, opening a direct view of the cellular ecosystems that drive disease. Protein panels, acquisition platforms, and tissue contexts differ from one cohort to the next, and a coherent picture of tissue biology can only emerge from models trained to bridge this heterogeneity. Foundation models offer the natural path forward, but their training depends on large, harmonized corpora that span platforms, panels, and cohorts. No public resource currently approaches that scale; available data sit in small, narrowly scoped releases with incompatible formats and conventions. We present $\texttt{spora}$, a large-scale multimodal SP dataset of 13,546 standardized multiplex images from 5,477 patients across 35 cohorts and five major SP technologies. Curated with biomedical experts, $\texttt{spora}$ provides model training-ready formats together with cell and tissue segmentations, structured clinical metadata, and, where available, paired H&E enabling the training of virtual staining models. It is complemented by $\texttt{spora [io]}$, a unified interface for tiling, sampling, and standardization across all cohorts and modalities, and $\texttt{spora [bench]}$, a benchmark suite spanning cell phenotyping, diagnostic and therapeutic prediction as well as virtual staining tasks across twelve cohorts that places current SP foundation models on common ground against task specific baselines.

[102] PACE: Pareto-Adaptive Compression for Efficient Native MLLMs cs.CV | cs.AIPDF

Lei Tian, Chunzheng Zhu, Kai Wu, Yifan Zhang, Haihua Yang

TL;DR: PACE面向无外置视觉编码器的原生多模态大模型，依据空间方向重要性和多目标Pareto预算分配压缩视觉token。

Details

Motivation: 直接把像素块送入统一Transformer的原生MLLM可能包含大量视觉token冗余，现有压缩方法主要针对ViT视觉编码器架构，未覆盖最有压缩需求的原生模型。

Result: 分析发现冗余沿高度和宽度轴独立且分布不均，早期层视觉token则难以区分且普遍冗余。PACE通过多模态RoPE的位置编码维度解耦，将注意力分成高度和宽度两个正交空间显著性估计，并用位置注意力信号门控，得到方向感知的重要度，额外开销很小。方法将token预算视为性能保持、推理效率和空间覆盖之间的Pareto优化，联合确定压缩深度和保留比例。在仅保留25%–50%视觉token时，多个基准表现接近无损，部分甚至提升，同时FLOPs降低超过2倍。

Insight: 原生模型的视觉冗余具有空间方向和层级结构，压缩时可利用该结构而非套用ViT剪枝。摘要未给出具体基准分数、不同保留比例的逐项结果和端到端延迟。

Abstract: Native multimodal large language models (native MLLMs) jointly encode visual and linguistic sequences within a unified autoregressive Transformer, emerging as a promising paradigm for multimodal understanding. Unlike conventional ViT-based MLLMs relying on external vision encoders, ViT-free native MLLMs tokenize pixel patches directly, potentially suffering severe visual token redundancy. A pilot study confirms this: unlike ViT-based models that degrade under compression, ViT-free MLLMs tolerate aggressive compression with near-lossless or even improved performance. Yet existing compression methods target ViT-based settings, leaving native MLLMs, which stand to gain the most, unexplored. To close this gap, we introduce an analysis pipeline centered on token discriminability and cross-dimensional heterogeneous redundancy, revealing that (i) redundancy is distributed independently and heterogeneously along height and width axes, and (ii) early-layer visual tokens are indiscriminable and uniformly redundant. Guided by these findings, we propose PACE (Pareto-Adaptive Compression for Efficient Native MLLMs), the first visual token compression framework tailored for native MLLMs with Multimodal-RoPE. Leveraging dimensional decoupling of positional encoding, PACE decomposes attention into two orthogonal spatial saliency estimates along height and width, gated by a positional attention signal, yielding direction-aware importance with negligible overhead. It further casts token budget allocation as Pareto optimization over performance preservation, inference efficiency, and spatial coverage, jointly identifying optimal compression depths and retention ratios. Experiments show PACE maintains near-lossless performance at only 25%-50% visual token retention, even improving several benchmarks, while delivering over 2x FLOPs reduction.

[103] Generative Text Modelling and Language-Guided Visual Attention for Deciphering Palimpsests cs.CVPDF

Robin Hollifeldt, Tianru Zhang, Jessica Ström, Ekta Vats

TL;DR: 该研究用GAN和扩散模型重建被覆盖的古抄本底层手写文字，并以视觉语言模型注意力将文本描述作为识别先验。

Details

Motivation: 重用抄本中，原有文字被擦除并覆盖，底层文字恢复对文化遗产保存重要，但材料特性使修复工作困难且耗时。

Result: 作者研究多种生成模型，并设计定制去噪扩散模型学习结构化噪声模式以重建底层文字；恢复图像随后由微调的注意力式手写文本识别网络处理。另提出语言引导识别框架，将词语或句子形式的类别描述作为先验，并使用预训练视觉语言模型（如CLIP）生成视觉注意力图，将描述与相关图像区域对齐。为推动后续研究，作者发布稀缺的合成重用抄本数据集、兼容CLIP的文字描述、预训练模型和代码。

Insight: 生成式图像修复可与语言提示区域注意力结合，辅助区分真实文字痕迹与材料噪声。摘要未给出恢复准确率、真实古抄本测试规模和专家校验结果，公开资源有助于研究但不能据此确认历史文本复原正确性。

Abstract: Recovering obscured text in palimpsests, i.e. the manuscripts where the original writing (under-text) was erased and overwritten for reuse, is vital for preserving lost information. Due to the nature of the material, such a task can be highly challenging and time-consuming. This paper presents an approach towards automated palimpsest handwritten text modeling for under-text recovery. We investigate generative models, including Generative Adversarial Networks (GANs) and diffusion models, and design a custom denoising diffusion model that learns structured noise patterns to reconstruct the under-text. The recovered text is subsequently processed by a fine-tuned, attention-based handwritten text recognition network. Furthermore, we investigate a language-guided recognition framework that leverages textual class descriptions as informative priors, enabling the text recognition model to distinguish meaningful features from noise. This is achieved by grounding task-specific descriptions, phrased in words or sentences, through visual attention maps generated by pretrained vision-language models like CLIP (Contrastive Language-Image Pre-training). To facilitate future research and benchmarking, we release new datasets for synthetic palimpsests (currently scarce), CLIP-compatible text descriptions, and pre-trained models with code.

[104] BiNar: Bitstream Native Restorer for Bitstream-corrupted Image Restoration cs.CVPDF

Fangcheng Li, Zhen Yu, Kejun Wu, Tianyi Liu

TL;DR: BiNar直接建模损坏压缩码流的序列字节语义，结合解码图像线索和视觉基础模型定位，指导扩散模型修复无法正常解码的图像。

Details

Motivation: 压缩码流中的损坏可能在像素域引起不可预测的错误传播和失真放大，甚至导致标准解码器完全无法解码，因此传统以像素为起点的图像恢复方法难以处理。

Result: BiNar由三部分构成：码流语义建模模块直接处理损坏码流中的连续字节，提取鲁棒信息并生成细粒度描述，不依赖像素解码；损坏自适应多模态聚合模块再从已解码的损坏图像中提取自适应视觉线索，并与码流描述融合；盲损坏定位模块利用视觉基础模型的分割先验定位需恢复区域。定位区域作为视觉提示、细粒度码流描述作为语言提示，共同条件化扩散图像修复模型。多种码流损坏条件实验显示，BiNar在感知质量和客观质量方面均超过所比较的先进恢复模型；消融支持细粒度码流语义的作用。

Insight: 从码流域提取的语义可在像素解码失败时提供额外恢复依据，再与损坏图像中仍可用的视觉线索结合。摘要未报告具体指标、码流格式和损坏类型覆盖，无法量化各模块带来的收益。

Abstract: Bitstream-corrupted image restoration (BCIR) aims to reconstruct faithful images from corrupted compressed bitstreams. However, corruption occurs in bitstream-domain may lead to unpredictable error propagation and distortion amplification in pixel-domain, and even result in complete undecodability by standard decoders, posing significant challenge to existing image restoration paradigms. To address this challenge, we propose Bitstream Native Restorer (BiNar), a generative restoration framework that leverages fine-grained bitstream-native semantics to guide bitstream-corrupted image restoration. In BiNar, Bitstream Semantic Modeling (BSeM) extracts robust informative semantics and models sequential bytes directly from corrupted bitstream, enabling to generate fine-grained bitstream descriptions without relying on pixel decoding. Based on the descriptions,the Corruption-adaptive Multimodal Semantic Aggregation (CMSA) further extracts adaptive visual modal cues in decoded corrupted images and aggregates to bitstream descriptions. The multimodal aggregated fine-grained descriptions can serve as coherent restoration guidance. Finally, Blind Corruption Localization (BCoL) exploits the segmentation prior of visual foundation models to identify and locate error regions that requires recovery in corrupted images. The located error regions and the fine-grained descriptions jointly serve as visual prompt and language prompt of diffusion models image restoration, respectively. Extensive experiments under diverse bitstream corruption conditions show that BiNar outperforms all State-of-the-arts image restoration models regardless of perceptual quality or objective quality. Ablation studies demonstrate the effectiveness of fine-grained bitstream-native semantics in BCIR task.

[105] Hierarchical and Query-Aware Visual Token Pruning for High-Resolution LVLMs cs.CVPDF

Luther Wei, Ling Zou, Kang-Yang Huang, Yu-Syuan Xu, Wen-Huang Cheng

TL;DR: 该训练免框架在执行LLM前按样本复杂度、局部裁剪相关性和token级文本图像对齐分配视觉token预算，在大幅压缩时保持LLaVA-NeXT性能。

Details

Motivation: 高分辨率视觉语言模型通过全局缩略图和多个局部裁剪保留细节，但产生数千视觉token，增加LLM推理成本。现有剪枝多不看查询；查询感知方法又可能只在粗裁剪层面工作，或等token进入昂贵LLM层后才计算相关性。

Result: 方法在任何LLM块运行前进行自上而下分层剪枝：样本级按视觉复杂度动态确定保留比例，裁剪级以句子级跨模态相似度分配预算，token级在共享空间中对齐投影视觉token与文本查询token，并用文本去偏归一化形成细粒度分数；最终结合局部和全局显著性选择token。在LLaVA-NeXT-7B上，保留25%和10%视觉token时分别保留未剪枝性能的98.0%和94.8%。保留25%时推理吞吐量提高2倍，POPE基准甚至比未剪枝基线高0.7分。

Insight: 将查询信息在进入LLM前就用于逐级分配视觉预算，可兼顾高分辨率细节与推理效率。摘要结果仅在LLaVA-NeXT-7B上报告，未给出各任务延迟、不同图像尺寸和更大模型上的表现。

Abstract: High-resolution Large Vision-Language Models (LVLMs) preserve critical, fine-grained visual details by encoding a global thumbnail alongside multiple local crops. However, the resulting thousands of vision tokens introduce severe computational bottlenecks during Large Language Model (LLM) inference. Existing token pruning methods rely primarily on query-agnostic visual saliency, while current query-aware approaches either operate at a coarse crop-level resolution or compute token relevance too late, after vision tokens have already entered the heavy LLM layers. To address this, we propose a training-free, hierarchical token pruning framework that utilizes query guidance prior to executing any LLM blocks. Our top-down hierarchy first operates at the instance-level, dynamically scaling the vision token retention ratio of each sample based on its visual complexity. Next, at the crop-level, we distribute this sample budget across local crops using sentence-level cross-modal similarity. Finally, at the token-level, we exploit projected vision tokens alongside textual query token embeddings in the shared space to measure fine-grained alignment, applying a text-debiasing normalization to produce high-fidelity query scores. These query-guided budgets and scores are fused with local and global visual saliency metrics to select the most critical vision tokens. Extensive experiments on LLaVA-NeXT-7B demonstrate that our plug-and-play framework retains 98.0% and 94.8% of unpruned performance at 25% and 10% vision token retention, respectively. Notably, at 25% retention, our method delivers a 2.0x inference throughput speedup and even surpasses the unpruned baseline on POPE by 0.7 points.

Dongzhihan Wang, Yi Du, Chen Wang, Xiuxian Li, Liang Xu

TL;DR: HAVANA将视频动作模型的预测性推理和反应式局部导航规划分成异步快慢系统，减少重复未来预测并在导航基准和真实机器人环境中测试。

Details

Motivation: 视觉语言导航既需要长时程、遵循指令的规划，也需要根据当前几何及时修正局部轨迹。现有视频动作模型在单一流程中联合预测视觉未来和局部轨迹，但前者可跨多次更新保持稳定，后者必须响应新观察。

Result: HAVANA由慢系统根据历史观察生成紧凑、语言条件化的潜在计划，快系统结合最新计划和当前观察生成局部轨迹。两者异步执行，使机器人无需反复预测未来即可快速适应；推理阶段不生成未来视频。在OmniNavBench上，HAVANA超过既有基线；真实部署还显示其可泛化至不同机器人具身和未见环境。

Insight: 把稳定的高层预测计划与需频繁更新的局部轨迹分离，能减少视觉未来生成成本并提高反应速度。摘要未给出基准分数、真实部署任务数量和延迟数值，跨具身泛化结论需结合完整实验范围解释。

Abstract: Vision-Language Navigation (VLN) requires long-horizon instruction-consistent planning and short-horizon local planning in current geometry. Video-action models (VAMs) offer a promising way to learn predictive representations for navigation. However, existing VAMs jointly predict visual futures and local trajectories in a single process, even though they operate at different paces: future prediction remains stable across multiple updates while local trajectories must react to fresh observations. To bridge this gap, we introduce HAVANA, a dual-system VAM that decouples predictive reasoning from reactive local planning. A slow system produces a compact language-conditioned latent plan from observation history, while a fast system combines the latest plan with the current observation to generate local trajectories. Their asynchronous execution enables rapid adaptation without repeated future prediction, and no future video is generated during inference. Experiments on OmniNavBench demonstrate state-of-the-art navigation performance over prior baselines, while real-world deployments further show robust generalization to unseen environments across different robot embodiments. Our code and checkpoints will be available at https://github.com/Anastasiawd/HAVANA.

[107] Precision Cotton Disease Management: Leveraging UAV Imagery and Custom YOLO Deep Learning Framework for Enhanced Detection cs.CVPDF

1. Anwar Iqbal*1 aiqbal. rime23smme@student. nust. edu. pk 2. Kunwar Faraz Ahmad Khan1 kunwar. faraz@smme. nust. edu. pk 3. Shahbaz Khan1 shahbaz. khan@smme. nust. edu. pk 4. Khawaja Fahad Iqbal1 fahad. iqbal@smme. nust. edu. pk 5. Sara Ali1, 2 sara. baber@smme. nust. edu. pk 6. Saqib Nazir1 saqib. nazir@smme. nust. edu. pk 7. Muhammad Azhar Javed3 muhammadazhar6231@gmail. com 1 Department of Robotics, Artificial Intelligence, SMME, National University of Science

TL;DR: Cotton-Vision是面向棉花病害早期识别的轻量无人机图像检测器，结合Ghost Bottleneck、ECA和CSP结构，在巴基斯坦七个棉田采集数据上取得较高检测精度。

Details

Motivation: 棉花病害常在人工巡田发现前造成产量损失，无人机图像和深度学习可提供规模化预警，但现有模型早期识别精度、跨域迁移和计算成本仍有不足。

Result: Cotton-Vision采用单阶段全卷积检测器，使用Ghost Bottleneck减少约20%参数、ECA自适应重标定通道、CSP融合改善梯度流。数据来自巴基斯坦旁遮普省七个棉田，含2,700多张高分辨率无人机图像；旋转、剪切、曝光和饱和度增强后形成11,704个训练样本。留出测试集上精确率97.7%、召回率77.7%、mAP@50为85.6%、mAP@50-95为76.6%。作者报告相较YOLOv5/8/11/12、RT-DETR和RF-DETR等基线，精确率、召回率及mAP提高10–20%；消融显示各模块有贡献，较基线架构提高12–16% mAP。

Insight: 轻量注意力和特征融合可在棉田无人机病害检测中兼顾精度与计算可行性。召回率77.7%意味着仍有未检出病害，且数据来自一个省份的七个棉田；摘要未报告外部地区验证，不能推断跨地区泛化。

Abstract: Cotton sustains the textile economies of over 80 countries with an annual yield of 113 million bales, yet 30-60% of yield losses are still attributed to viral, bacterial, and fungal diseases that remain undetected until manual field scouting. To address this critical issue, the integration of Unmanned Aerial Vehicle (UAV) imagery with deep learning offers a scalable early-warning system; however, existing models suffer from low early-stage accuracy, poor transferability, and high computational cost. In this paper, we propose Cotton-Vision, a fully convolutional single-stage detector that synergizes three lightweight architectural innovations: Ghost Bottleneck modules for a 20% parameter reduction, Efficient Channel Attention (ECA) for adaptive feature recalibration, and Cross-Stage Partial (CSP) fusion for enriched gradient flow. The network is trained on a novel dataset collected across seven cotton fields in the Punjab province of Pakistan, comprising over 2,700 high-resolution drone images. Extensive data augmentation (rotation, shear, exposure, and saturation) expands the dataset to 11,704 training samples. On a held-out test set, Cotton-Vision achieves a precision of 97.7%, a recall of 77.7%, an mAP@50 of 85.6%, and an mAP@50-95 of 76.6%. These results represent a 10-20% improvement in precision, recall, and mAP over state-of-the-art baselines, including YOLOv5, YOLOv8, YOLOv11, YOLOv12, RT-DETR, and RF-DETR. Through a comprehensive ablation study, we demonstrate that each proposed architectural enhancement contributes to the model’s superior performance, yielding a 12-16% mAP improvement over the baseline architecture. Our findings confirm that this synergy of lightweight and attention-based modules provides a robust, computationally viable solution for early cotton disease detection.

[108] CRISP: Progressive Illumination Curriculum for Recurrent Low-Light Video Enhancement with Semantic Priors cs.CVPDF

Ruirui Lin, Xin Yan Lim, Yini Li, Zhiming Liu, Nantheera Anantrasirichai

TL;DR: CRISP通过逐步增加照明恢复难度的训练课程和两帧循环结构增强低照度视频，兼顾亮度、色彩、空间细节和时间连贯性。

Details

Motivation: 低照度视频增强需恢复亮度、颜色和精细结构，同时在噪声及运动存在时保持时间一致性，严重退化区域尤其难以兼顾这些目标。

Result: CRISP使用两帧循环架构在相邻输入间传播信息，并采用渐进照明训练课程：监督目标从低照输入与参考间的中间值逐步过渡到完整参考，渐进提升亮度和色彩修正要求。该课程稳定循环训练且不增加推理成本，在消融组件中带来最大性能提升。可选模块通过AdaLN式调制加入冻结DINOv3语义特征以支持结构恢复，无需微调基础模型。在DID和BVI-RLV上，CRISP分别取得所测方法中的最高PSNR和最高SSIM，同时保持较高计算效率。

Insight: 逐步增加恢复难度可稳定循环模型训练，冻结语义先验则作为结构补充。摘要未给出具体PSNR/SSIM数值、推理速度或不同光照强度的分项结果；“效率”需结合正文硬件设置解读。

Abstract: Low-light video enhancement requires the recovery of brightness, color, and fine spatial structures while preserving temporal coherence in the presence of noise and motion. Existing methods often struggle to balance these objectives, particularly in severely degraded regions. This paper presents CRISP, a recurrent framework trained using a progressive illumination curriculum. Its two-frame recurrent architecture propagates information across successive frames to exploit temporal context beyond the immediate input pair. During training, supervision progressively transitions from intermediate targets interpolated between the low-light input and the reference to the full reference, gradually increasing the required brightness and color correction. This curriculum stabilizes recurrent training without introducing additional inference cost and yields the largest performance improvement among the components evaluated in the ablation study. To further support structural reconstruction, the framework optionally incorporates frozen DINOv3 semantic features through adaptive layer normalization (AdaLN)-style modulation, without fine-tuning the foundation model. Experiments on DID and BVI-RLV demonstrate that CRISP achieves the highest PSNR on DID and the highest SSIM on BVI-RLV among the evaluated methods, while maintaining computational efficiency.

[109] PACE-Mem: Prefix-Aware Complementary Evidence for Multimodal Agents cs.CVPDF

Shidu Ren, Rong-cheng Tu, Hang Zhou, Xiao Luo

TL;DR: PACE-Mem用最大边际相关选择互补记忆证据，并依据选择分数轨迹自适应确定检索前缀长度，改善长视频记忆问答。

Details

Motivation: 多模态智能体长期记忆会积累许多相似场景、动作和实体观察；检索的相关记忆可能高度重叠，而固定检索深度可能对不同问题返回过少或过多证据。

Result: PACE-Mem针对结构化记忆，用查询引导池化将片段内记录汇总为相关性分数和表示，再以最大边际相关在查询相关性与和已选证据重叠之间折衷。基于该分数轨迹，前缀长度选择算法比较相邻分数下降并按长度归一化，同时考虑分数变化和已累积证据量。该训练免插件不改变记忆构建和答案生成。在三个长视频基准和两个智能体框架上，相较平均检索深度相近的最强Adaptive-k基线，问答准确率平均提高1.6%。分析显示，同时增加和删除证据与较大净准确率收益相关，单独增删的收益较小。

Insight: 检索证据的数量和互补性应联合决定；固定k容易无法适配不同问题。摘要未列出各基准的分项成绩、检索时延和“1.6%”口径，结论限于片段级智能体记忆。

Abstract: As multimodal agents accumulate experience, their long-term memories contain many similar observations of recurring scenes, actions, and entities. Answer quality depends on the evidence retrieved from these memories into the model’s context. However, query-relevant memories may contain overlapping information, and a fixed retrieval depth may return too little or too much evidence for a given question. To address these challenges, we propose PACE-Mem, Prefix-Aware Complementary Evidence, which constructs a complementary evidence sequence and determines its length from the same selection trajectory. For structured memories, query-guided pooling combines records within each clip to obtain relevance scores and representations. Using these relevance scores and representations, maximal marginal relevance (MMR) selection balances query relevance against overlap with previously selected evidence. From the resulting score trajectory, a prefix-length selection algorithm determines retrieval depth by comparing length-normalized adjacent declines, accounting for both score changes and the amount of evidence accumulated. PACE-Mem is a training-free, plug-in retrieval method for clip-based agent memory systems, leaving memory construction and answer generation unchanged. Across three long-video benchmarks and two agent frameworks, PACE-Mem improves question-answering accuracy by an average of 1.6% over Adaptive-k, the strongest baseline at comparable mean retrieval depths. Further analysis reveals that jointly adding and removing evidence is associated with larger net accuracy gains than either change alone.

[110] ZeroDSS: Semantic Segmentation from Class Names Alone cs.CV | cs.AIPDF

Tian Ming, Marzi Heidari, Yuhong Guo

TL;DR: ZeroDSS只使用类别名称和通用生成先验生成训练图像及掩码，并通过模型反馈迭代重平衡类别和修正伪掩码，完成无目标域数据的语义分割。

Details

Motivation: 语义分割依赖真实像素级标注，在新领域获取密集标签成本高。研究关注无数据适配设定：训练时没有目标域图像、掩码或描述，仅知道预先给定类别集合。

Result: ZeroDSS将合成监督视为可调训练分布，而非真实数据的直接替代。系统通过LLM引导组合提示生成结构化文本分布并自适应平衡类别，再利用扩散注意力生成图像—掩码对。闭环流程先用置信度加权监督和双尺度一致性训练分割器，再依据模型反馈针对表现较差类别调整分布并修正扩散掩码，最终用EMA教师—学生框架迭代更新监督和模型。仅用合成训练数据，DeepLabV3在MS COCO上达到35.0% mIoU；DeepLabV3和Mask2Former在PASCAL VOC上分别为66.3%和63.5%。

Insight: 按模型反馈反复塑造合成数据分布，有助于缩小合成到真实图像域的差距，且不需要目标域图像或标注。摘要没有列出生成模型、计算成本及类别分项表现，成绩不应直接与使用真实域数据的方法混为同条件比较。

Abstract: Semantic segmentation depends heavily on real images with pixel-level annotations, making dense supervision a major obstacle to deployment in new domains. We study task-specific data-free semantic segmentation, where training uses generic pretrained generative priors but has access to no target-domain images, masks, or captions–only a predefined category set. We introduce Zero-Data Semantic Segmentation (ZeroDSS), a feedback-driven approach that treats synthetic supervision as an adaptively controllable training distribution rather than a direct surrogate for real data. ZeroDSS constructs a semantically structured prompt distribution through LLM-guided compositional generation and adaptive class rebalancing, then converts the prompts into image-mask pairs using diffusion attention. Its feedback-driven distribution refinement consists of a two-phase closed-loop training procedure with a single loop-back step. In Phase I, a segmenter is trained using confidence-weighted supervision and dual-scale consistency. In Phase II, model feedback is used to adaptively target underperforming categories and refine diffusion-derived masks using model predictions. After this refined initial training, a teacher-student framework is utilized in the final phase to iteratively improve the fused supervision and segmenter through exponential moving average (EMA) updates. Using synthetic training data only, ZeroDSS achieves 35.0% mIoU on MS COCO with DeepLabV3, and 66.3% and 63.5% mIoU on PASCAL VOC with DeepLabV3 and Mask2Former, respectively. These results show that explicitly constructing and iteratively refining the synthetic supervision distribution can substantially reduce the synthetic-to-real gap in dense prediction.

[111] S3D-Mosaic: Subgradient-Guided Cross-Band Structure Transfer for Spectral Demosaicing cs.CVPDF

Weiqiang Zhao, Xian Li, Yuzhe Gui, Tianzhu Liu, Yanfeng Gu

TL;DR: S3D-Mosaic用子梯度引导的跨波段结构迁移进行多光谱马赛克去马赛克，以对比度幅值不变的方式传递邻近波段的空间结构。

Details

Motivation: 多光谱滤光片阵列以稀疏方式采集各波段，导致严重空间混叠；缺失结构可从相邻波段推断，但波长相关对比度变化使直接迁移不可靠。既有模型方法缺少幅值不变机制，学习方法则需要特定布局训练或逐图调参。

Result: S3D-Mosaic构造子梯度引导正则化框架，并以广义余弦相似解释，实现凸的幅值不变信息迁移。紧凑的方向稀疏等变差分卷积DSE-DConv以低成本提取波段级方向选择性高频结构。将其用于谱相邻波段后得到跨波段自结构相似性正则，使用双向扫描ADMM优化，解耦双向依赖并传播对幅值变化稳健的结构信息。DSE-DConv的两个参数只需自监督训练一次；在CAVE和NTIRE 2022上超过先进无监督方法，并在固定重建超参数下泛化到不同MSFA布局、全色马赛克融合和真实无人机数据。

Insight: 迁移空间结构而非直接迁移强度幅值，可缓解跨波段对比差异导致的失真；固定超参数跨布局泛化是重要优势。摘要未提供重建指标和不同传感器布局的具体误差，结论限于所测数据。

Abstract: Multispectral filter array (MSFA) imaging enables snapshot spectral acquisition, but its per-band sparse sampling causes severe spatial aliasing. Although missing spatial structures can be inferred across bands, wavelength-dependent contrast variations make direct cross-band transfer unreliable. Existing model-based methods lack explicit magnitude-invariant mechanisms, and learning-based methods require layout-specific training or per-image tuning, limiting generalization. To address these issues, we propose S3D-Mosaic, a model-based spectral demosaicing method built upon a subgradient-guided cross-band structure transfer framework. We first establish a subgradient-guided regularization framework with a generalized cosine-similarity interpretation, providing a convex mechanism for magnitude-invariant information transfer. Then we design a compact Directional Sparse Equivariant Difference Convolution (DSE-DConv) for band-level extraction of orientation-selective, high-frequency structure features at low computational cost. Instantiating the framework with DSE-DConv on spectrally adjacent bands yields a cross-band self-structure similarity regularizer for spectral demosaicing, which we optimize via a dual-scan ADMM algorithm that decouples the bidirectional dependency while propagating magnitude-robust structural information across bands. With DSE-DConv’s two parameters trained once via self-supervision, S3D-Mosaic outperforms state-of-the-art unsupervised methods on CAVE and NTIRE 2022, and generalizes robustly to varying MSFA layouts, pan-mosaic fusion, and real-world UAV data with fixed reconstruction hyperparameters. The source code is available at \url{https://github.com/bestwishes43/S3D-Mosaic}.

[112] BodyBridge: Grounded Body Layers for Sim-to-Real Translation of Vulnerable Road Users cs.CVPDF

Mark Azer, Fabian B. Flohr

TL;DR: BodyBridge以像素对齐的规范人体部位对应和精确掩码控制视频扩散模型中的弱势道路使用者，改善仿真到真实转换的姿态一致性。

Details

Motivation: 仿真到真实驾驶视频扩散模型可生成逼真场景，但通常把行人和骑行者作为通用物体类别，无法表达影响安全的关节姿态及个体身份。

Result: BodyBridge提出Grounded Body Layers表示，结合场景几何、每像素规范人体对应和精确VRU掩码。该表示向视频扩散模型提供显式关节结构，在控制分支后训练时强化VRU区域，并支持推理时与场景几何控制相加组合；还可按关节评估生成结果并分类身份错误。在CARLA到真实域转换中，相较已发布深度条件基线，近距离PCK@0.5提高22%；用规范人体对应替换法线图身体控制后，身体翻转减少18%；场景组合控制使近距离朝向翻转减少27%。

Insight: 人体对应表示补足场景级控制未包含的姿态和身份结构，有助于生成符合安全分析需要的VRU视频。摘要未给出绝对指标、真实数据规模和远距离效果，代码、数据及模型计划公开。

Abstract: Simulation-to-real driving video diffusion models can translate photorealistic scenes, but they typically represent pedestrians and cyclists, vulnerable road users (VRUs), as generic object classes. This is insufficient for vulnerable road users whose safety-relevant appearance depends on articulated pose and identity. We present BodyBridge and Grounded Body Layers (GBL), a simulator-derived, pixel-aligned representation that combines scene geometry, canonical per-pixel body correspondence, and exact VRU masks. GBL provides explicit articulation to the video diffusion model, emphasizes VRU regions during control-branch post-training, and supports additive composition with scene-geometry control at inference. It also enables per-joint evaluation of generated VRUs and a failure taxonomy for identity errors. On CARLA-to-real translation, BodyBridge improves near-range PCK@0.5 by 22% over the released depth-conditioned baseline. Replacing normal-map body control with canonical correspondence reduces body flips by 18%, while composition reduces near-range facing flips by 27%. These results show that explicit body grounding complements scene-level control and improves pose-faithful VRU generation. All code, data, and model checkpoints will be made publicly available to support reproducible, pedestrian-aware autonomous driving research.

[113] Do 3D Large Language Models Really Use 3D Evidence? cs.CV | cs.AIPDF

Yifei Zhao, Fanyu Zhao, Han Zhang, Yixuan Lin, Yuyang Hou

TL;DR: UNICORNS评测3D大语言模型面对指代歧义、证据不足和错误预设时能否依据场景证据作答；HORN后训练显著提升挑战类别表现。

Details

Motivation: 现有3D场景问答基准通常评估可回答问题，很少检查模型能否处理指代歧义、场景证据不足或问题中含有错误前提，因此无法判断答案是否真正依据3D证据。

Result: UNICORNS含557个ScanNet扫描中的11,387个问题，覆盖四种证据条件、八类挑战及可回答控制题；只有满足正确响应行为且不包含场景不支持主张才算通过。六个开源3D-LLM中，最强零样本模型在可回答控制题准确率为61.22%，八类挑战仅40.94%。作者提出义务感知后训练方法HORN，结合组相对任务奖励、叶节点条件化义务奖励和监督控制回放。实验中HORN挑战类别成绩为75.1±1.0%，总体为77.2±0.8%，超过所有基线且未损害可回答控制准确率。

Insight: 可靠的3D问答不仅要回答有证据问题，还需在证据不足或前提错误时采取合适行为。摘要未提供各挑战类别分项成绩和模型训练成本，结论适用于ScanNet及所测3D-LLM。

Abstract: Recent 3D Large Language Models (3D-LLMs) have shown strong performance on scene question answering with extensive post-training. Yet, existing benchmarks rarely evaluate how 3D-LLMs respond to referential ambiguity, insufficient scene evidence, or false presuppositions, obscuring whether their responses are grounded in 3D evidence. In this work, we introduce UNICORNS, a benchmark for evaluating the reliability of 3D-LLM responses across four evidence conditions. UNICORNS contains 11,387 questions across 557 ScanNet scans, covering eight challenge categories alongside answerable controls. A response passes only if it satisfies the required response behavior and contains no scene claim unsupported by the available evidence. Evaluation of six open-sourced 3D-LLMs reveals a pronounced performance gap: even the strongest zero-shot model achieves 61.22% accuracy on answerable controls, but only 40.94% across the eight challenge categories. To address this gap, we further propose HORN, an obligation-aware post-training method combining group-relative task rewards, leaf-conditioned obligation rewards, and supervised control replay. Extensive experiments demonstrate that HORN reaches $75.1 \pm 1.0%$ across the challenge categories and $77.2 \pm 0.8%$ overall, outperforming all baselines without degrading answerable-control accuracy. Our findings enable a systematic evaluation and improvement of evidence-grounded 3D reasoning for 3D-LLMs (code: https://anonymous.4open.science/r/unicorns-6B38/).

[114] HumanSkill-4D: Training-Free Language Querying of Animatable Gaussian Avatars in Space and Time cs.CV | cs.AIPDF

unsu Kim, Naeun Kim, Dongyoon Han, Seungryul Baek

TL;DR: HumanSkill-4D无需训练即可通过语言查询从可动画高斯人形中选择身体或服装部件，并按姿态与动作条件定位对应时间片段。

Details

Motivation: 语言条件3D分割难以区分相邻身体部位和服装，也可能需要间接指代；细粒度动作在RGB外观上相似，导致基于语言查询的动作区间选择困难。

Result: HumanSkill-4D为每个头像建立语义记忆，将身体及服装类别关联到持久高斯区域和外观描述，并在不同头像间共享多模态大模型和几何算子。模型按外观、关系或功能选择类别；时间定位则把算子组合成程序，检查运动记忆中保存的驱动关节轨迹是否满足查询条件。类别选择和算子输出可追踪，组合或增加算子可扩展至新动作条件。作者提出三个由人工标注的基准，实验显示该方法在部件分割及3D、4D指代分割上超过基线，同时记忆紧凑，每个头像仅需几分钟设置。

Insight: 将静态语义区域记忆与可组合运动条件算子结合，可把语言查询扩展到空间部件和时间动作定位。摘要未给出基准名称、具体指标和设置时间分布，训练免方法仍需每个头像建立记忆。

Abstract: We introduce avatar referring segmentation, which selects regions of animatable Gaussian avatars from language queries. A query can also specify a pose or motion to select matching intervals. This lets users interact with avatars for local editing and motion inspection. However, making these selections with existing language Gaussian splatting methods poses two challenges: adjacent parts and garments are hard to separate and may be referred to indirectly (region ambiguity), while fine-grained motions can look alike in RGB (motion ambiguity). We address these challenges with HumanSkill-4D, a training-free framework that shares a multimodal large language model (MLLM) and geometric operators across avatars. For each avatar, a semantic memory links body-part and garment categories to persistent Gaussian regions and appearance descriptions. The MLLM uses category names and descriptions to select categories by appearance, relation, or function. To find matching intervals, it composes operators into programs that check query conditions on driving joint trajectories stored in a motion memory. These category choices and operator outputs make selections traceable, while composing or adding operators supports new motion conditions. To evaluate this task, we introduce three human-annotated benchmarks and show that HumanSkill-4D outperforms the baselines in part segmentation and 3D and 4D referring segmentation, with compact memories and only minutes of setup per avatar.

[115] Investigating Length Bias and Robustness in LVLMs for Multiple-Choice Question Answering cs.CVPDF

Md. Atabuzzaman, Hani Alomari, Chris Thomas

TL;DR: 该研究发现大型视觉语言模型在多选题中系统偏好更长的答案选项，视觉证据变模糊时偏差加剧，推理期缓解方法仍不足以部署。

Details

Motivation: 大型视觉语言模型在多项选择题上成绩较强，但模型可能利用答案长度等语言先验，而非真正依据图像判断；这类偏差会影响评测可靠性。

Result: 作者在细粒度视觉领域的受控基准上研究答案长度偏差。即使较长选项与图像不符，模型仍常偏好长答案：正确答案较短时准确率下降2%–38%，较长时提高1%–70%，且偏差跨模型规模存在。逐层机制分析显示，偏差在后层出现，表现为视觉证据被语言端长度先验压制。模糊图像实验显示视觉依据变弱时偏差增大；弃答测试表明模型也难以拒绝错误预测。多种推理期缓解方法只能部分改善，且均未能在无需不可得预言机信息的情况下实现可部署缓解。

Insight: 多选题分数可能受答案长度先验影响，视觉证据弱时尤其明显；仅调整推理期规则可能不够，需考虑架构和训练层面的偏差控制。摘要未提供各模型具体结果和基准规模。

Abstract: Large Vision-Language Models (LVLMs) demonstrate strong performance on multiple-choice question answering (MCQA), yet their reliability under biased conditions remains underexplored. We investigate length bias, the systematic preference for longer answer options, through controlled benchmarks spanning fine-grained visual domains. LVLMs consistently favor longer options even when visually inconsistent, with accuracy dropping 2–38% when correct answers are shorter and increasing 1–70% when longer. This bias persists across model scales. Through layer-wise mechanistic analysis, we reveal that length bias emerges from late-layer suppression of visual evidence in favor of language-side length priors. Blurry-image experiments show that bias intensifies when visual evidence weakens, while abstention tests reveal that LVLMs struggle to refuse incorrect predictions. We propose and evaluate multiple inference-time mitigation strategies; while these provide partial improvements, none achieves deployable bias mitigation without relying on oracle information unavailable at inference time. These results underscore that inference-time interventions are insufficient, pointing to the need for architectural solutions and bias-aware training.

[116] VisConf: Quality-Aware Consensus for Best-of-N in Vision-Language Model Reasoning cs.CVPDF

Muhammad Usman, Muhammad Qasim Ayub, Usman Ahad, Muhammad Tahir

TL;DR: VisConf结合视觉参与度和候选校准的Self-Certainty评估VLM推理轨迹质量，再以排序加权共识汇总Best-of-N候选答案。

Details

Motivation: Best-of-N通过采样多条推理轨迹提升文本LLM表现，但对VLM而言，候选轨迹是否真正使用了图像也很重要；仅看答案频次或模型置信度可能挑中没有视觉依据的推理。

Result: 作者发现概率分布置信度与视觉注意力是互补质量信号，各自单独使用都有局限，同时两者都较强的轨迹更可能正确。VisConf将视觉参与度与候选校准Self-Certainty相乘，惩罚任一信号薄弱的回答；Rank-Weighted Consensus再将同一问题内的VisConf排名转换为答案投票权重。在三种VLM和MathVista、MMMU-Pro、MMStar三个基准上，VisConf结合RWC持续超过强Best-of-N基线。候选集中存在正确但数量偏少的轨迹时，收益尤其明显。

Insight: VLM候选选择应同时考察模型自信和是否利用视觉输入，避免纯文本上自洽但忽略图像的回答主导投票。摘要未给出各基准数值、额外视觉注意力计算成本及不同采样数下的表现。

Abstract: Best-of-N methods improve text-only LLM reasoning by sampling multiple trajectories before selecting an answer. For Vision-Language Models (VLMs), answer frequency and model confidence alone can be insufficient, as sampled generations also vary in their engagement with the visual input. We find that distributional confidence and visual attention are complementary indicators of rollout quality. Each is limited on its own, while trajectories that show strong signals in both are more likely to be correct. Motivated by this, we introduce VisConf, a rollout-quality score that multiplicatively combines visual engagement with candidate-calibrated Self-Certainty, penalizing rollouts that are weak in either signal. We incorporate Rank-Weighted Consensus (RWC), a tuning-free aggregation rule that converts within-question VisConf ranks into answer-level vote weights. Across three VLMs on MathVista, MMMU-Pro, and MMStar, VisConf with RWC consistently outperforms strong BoN baselines. Our analysis shows particularly large gains when correct trajectories are present but underrepresented in the candidate pool.

[117] Write Once, Query Many: Evolving Compact Reusable Memory for Long Videos cs.CV | cs.LGPDF

Kerui Chen, Jinglu Wang, Yun Wang, Xinjiang Wang, Yan Lu

TL;DR: Video-EMem以时间锚定的文字和稀疏图像建立可进化长视频记忆，将一次视频写入成本摊到多次查询，并分别优化写入保真度和读取问答。

Details

Motivation: 长视频超过多模态大模型的上下文容量，逐个查询进行智能搜索会重复支付视频处理和定位成本；可复用记忆既要节省存储，也要保留构建时未知问题所需的信息。

Result: Video-EMem使用跨时间尺度的感知观察，结合紧凑自由文本和稀疏图像，保留语义与互补视觉细节，并支持不依赖固定语义模式的时间导航。Writer一次性整合重叠的多尺度观察，Reader只从生成记忆中检索和组织证据，底层模型保持冻结。工作流演化将写入器的评估与具体查询无关，关注源内容保真和覆盖；读取器则按证据定位与利用中的问答失败优化。在三个长视频基准上，进化后平均准确率比未进化流程高7.2%。相较原视频平均节省约99.86%存储；与直接输入2,048帧相比，Video-EMem使Qwen3-VL-30B在三项长视频问答基准平均提高2.6%。使用与先进智能体方法相同的o3答案模型时，平均准确率提高2.5%，且单次查询延迟更低。

Insight: 将写入质量和读取利用分开优化，有助于记忆针对开发查询之外的问题泛化；一次编码、多次查询可摊薄成本。摘要未给出记忆构建时间及各基准逐项数值，存储缩减不代表全部推理成本同比下降。

Abstract: Long videos often exceed the context capacity of multimodal large language models (MLLMs), while agentic search repeats costly video processing and grounding for each query. Reusable memory can amortize this cost, but must preserve information for questions unknown during construction. We propose Video-EMem, an evolvable video memory harness built on time-anchored perceptual observations. This representation combines compact free-form text with sparse images to preserve semantic content and complementary visual details across temporal scales, supporting temporal navigation without a fixed semantic schema. A Writer encodes each video once by consolidating overlapping multiscale observations, and a Reader retrieves and organizes evidence solely from the resulting memory. With the underlying models frozen, we evolve their workflows and operators using distinct feedback: query-independent source fidelity and coverage for the Writer, and QA failures in evidence localization and utilization for the Reader. Experiments support this separation for generalization beyond development queries, with our evolution strategy improving average accuracy by 7.2% over the unevolved harness. While reducing storage by about 99.86% on average relative to raw videos, Video-EMem improves Qwen3-VL-30B by an average of 2.6% across three long-video QA benchmarks over direct input of 2,048 frames. Using the same o3 answer model as the state-of-the-art agentic method, it improves average accuracy by 2.5% with lower per-query latency.

[118] Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing cs.CV | cs.AI | cs.ROPDF

Yujia Yang, Quan Yuan, Jiawei Fan, Yang Li, Tiange Fu

TL;DR: Scene-Adaptive VLA根据场景复杂度与模型决策偏差动态路由激活层，在Bench2Drive上降低推理计算开销约53.5%，且驾驶和语言能力仅有轻微退化。

Details

Motivation: 用于自动驾驶的视觉语言动作模型参数量大，车载实时推理消耗算力与能量并影响续航。已有高效VLA驾驶方法多调整推理顺序或token，结构层面的场景自适应尚未充分探索。

Result: Scene-Adaptive VLA利用道路场景时间连续性和模型预测不确定性，以帧级场景变化及决策偏差表征复杂度，并编码为场景感知token；两个专用查询经预算条件层路由，确定激活参数比例及所选层组合，整体端到端优化。在Bench2Drive上适配先进VLA后，推理时计算开销减少53.5%，驾驶和语言能力仅有轻微退化。

Insight: 按实时场景决定使用哪些网络层，可从模型结构层面分配自动驾驶推理预算。摘要未给出具体驾驶指标、不同计算预算下的性能曲线和硬件延迟，故53.5%指计算开销而非一定等比例降低端到端延迟。

Abstract: Vision-Language-Action (VLA) models offer a promising paradigm for autonomous driving. However, the massive parameter scale of VLAs strains onboard computational resources during real-time inference. This directly increases energy consumption and limits vehicle battery range, making it critical to minimize computational overhead without compromising driving capability and explainability. Current research on efficient VLA-based driving focuses on sequential adaptation (\textit{e.g.}, Chain-of-Thought and token adjustment), leaving structural adaptation underexplored despite parameter redundancy across dynamic driving scenes. To address this gap, we propose Scene-Adaptive VLA, an efficient framework leveraging dynamic layer routing to adjust active parameters based on real-time scenes. Inspired by the temporal continuity of scenes and model predictive uncertainty, our approach characterizes model-perceived scene complexity via frame-level temporal scene variations and decision deviations. To uncover latent scene-to-parameter correspondences, we encode this characterization into scene-aware tokens, alongside two specialized queries. A budget-conditioned layer routing mechanism evaluates these refined queries to allocate a budget that determines the active parameter ratio and subsequently selects appropriate layer combinations. The entire framework is optimized end-to-end. Extensive evaluations on the Bench2Drive benchmark show that when adapted to a state-of-the-art VLA model, our approach significantly reduces inference-time computational overhead by 53.5% with negligible degradation in driving and language capabilities. Code is available at https:github.com/yang-luvu/Scene-Adaptive-VLA.

[119] Dynamic self-ensemble on multi-exit residual vision transformer for small-data image classification cs.CVPDF

Stephen Ekaputra Limantoro, Yan-Kai Wang, Ching-Chun Huang, Hong-Han Shuai, Wen-Huang Cheng

TL;DR: ReSEViT以多阶段残差ViT和动态自集成利用中间层特征，目标是在小数据图像分类中减少单一分类器导致的错误和方差。

Details

Motivation: 视觉Transformer通常需要大规模数据才能取得强图像分类表现，而医疗影像等应用可能只有有限标注。只依赖单个分类器也可能无法适应不同层级特征形成的复杂决策边界。

Result: ReSEViT将多阶段残差ViT与自集成学习结合，在同一模型中集成多个阶段分类器，并通过动态学习器选择DLS自适应调整各阶段重要性。每个阶段以类似ResNet的残差token嵌入建模有效局部结构，并保留Transformer模块。作者认为残差token与多阶段集成可减少小数据分类中的冗余、预测错误和方差，同时维持模型规模效率。

Insight: 多层特征的动态组合可为小样本分类提供比单一输出头更稳健的决策。摘要未提供数据集、准确率、标注规模和对比结果，当前只能确认架构设计和作者对预期效果的概述。

Abstract: Vision Transformer (ViT) has garnered increased attention for image classification in recent years. However, ViT is notorious for its data-hungry problem as it achieves high-performance results when only trained on large-size datasets. In this paper, we study how ViT specifically works on small-scale datasets with limited data where the availability of image data is often constrained in practice, especially in fields such as medical imaging. To mitigate the issue, we designed a multi-stage residual ViT, dubbed ReSEViT, with the self-ensemble method to utilize multi-level information from intermediate features. Due to the complex decision boundaries for classifications, relying on a single classifier can lead to prediction errors. Hence, we propose a new learning method for ViT, named self-ensemble learning, which ensembles multi-stage classifiers in a single model. Our self-ensemble comprises a novel approach, Dynamic Learner Selection (DLS), to adaptively select the importance of different stages. ReSEViT incorporates residual token embeddings which is a ResNet-like block to model effective local structures alongside with transformer blocks in each stage. The integration of residual tokens and self-ensemble learning into a multi-stage ViT highlights the reduction in redundancy, prediction error and variance while maintaining an efficient model size for small-scale image datasets.

[120] SCOPE-VSR: A Broad-Coverage Training Dataset for Video Super-Resolution cs.CVPDF

Gengyu Liao, Yechun Yong, Zheng Chen, Jianze Li, Ziqing Zhang

TL;DR: SCOPE通过面向任务的视频筛选和分层语义重平衡构建约27K条开放域超分视频训练集，并提出一步视频恢复模型GraCo-VSR。

Details

Motivation: 开放域视频超分辨率需在场景和主体多样变化下产生时间连贯细节，但现有训练数据语义分布不均、覆盖有限，限制模型泛化。

Result: 作者提出SCOPE结构化数据整理流水线，将任务导向的视频筛选与分层语义重平衡结合，构建SCOPE-VSR训练集和SCOPE-test测试集；训练集含约2.7万段视频，覆盖多种真实场景及主体。配套GraCo-VSR是单步视频恢复框架，使用可靠性意识设计和显式监督，减少退化造成的运动与纹理引导失真，使模型更有效利用多样训练内容。评测显示其重建保真度和感知质量表现较强。

Insight: 结构化筛选和语义平衡旨在解决开放域VSR数据分布瓶颈，可靠性机制则减少运动/纹理引导受损。摘要未提供具体指标、模型对比和测试集规模，因此“表现较强”无法量化。

Abstract: Open-domain video super-resolution (VSR) requires temporally coherent details across widely varying scenes and subjects. However, existing training datasets are often semantically imbalanced and insufficiently diverse. We address this data bottleneck with a structured curation pipeline for open-domain VSR, SCOPE. In this pipeline, we combine task-oriented video selection with hierarchical semantic rebalancing. Applying this pipeline to a large-scale collection of data sources, we construct the SCOPE-VSR dataset for training and SCOPE-test for testing. SCOPE-VSR has approximately 27k videos and provides balanced and diverse coverage of various real-world scenarios and subjects. We further develop GraCo-VSR, a novel framework for one-step video restoration. GraCo-VSR adopts a reliability-aware design with explicit supervision to reduce degradation-induced distortions in motion and texture guidance, allowing the model to benefit more effectively from the diverse training content in SCOPE-VSR. Evaluations show that our model achieves strong reconstruction fidelity and perceptual quality. Our code is available at: https://github.com/gyliao2768/SCOPE-VSR.

[121] IntriAdapt: Generalizable Intrinsic-Guided Test-Time Adaptation for Metric Geometry Estimation cs.CVPDF

Jiaxing Chen, Zizheng Yan, Jingyu Zhuang, Jinwei Chen, Shuguang Cui

TL;DR: IntriAdapt把已知相机内参作为几何一致性验证信号，在冻结几何基础模型上优化轻量LoRA适配器以改善度量几何估计。

Details

Motivation: 从图像恢复度量三维几何本质上受透视投影歧义影响。现有模型推理时使用给定或预测内参，却通常不以这些参数监督适配；测试时适配则常需外部深度测量。

Result: IntriAdapt比较预测点云的视线方向与已知内参对应视线之间的差异，以闭式几何残差衡量不一致，并在冻结几何基础模型上优化轻量LoRA适配器，使预测符合相机模型，无需深度传感器或外部先验。方法也扩展至视频：在短序列上适配后，可迁移到同一相机拍摄的其他序列；测试图像通过前馈处理，不必逐帧重新优化。在十个单图像和七个序列基准、多个基础模型上的实验显示，其度量精度和相对几何均持续超过未适配模型及使用真实内参的对照。

Insight: 相机内参可作为内生几何校验信号，替代额外深度监督进行测试时适配；短序列适配可复用于同相机数据。摘要未给出各基准数值、未知/变化内参效果和校准成本。

Abstract: Recovering metric 3D geometry from images is fundamentally ill-posed due to perspective projection ambiguity. Current geometry models incorporate supplied or predicted intrinsics during inference, but do not use them to supervise model adaptation. To refine these predictions, test-time adaptation (TTA) typically relies on external depth measurements. In this paper, we instead use camera intrinsics as an output-side geometric verifier. Comparing the point cloud’s viewing rays with those defined by the known intrinsics yields a direct measure of geometric inconsistency. We propose \textit{IntriAdapt}, a TTA framework that minimizes this closed-form geometric residual. By optimizing a lightweight LoRA adapter on a frozen geometry foundation model, our method aligns the prediction with the specific camera model without depth sensors or external priors. We further extend this approach to video sequences, demonstrating that adapting a short sequence generalizes to other sequences captured by the same camera. The adapted model processes held-out images in a feed-forward pass, without repeating optimization for each frame. Experiments on ten single-image and seven sequence benchmarks across multiple base models show that \textit{IntriAdapt} consistently outperforms both unadapted and oracle-intrinsics baselines in metric accuracy and relative geometry.

[122] From Views to Worlds: Active Exploration over 3D Worlds for Vision-Language Models cs.CVPDF

Qijian Tian, Jiayu Ying, Ke Fan, Lizhuang Ma, Xin Tan

TL;DR: V2W让视觉语言模型主动探索由多视角图像重建的三维世界，通过交互获取相机姿态、几何和语义空间证据以改进空间推理。

Details

Motivation: 被动的二维视角理解限制视觉语言模型空间推理；判断三维关系常需主动获得当前画面之外的空间证据，如相机位置和跨视角关系。

Result: V2W以多视角图像为输入，先重建包含几何、相机姿态和语言对齐语义的显式三维世界，再通过视觉及语言接口提供给模型。模型与这些接口多轮交互，逐步采集空间证据，从视角中心感知转向世界中心推理。方法可免训练地接入现有VLM，也可通过强化学习优化探索策略。空间心智建模基准实验显示，免训练设置下V2W显著提升VLM基线；学习探索策略后达到先进表现。

Insight: 主动探索使模型能请求并核查二维画面未直接呈现的几何和视角关系。摘要未给出基准名称、分数、三维重建误差和交互成本；效果依赖输入多视角质量及重建世界准确性。

Abstract: Vision-language models (VLMs) have achieved remarkable progress in 2D visual understanding tasks, yet the passive view-centric perception paradigm limits their capacity for 3D spatial reasoning. Such reasoning often requires actively acquiring spatial evidence beyond the currently visible views, such as camera poses and cross-view spatial relationships, to infer 3D relationships that are not explicitly represented in 2D images. However, existing methods either lack active exploration or merely provide additional views during inference, without explicitly acquiring such spatial evidence beyond 2D images. In this paper, we propose View-to-World (V2W), an active exploration framework that enables VLMs to actively explore 3D worlds and lifts VLMs from passive view-centric perception to active world-centric spatial reasoning. Given a spatial reasoning task with multi-view images, V2W first reconstructs an explicit 3D world with geometry, camera poses, and language-grounded semantics, and exposes it to VLMs through visual and linguistic interfaces. Through multi-turn interaction with these interfaces, VLMs iteratively acquire spatial evidence from the constructed 3D world, enabling active world-centric spatial reasoning. V2W can be integrated with existing VLMs in a training-free manner and can be further enhanced by optimizing the exploration policy via reinforcement learning. Experiments on spatial mental modeling benchmarks demonstrate that V2W substantially improves VLM baselines in the training-free setting and achieves state-of-the-art performance with the learned exploration policy, highlighting the importance of active exploration for spatial intelligence. Project page: \url{https://fangzhou2000.github.io/projects/v2w}.

[123] MESS: Multi-Exposure Sequence Synthesis for Generalizable Image Enhancement cs.CVPDF

Ruodai Cui, Shuaizheng Liu, Rongyuan Wu, Lei Zhang

TL;DR: MESS从高位深RAW图像模拟不同曝光序列，再训练扩散模型将普通8位RGB图像扩展为多曝光样本，构建50万对训练数据支持多场景图像增强。

Details

Motivation: 低照度增强、曝光校正和逆光增强模型跨场景泛化有限，主要瓶颈是高质量、丰富多样的配对训练数据难以采集。既有合成方法也难以覆盖复杂光照和相机成像流程。

Result: MESS先通过模拟ISP流水线，对高位深RAW图像渲染不同曝光，构造多曝光序列作为监督；再训练扩散式RGB到多曝光序列生成器，从单张8位RGB扩展到野外普通图像。每张源图随机抽取一种合成曝光作为退化输入，原高质量RGB作为目标。作者构建50万对训练样本并训练轻量增强网络，报告其在多种图像增强任务上的泛化能力明显超过已有模型。代码和数据公开。

Insight: 先从RAW获取成像合理的多曝光监督，再扩展到普通RGB数据，可缓解高质量配对采集成本。摘要未给出具体任务指标、RAW数据规模及不同相机ISP迁移结果，泛化提升以作者总体实验报告为限。

Abstract: Despite the significant progress in image enhancement under adverse illuminations, such as low-light image enhancement, exposure correction, and backlit image enhancement, existing models show limited generalizability to different scenarios. A major bottleneck lies in the lack of large-scale and diverse training data, as these tasks typically rely on manually captured, precisely aligned image pairs that are expensive to scale up. Although synthetic data have been explored as an alternative, existing synthesis methods are limited in modeling complex adverse illumination conditions and camera imaging pipelines. To address this issue, we propose a novel framework to synthesize realistic training pairs for generalizable image enhancement in the wild. Specifically, we first derive multi-exposure sequences (MES) from high-bit RAW images by rendering them under different exposure settings through an emulated ISP pipeline. Using these RAW-derived MES as supervision, we train a diffusion-based RGB-to-MES generator to synthesize plausible exposure sequences from a single 8-bit RGB image, extending the synthesis pipeline from limited RAW collections to large-scale in-the-wild RGB data. For each source image, we randomly sample one synthesized exposure variant as the degraded input and use the original high-quality RGB image as the target. With the proposed MES Synthesis (MESS) approach, we construct a dataset of 500K training pairs and train a lightweight network, which demonstrates significantly better generalization performance than existing models across various image enhancement tasks. Code and data are available at https://github.com/PolyU-VCLab/MESS.

[124] CorrFlow: Self-Correcting Flow Matching cs.CV | cs.AIPDF

Hangwei Zhang, Nathan Rouillé, Chuanxia Zheng

TL;DR: CorrFlow在流匹配采样中使用轻量修正器，根据基础模型同次前向的干净预测、条件、噪声和隐藏状态预测速度残差。

Details

Motivation: 流匹配模型反复预测速度生成样本，但通常不会利用产生这些预测的中间表示重新检查和修正当前预测。

Result: CorrFlow基于整流流插值下单步干净预测误差等于速度残差按噪声水平缩放这一关系，训练修正器根据基础模型干净预测、条件、噪声水平及同次前向隐藏状态预测残差。每步将修正加入基础速度，再进行采样更新；无需额外运行基础模型，每步仅增加一次轻量修正器前向。理论分析给出特征复用优于原始输入仿射预测器的条件，并表明在匹配训练计算量下，训练修正器比继续训练基础模型能更大幅度降低残差风险。类别条件图像生成、文生视频和图生视频实验显示，相比相应基础模型，视觉质量及物理一致性均有提升。

Insight: 复用已有隐藏状态可让流匹配模型自校正，而不增加基础模型调用。摘要未提供具体质量指标、额外推理延迟和理论条件范围，不能判断所有采样器和任务都获得相同收益。

Abstract: Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information through a lightweight refiner. Under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level. Motivated by this identity, the refiner learns to predict residual corrections from the base model’s clean prediction, conditioning, noise level and hidden states from the same forward pass. The correction is added to the base velocity before each sampling update, so subsequent predictions are made from the corrected state. This requires no additional base-model evaluations and only one lightweight refiner evaluation per step. Our theoretical analyses establish conditions under which feature reuse outperforms raw-input affine predictors and refiner training achieves greater residual-risk reduction than continued base-model training under a matched training compute budget. Experiments on class-conditional image generation, text-to-video generation and image-to-video generation show consistent improvements over the corresponding base models, including gains in visual quality and physical consistency. Project Page: https://fogradio.github.io/corrflow-project

[125] Design Anything, Design at Scale cs.CVPDF

Xuchao Ren, Siyuan Feng, Jing Huang, Yanhua Yao, Jing Wang

TL;DR: Design Anything通过外置意图—执行图和确定性控制器保存跨多产物设计项目的要求与决策，提升多阶段设计任务约束满足率。

Details

Motivation: 设计项目常由明确的目标、约束和交付物组成，后续工作依赖早期选择；长流程智能体难以保证各产物在同一项目内持续符合设计意图。

Result: 该图结构框架将设计意图保存在模型之外，用意图—执行图把交付物绑定到生成它的工作流步骤。任务契约保留约束和决策，选定设计显式作为后续生成输入；确定性控制器依据绑定和执行记录检查输出类型及数量、追踪未完成交付并指导后续回合。DesignAnything-Bench含19个领域的228项任务，涵盖38个项目系列及6、14、24个产物规模。固定LLM时，DeepSeek V4.1 Flash High和GPT-5.6 Sol High分别满足85.0%和92.5%的事实约束，超过同模型最强基线15.4%和14.4%；24产物任务优势分别为25.1%和23.0%。18/19领域的约束满足率领先或持平；专业设计师盲评也偏好其完整交付集。

Insight: 将任务状态、决定和产物依赖显式存储，可减少长项目中意图漂移及漏交付。摘要未说明盲评人数、偏好效应量和不同领域的细节，但基准规模覆盖多种设计任务。

Abstract: As multimodal resources are transformed into media outputs, agentic AI is expected to stay aligned with design intent over long horizons. Designers follow briefs specifying goals, constraints, and deliverables, with later work building on earlier decisions. Recent work reuses design experience across tasks, yet alignment among artifacts within one project remains challenging. Agents therefore need to preserve and apply this intent across domains and as projects grow. We introduce Design Anything, a graph-based harness that preserves design intent outside the model and binds deliverables to the workflow steps that produce them through an intent-execution graph. A task contract retains constraints and decisions, selected designs become explicit inputs to later generations, and a deterministic controller uses bindings and execution records to check output types and counts, identify outstanding deliverables, and guide subsequent turns. We also introduce DesignAnything-Bench, with 228 tasks across 19 design domains, including 38 project families at scales of 6, 14, and 24 artifacts. With the LLM held fixed, Design Anything satisfies 85.0% and 92.5% of the brief’s factual constraints with DeepSeek V4.1 Flash High and GPT-5.6 Sol High, respectively, exceeding the strongest same-model baselines by 15.4% and 14.4%. These gains reach 25.1% and 23.0% on 24-artifact tasks. Under each model, Design Anything leads or ties in constraint satisfaction in 18 of 19 domains. Professional designers also prefer its complete deliverable sets over those of same-model baselines in a blind evaluation.

[126] ATLAS: Learning to Adapt LVLM Attention for Reasoning Video Segmentation cs.CVPDF

Hamidreza Dastmalchi, Ali Cheraghian, Aijun An, Abdelwahed Khamis, Lars Petersson

TL;DR: ATLAS学习把视觉语言模型分散、噪声较大的跨层注意力转化为物体中心定位热图，再用SAM2传播到视频帧，实现推理视频分割。

Details

Motivation: 推理视频对象分割需根据自然语言描述在整段视频中定位目标。大型视觉语言模型注意力中隐含定位信号，但跨层、头和模态分散且噪声大，不能直接用于分割；手工选择注意力启发式或细化计算成本高。

Result: ATLAS将所有层、注意力头和模态的注意力作为结构化特征，融合保留空间信息的高分辨率帧级注意力与编码时间上下文的低分辨率视频级注意力。轻量U-Net使用时间注意力和FiLM调制，生成粗分割热图，再通过SAM2在视频中传播。在具有挑战性的指代表达和推理视频分割基准上，ATLAS超过现有冻结骨干方法。

Insight: 学习融合注意力而非手工挑层可更有效地利用VLM内部定位线索，并将粗定位转成时序掩码。摘要未列出基准名称、数值成绩和视频传播误差，无法评估推理成本及长视频漂移。

Abstract: Reasoning Video Object Segmentation (RVOS) aims to localize and segment objects referred to by natural-language expressions throughout a video. Although large vision-language models (LVLMs) implicitly encode localization cues in their attention maps, these attentions are often noisy, diffuse, and scattered across layers, heads, and modalities, limiting their direct use for segmentation. Existing methods typically rely on handcrafted attention-selection heuristics or computationally expensive attention refinement. We propose ATLAS, an attention adaptation framework that learns to transform raw LVLM attentions into object-centric localization cues. ATLAS treats attentions from all layers, heads, and modalities as structured features, combining high-resolution frame-level attentions that capture spatial information with low-resolution video-level attentions that encode temporal context. A lightweight U-Net equipped with temporal attention and FiLM-based modulation fuses these complementary representations to generate coarse segmentation heatmaps, which are propagated across the video using SAM2. Extensive experiments on challenging referring and reasoning video segmentation benchmarks show that ATLAS outperforms existing frozen-backbone methods. Code and models are available at \url{https://hamidreza-136.github.io/atlas/}.

[127] Video-Language Entailment for Video Similarity Estimation cs.CVPDF

Benedetta Liberatori, Lorenzo Vaquero, Elisa Ricci, Paolo Rota

TL;DR: VICE用双分支视频—文本蕴含分数估计概念条件视频相似性，使模型逐个分析视频与另一视频说明，而非同时输入两个视频比较。

Details

Motivation: 当前大型多模态模型主要按单视频目标训练，视频语义相似性估计仍困难。除了整体相似度，概念相似性还需比较活动、地点等特定概念，但尚缺少利用LMM显式进行跨视频推理的方法。

Result: VICE（Video Caption Entailment）采用双分支结构，每个分支联合处理一个视频和另一个视频的描述，按指定概念计算视频—文本蕴含匹配，再融合两个方向的输出得到相似分数，并生成文字解释。该设计复用模型原有单视频理解能力，支持细粒度视频与文本间比较。在ConViS-Bench和StepDiff上，VICE超过同时处理两个视频的模型，显示单视频推理式比较可用于概念相似估计。

Insight: 把跨视频比较转化为两次单视频—文本蕴含判断，可利用现有单视频理解能力进行概念级相似度推理。摘要未提供指标数值、基线模型清单及不同概念表现，文字解释质量也未单独评估。

Abstract: Comparing videos to assess their semantic similarity remains a challenging task for current Large Multimodal Models (LMMs), which are primarily trained with single-video objectives. While most prior works have focused on computing global similarity, recent studies have introduced the more challenging problem of concept-based similarity, which measures how similar two videos are with respect to specific concepts (e.g., activity, location). However, no dedicated approaches have yet leveraged LMMs to explicitly address this task. In this paper, we aim to fill this gap introducing Video Caption Entailment for video similarity estimation (VICE): a dual-branch design that leverages LMMs for cross-video reasoning. Each branch jointly processes one video and the caption of the other, measuring their alignment for a specific concept via video-language entailment. The outputs from both branches are fused to produce a final similarity score, along with a textual explanation of the model reasoning. This approach leverages the model’s native single-video understanding while enabling fine-grained comparative reasoning between video and text. Experiments on ConViS-Bench and StepDiff show that VICE outperforms models that process both videos simultaneously, highlighting the promise of single-video reasoning for concept-based similarity estimation.

[128] UniAVC: Unified Audio-Video Creation via Progressive Training and Asymmetric Scheduling cs.CV | cs.MM | cs.SD | eess.ASPDF

Kien T. Pham, Yuhao Liu, Chong Hou Choi, Zhenqi He, Yanghao Wang

TL;DR: UniAVC先用充足单模态数据训练音频和视频生成/编辑能力，再用配对数据学习跨模态耦合，并以独立噪声调度支持非对称生成和编辑。

Details

Motivation: 统一音视频创作要求一个模型完成文本条件联合合成、双向跨模态生成和指令编辑，但联合生成和编辑的配对数据稀缺。若直接用稀缺配对数据从头学单模态能力会浪费监督；两模态共用噪声日程也忽视视频与音频信息密度及去噪节奏差异。

Result: UniAVC先分别用单模态数据把每个分支训练到具备完整生成和编辑能力，再将配对数据用于跨模态对应学习。音频与视频共享时间步但各自使用解耦噪声调度，使空间和运动结构密集的视频可在较长高噪声阶段组织，而较稀疏音频可更早稳定；任务还让一种干净模态直接监督另一模态生成或编辑。设计支持意图感知编辑：保持一种模态不变并用其约束另一种；也支持推理后自我细化，仅重生成不满意分支。UniAVC从单套权重覆盖多种创作任务，无需任务专用适配器。

Insight: 分阶段训练把宝贵配对数据留给跨模态关联，模态特定时间表则处理音视频去噪动态不对称。摘要未给出任务基准、生成质量分数和训练成本，方法效果目前以架构功能陈述为主。

Abstract: Unified audio-video creation requires a single model capable of text-conditioned joint synthesis, bidirectional cross-modal generation, and instruction-guided editing across both modalities. Existing methods rely heavily on paired corpora, which are scarce for joint generation and even rarer for editing. Consequently, they are spent learning single-branch generative and editing capabilities that abundant unimodal data could readily teach, rather than on the cross-modal coupling. Furthermore, prevailing systems apply a single shared noise schedule to both modalities, neglecting two critical limitations. First, due to the significant discrepancy in information density, video requires prolonged high noise to organize spatial and motion structure, whereas the sparser audio signal can settle earlier during the synthesis process, the dynamics a uniform schedule might compromise. Second, lockstep denoising leaves asymmetric workflows, such as conditioning on one modality to synthesize or edit another, untrained. We present UniAVC, which addresses both obstacles. It first trains each branch on unimodal dataset to full generative and editing competence, and subsequently dedicates paired data to exploit cross-modal correspondence. Modality-decoupled schedulers meanwhile warp the noise schedule independently from a shared timestep, while subtasks with one clean modality directly supervise asymmetric configurations. These settings respectively enable intent-aware editing, which preserves one modality intact while enlisting it to constrain the other, and post-inference self-refinement, which regenerates only an unsatisfactory branch against an accepted one. UniAVC thus spans multiple creation tasks from a single set of weights without task-specific adapters.

[129] SwitchSplat: Learning Discrete Object States from Casually Captured Scenes cs.CV | cs.GRPDF

Yizheng Xie, Yalda Foroutan, George Shramko, Daniel Rebain, Shrisudhan Govindarajan

TL;DR: SwitchSplat用共享静态底景与每个可交互物体的二元状态槽表示高斯场景，使模型可从随手拍摄视频学习物体开关状态并组合未见配置。

Details

Motivation: 3D Gaussian Splatting重建通常假设世界静态，但随手拍摄常包含开门、拨开关或移动物体等离散状态改变。时间条件动态方法可能把“何时观察”与“场景处于何状态”混在一起，只能重放见过序列而无法组合新状态。

Result: SwitchSplat将场景分成共享静态底座和独立对象槽；每个槽存放一个可交互对象的两种二元状态。高斯到槽的分配完全可微，因此模型可直接从视频学习状态几何，无需人工标注。推理时用户可自由组合不同槽状态，并从任意视角渲染结果。在新建的随手拍摄数据集上，SwitchSplat的新视角合成质量超过静态及动态基线，并能渲染采集时未出现过的状态组合。

Insight: 显式离散状态分解使可编辑3D场景从时间序列重放迈向对象状态组合。摘要没有提供数据集规模、渲染指标及可编辑对象种类，复杂连续动作和多状态对象是否支持尚不清楚。

Abstract: 3D Gaussian Splatting has made high-fidelity 3D reconstruction fast and accessible, yet existing representations assume a static world. This assumption breaks down in casual captures, where a user naturally interacts with the environment – opening doors, toggling switches, or moving objects – introducing discrete state changes that no current framework can model. Time-conditioned dynamic methods such as 4DGS conflate when a state was observed with what state the scene is in, limiting them to replaying seen sequences rather than composing arbitrary object configurations. We present SwitchSplat, the first 3DGS representation to address this gap via a base+slot decomposition: a shared static base captures the invariant scene, while each slot independently holds the two binary states (on/off) of one interactable object. Because the assignment of Gaussians to slots is fully differentiable, the model discovers the geometry of each state directly from video, without manual annotation. At inference, the user can freely combine slot states and render the resulting scene from any viewpoint. We validate SwitchSplat on a new casual-capture dataset and show that it outperforms static and dynamic baselines in novel-view synthesis quality, while uniquely supporting rendering of state combinations never seen during capture. Project page: https://switchsplat.github.io

[130] VTF900: A 900-Subject Visible-Thermal Face Recognition Dataset and Benchmark cs.CVPDF

Jacob Abaare, Bianca Ianetta, Cedric Nimpa Fondje, Shuowen Hu, Benjamin S. Riggan

TL;DR: VTF900提供900名受试者的可见光—热红外配对人脸识别基准，支持大图库验证和识别，并显示小规模图库成绩可能无法预测部署规模表现。

Details

Motivation: 可见光—热红外人脸识别的现有基准通常只有约100人图库，难以评估大规模识别能力或提供运营相关的性能统计。

Result: VTF900包含900名受试者及覆盖表情、姿态、遮挡变化的可见光和热红外配对图像，支持1:1验证和1:N闭集识别，并专门考察跨光谱、跨数据集条件。作者在ARL-VTF上训练后，对卷积基线、视觉Transformer和域自适应网络进行基准分析。全900人图库、正面条件下ViT-B达到97.68% AUC和51.10% TAR@1% FAR；离姿态条件下DPIT-ResNet50等域自适应方法更稳健，达到94.68% AUC和36.08% TAR@1% FAR。结果显示，小图库高AUC不必然转化为部署规模可靠识别，姿态变化会放大退化。

Insight: 大规模图库和离姿态、跨光谱评测可揭示小样本基准掩盖的性能退化。摘要未报告受试者人口构成、图库构建细节和隐私控制，也没有覆盖真实监控环境误识风险。

Abstract: Despite advancements in visible-thermal face recognition, where thermal probe (query) face images are matched against visible enrollment templates, existing benchmarks are currently limited by small gallery sizes (N $\approx$ 100 subjects). This limitation inhibits the ability to accurately evaluate visible-thermal facial recognition at scale and to report significant recognition performance metrics, which have better operational relevance. Therefore, we introduce VTF900, a large-scale visible-thermal face recognition benchmark with 900 subjects and paired imagery that includes expression, pose, and occlusion variations. VTF900 is designed to support large-scale (i.e., large gallery) evaluations across verification (1:1 matching) and closed-set identification (1:N matching) for visible-thermal face recognition. Furthermore, VTF900 specifically addresses the challenging combination of cross-spectral and cross-dataset analysis. In this paper, after training models on the ARL-VTF dataset, we provide benchmark analyses across convolutional (baseline), vision transformers, and domain-adaptive neural network architectures to assess how these methods scale. At full gallery scale (N = 900), transformer models achieve strong discriminative performance under frontal conditions (e.g., ViT-B: 97.68% AUC, 51.10% TAR@1%FAR), while domain-adaptive methods such as DPIT-ResNet50 demonstrate greater robustness under off-pose conditions (94.68%AUC, 36.08% TAR@1%FAR). Our results show that high AUC at small scale does not translate to reliable performance at deployment scale, and that pose variation significantly amplifies performance degradation. VTF900 exposes these limitations, enabling more rigorous evaluation of cross-spectral face recognition systems.

[131] SpriteStage: Long Horizon Animation Generation via Progressive Scene Programming and Replay cs.CV | cs.AIPDF

Zhaoqi Wang, Zijian Zhang, Jiamou Liu

TL;DR: SpriteStage通过逐场景编程、执行反馈和引擎重编排生成长篇动画，使后续场景基于前序已接受的世界状态与叙事结果。

Details

Motivation: 长动画需跨场景协调动作、对白、镜头覆盖和剧情发展。若一次性写完整故事后才执行，后续场景可能依赖前面实际未发生的事件；可执行场景程序则让这些决定可检查。

Result: SpriteStage在逐步演绎与引擎重编排间循环：每个场景从已接受的世界检查点和叙事快照生成，放在共享空间舞台执行，并依据运行及视觉反馈修订相关动作或故事节点；接受的结果成为下一场景起点。引擎重编排将草稿映射到制作资产和引擎能力，同时保留事件、几何和时间。对整体序列的检查及冻结回放将已接受表现带入最终视频。三部约9.9至14.3分钟的完整动画展示了该框架通过场景编程和引擎执行生成长篇叙事的效果。

Insight: 逐场景执行并在接受后推进，可减少后续剧情依赖未发生事件的风险，也支持审查和修订。摘要以三部作品展示框架可行性，未提供自动评价指标、人工制作成本或与端到端视频生成基线的比较。

Abstract: Video generation models synthesize visual sequences from text and image inputs. Extending video generation to long animations requires actions, dialogue, camera coverage, and narrative developments to remain coordinated across scenes. Executable scene programs make these decisions available for inspection. Constructing an entire story before observing its execution can leave later scenes dependent on unrealized events. We present SpriteStage, a progressive scene programming framework that builds animation through iterative enactment and engine reorchestration. Iterative enactment generates each scene from an accepted world checkpoint and narrative snapshot, executes it on a shared spatial stage, and uses runtime and visual feedback to revise the relevant actions or story beats. The accepted outcome becomes the starting state for the next scene. Engine reorchestration maps the resulting draft to production assets and engine capabilities while retaining its events, geometry, and timing. Inspection of the assembled sequence and frozen replay carry the accepted performance into the final video. Three complete animations, ranging from approximately 9.9 to 14.3 minutes, demonstrate the framework’s effectiveness in producing extended narratives through scene programming and engine execution.

[132] Multi-Speed Field Evaluation of a Semi-Autonomous Precision Sprayer for Weed Control in Vegetable Crops cs.CVPDF

Md Khairul Islam, Boyang Deng, Yuzhen Lu, Daniel Brainard, Ajit Srivastava

TL;DR: 该研究在真实蔬菜田评估半自主视觉喷药系统在三种行驶速度下的检测、跟踪、喷射和作物损伤表现，发现速度上升显著削弱实际控制效果。

Details

Motivation: 机器视觉定点喷药可减少除草剂使用并保护作物，但田间速度如何影响检测—跟踪—喷射的完整闭环尚未充分评估。

Result: 作者构建五季数据集，包含10,599张图像、84,311个实例及七类目标，并将YOLOv11-medium、ByteTrack和12喷嘴定制喷药器部署在半自主电动拖拉机上，针对萝卜田在0.13、0.27和0.45米/秒速度测试。最低速度下检测mAP@50为94.6%，召回率93.2%；速度增加后除草目标命中率由78.3%降至35.0%，作物接触率则由20.1%升至28.7%。系统未达到预设商业可行阈值。

Insight: 单看目标检测准确率不足以证明精密喷药可用，运动下的感知、喷射时机补偿及靠近作物时的执行精度都需改进。摘要明确说明当前系统不满足商业可行性门槛，结果限于三种低速田间设置。

Abstract: Machine vision-guided precision spraying is promising for sustainable vegetable production by enabling site-specific weed targeting, reducing herbicide application while protecting crop yield and quality. However, the effects of operating speed of sprayer systems on the integrated detection-tracking-actuation pipeline remain insufficiently characterized under field conditions. This study evaluated a semi-autonomous, tractor-based spraying system for targeted weed control in radish and quantified performance across three operating speeds. A five-season dataset comprising 10,599 images and 84,311 annotated instances across seven classes was curated for model development. A YOLOv11-medium model was integrated with ByteTrack tracking and a customized 12-nozzle sprayer on a semi-autonomous electric tractor. Field trials were conducted at 0.13, 0.27, and 0.45 m/s (i.e., 0.3, 0.6, and 1.0 mph) to characterize speed-dependent changes in weed detection, targeting, actuation, and crop damage. At the lowest speed of 0.13 m/s, the system achieved 94.6% mAP@50 and 93.2% recall for weed and crop detection. However, increasing platform speed substantially reduced system performance, with weed hit rate declining from 78.3% to 35.0% and crop contact rates increasing from 20.1% to 28.7%. Consequently, the system did not meet predefined commercial-viability thresholds. These results demonstrate that practical deployment requires improvements beyond weed detection alone, particularly coordinated improvements in motion-robust plant perception, spray-timing compensation, and crop-adjacent actuation precision. The identified performance bottlenecks provide specific engineering targets for advancing vision-guided precision spraying toward practical, high-speed vegetable weed management. The software programs developed in this study have been made publicly available to support future precision weeding research.

[133] Meta Inverse Prompting for Video Generative Models cs.CVPDF

Yerin Jung, Jongheon Jeong

TL;DR: MVIP将视频逆提示建模为元学习，从给定视频快速恢复可控文本条件，避免逐视频昂贵迭代优化。

Details

Motivation: 视频生成和编辑通常是从条件到视频的单向流程，但媒体生产常需从已有视频反推出可控文本条件。图像逆提示不能直接扩展到视频，因为还要表达时间动态，并且逐视频优化计算成本高。

Result: Meta Video Inverse Prompting通过元学习在时间维度优化元提示，从而适配高效提示推断。方法摊销逆推理成本，测试时无需逐视频反复迭代优化。实验显示，MVIP恢复的逆提示质量更高、重建保真度显著优于基于图像的逆向方法和迭代字幕基线，同时效率更高。

Insight: 元学习可将视频条件反演成本从每条视频的逐例优化转移到模型训练阶段。摘要未给出数据集、量化保真度和速度指标，也未说明逆提示对不同视频生成器的迁移能力。

Abstract: Recent advances in video generative models have enabled flexible conditional generation and editing pipelines, such as text-to-video, image-to-video, and reference-guided synthesis. However, these workflows are inherently one-directional, mapping conditions to videos without providing a mechanism to invert this process. In complex media production settings, the ability to recover controllable textual conditions from a given video is highly desirable, yet remains underexplored. While inverse prompting has been studied for image generation, directly extending these approaches to video is challenging due to the need to capture temporal dynamics beyond static appearance, as well as the substantial computational cost of per-video optimization. In this work, we propose the first practical framework for efficient video inverse prompting. Specifically, we formulate inverse prompting as a meta-learning problem and introduce Meta Video Inverse Prompting (MVIP), which learns a meta-prompt optimized over temporal dimensions to enable fast and effective prompt inference. Our approach amortizes the inversion process, eliminating the need for costly iterative optimization at test time. Experimental results demonstrate that MVIP recovers higher-quality inverse prompts with significantly improved reconstruction fidelity compared to image-based inversion and iterative captioning baselines, while achieving substantial gains in efficiency. Code and video results are available at https://ueriniuno.github.io/mvip-page/.

[134] Visual Sufficiency Landscapes: Understanding Query-Conditioned Reliability in Visual Token Pruning cs.CVPDF

Zhenghan Li, Kexin Sun, Wei Zheng, Peng Hu, Xiao Wei

TL;DR: 该研究定义视觉充分性景观，测量在固定层和token预算下有多少视觉token子集可维持密集模型行为，并用其预测剪枝方法的失败风险。

Details

Motivation: 不同视觉token剪枝方法可能保留不同子集却获得相近表现，因此需要理解可成功子集的数量、该数量如何受问题影响，以及是否能预测某个剪枝方法会失败。

Result: 视觉充分性概率定义为均匀抽取的固定预算token子集中，同时保持密集模型答案、参考答案似然和输出分布的比例；独立重复抽样将输入差异与估计噪声分开。在Qwen2.5-VL-7B、InternVL2-8B和LLaVA-v1.5-7B上，GQA和TextVQA的输入间差异超过重复误差，同一GQA图像不同问题也呈现不同景观。充分性概率可改善对完整剪枝方法失败的预测：在LLaVA上加入该概率后，FastV、VisionZip和SparseVLM的AUC分别提高0.1151、0.0683和0.1401，同时降低log loss与Brier分数。若在原始图像上训练预测器，迁移到新图像时log loss分别降低9.0%、13.0%和14.1%。

Insight: 某图像可剪枝性不仅取决于模型和预算，也取决于具体查询；衡量成功子集丰度可识别高风险剪枝条件。摘要未给出各层、预算的完整景观和抽样成本，结论主要覆盖三种模型。

Abstract: Different visual token pruning strategies retain different subsets while maintaining similar task performance. This motivates a combinatorial question: how many visual subsets can preserve model behavior, and how does their abundance depend on the query? We introduce visual sufficiency landscapes to characterize this abundance. At a fixed layer and budget, visual sufficiency probability is the fraction of uniformly sampled subsets that jointly preserve the dense model’s answer, reference-answer likelihood, and output distribution. Independent sampling replicates separate input variation from estimation noise. Across Qwen2.5-VL-7B, InternVL2-8B, and LLaVA-v1.5-7B, differences between inputs on GQA and TextVQA exceed replication error, and different questions about the same GQA image produce distinct landscapes. Sufficiency probability also improves failure prediction for full pruning methods. On LLaVA, adding this probability increases AUC by 0.1151 for FastV, 0.0683 for VisionZip, and 0.1401 for SparseVLM, while reducing log loss and Brier score. Several architecture-adapted configurations on the other two models reproduce these gains. Under controlled selection rules, predictors trained on the original images reduce log loss on new images by 9.0%, 13.0%, and 14.1%, respectively. These results establish successful-subset abundance as a measurable dimension of local prunability and connect it to the identification of high-risk pruning conditions.

[135] ReTrace: Visual Backtracking for Efficient Multimodal Large Language Models cs.CV | cs.AIPDF

Guangzong Si, Dong Wang, Yifan Yu, Miao Liu, Panwang Pan

TL;DR: ReTrace在压缩视觉token时保留暂时停用的token，并在中间层视觉回溯窗口恢复其计算，避免将临时节省计算等同于永久删除信息。

Details

Motivation: 视觉token压缩常在浅层剪掉低显著性token，将计算暂时停用和永久信息删除混为一谈。但模型后续可能重新需要早期看似不重要的视觉上下文。

Result: 逐层注意力分析发现，视觉枢轴的位置较稳定，但中间层会主动从早期低重要度token中取回上下文。ReTrace据此引入休眠token状态，将非枢轴token缓存到昂贵层计算之外，在回溯窗口恢复；待信息被吸收后再剪枝。在Qwen2.5-VL和InternVL3上，实验显示FLOPs及KV缓存几乎减少90%，预填充加速2.83倍，并在复杂视觉语言基准上保持较强性能。

Insight: 视觉信息可能先暂不参与计算、随后又对推理有用；压缩系统需允许回溯再激活，而非一次性删除。摘要未给出各基准保留率和具体硬件配置，“近90%”限于报告模型及设置。

Abstract: Multimodal Large Language Models (MLLMs) encode images into dense visual token sequences, creating substantial attention and KV-cache overhead. Existing visual token compression methods reduce this cost by pruning low-saliency tokens in shallow layers, but this practice couples temporary computational deactivation with irreversible information deletion. We show that this coupling breaks an intrinsic inference pattern in MLLMs. Through layer-wise attention analysis, we identify \textbf{Visual Backtracking}: visual pivots remain spatially stable, yet in middle layers they actively retrieve context from tokens that appeared unimportant earlier. Thus, early low-saliency tokens can be inactive for computation while still being useful for later reasoning. Motivated by this finding, we propose \textbf{ReTrace}, a backtracking-aware compression framework that introduces a dormant token state. ReTrace caches non-pivot tokens outside expensive layer-wise computation, restores them during the backtracking window, and prunes only after their information has been absorbed. Experiments on Qwen2.5-VL and InternVL3 demonstrate nearly 90% reductions in FLOPs and KV cache, a 2.83$\times$ prefilling speedup, and strong performance retention across complex vision-language benchmarks.

[136] IECLA: Interpretable Emotion Contribution Learning for Affective Image Editing cs.CV | cs.AIPDF

Yijie Guo, Dexiang Hong, Weidong Chen, Xinyan Liu, Zhendong Mao

TL;DR: IECLA先定位哪些图像区域促成或抵触当前情绪，再根据目标情绪驱动局部编辑，以兼顾情绪表达与原图内容结构保持。

Details

Motivation: 情感图像编辑需引发目标情绪同时保留源图内容及结构；复杂场景中仅指定目标情绪无法说明哪些区域应改动、哪些应保持。

Result: IECLA通过反事实干预为对象级情绪贡献构造带符号监督，并训练上下文条件贡献预测器，定位支持或反对源情绪的证据。推理时图像级评估器直接从输入估计源情绪，预测器识别可编辑区域及保留依据，无需反复执行干预。目标感知编辑智能体将证据和期望情绪转成可由现有编辑器执行的局部、内容感知操作。实验显示，IECLA在域内和跨数据集测试中更忠实地预测反事实贡献；外部VLM和独立人工参考支持其局部证据质量。跨六种编辑骨干，方法均强化目标情绪表达并保持图像内容和结构；用户研究也显示目标表达和自然度感知提高。

Insight: 把情绪贡献归因与目标编辑分开，可让编辑决策具有对象级依据，并明确内容保持区域。摘要未给出具体指标和人工研究规模，编辑效果仍依赖贡献预测及下游编辑器质量。

Abstract: Affective image editing (AIE) aims to evoke a target emotion while preserving the content and structure of a source image. Existing methods improve emotional expression or visual fidelity, but the target emotion alone does not reveal which source regions should be changed and which should remain intact, especially in complex scenes. We introduce Interpretable Emotion Contribution Learning for Affective Image Editing (IECLA), a framework that separates source-emotion attribution from target-aware editing. IECLA constructs signed object-level supervision through counterfactual interventions and trains a context-conditioned contribution predictor to localize evidence that supports or opposes the source emotion. At inference, an image-level evaluator estimates the source emotion directly from the input, while the predictor identifies editable and preservation evidence without repeated interventions. A target-aware editing agent then translates this evidence and the desired emotion into localized, content-aware operations for existing editors. Extensive evaluations show that IECLA predicts counterfactual contributions more faithfully on both in-domain and cross-dataset tests, while external VLM and independent human references support the quality of its localized evidence. Across six editing backbones, IECLA consistently strengthens target-emotion expression while preserving image content and structure, and user studies further show improved perceived target expression and naturalness.

[137] MoCKEFA: Model-Conditioned Keyframe Extraction and Fair Attribution for Video Model Explanation cs.CVPDF

Lahiru Wijayasingha

TL;DR: MoCKEFA先按模型预测选择少量关键帧，再用Shapley值在关键帧间公平分摊视频预测重要性。

Details

Motivation: 少数关键帧可能足以保留模型对视频的判断，但传统逐帧归因容易把重要性扩散到大量近重复帧，难解释模型真正依据的时刻。

Result: MoCKEFA以模型对视频的预测为条件提取关键帧，作者称这是该类条件归因方法的首次实现；随后用博弈论公平的Shapley值将视频对模型判断的贡献分配到关键帧。定性与定量实验显示，关键帧归因优于传统帧归因。研究还初步探索通过相邻关键帧差异突出视频中的重要变化。

Insight: 按具体模型预测筛选可保留其决策所需的紧凑视频证据，再分配帧级贡献。摘要没有给出模型、数据集和具体指标，关键帧选择的稳定性及归因计算成本仍不清楚。

Abstract: We show that a video can be represented with just a few keyframes without sacrificing a given model’s belief about the video. Our method is the first of its kind that extracts these keyframes conditioned on the model’s prediction, which is massively helpful for building explainable video models. We then fairly distribute the importance of the video to the model among the keyframes with game theoretically fair Shapley values. Through qualitative and quantitative experiments, we show that keyframe attribution is superior to the traditional frame attribution. Further, we conduct preliminary exploration on highlighting the important changes among the frames by looking at the differences between adjacent keyframes.

[138] Sufficient Frame Subsets for Video Understanding cs.CVPDF

Lahiru Wijayasingha

TL;DR: 该方法提取能单独保留视频模型预测的充分帧子集，用于解释模型在何时获得关键证据。

Details

Motivation: 理解视频模型何时观察到关键内容有助于解释其决策，但现有归因容易把证据分散到许多近重复帧。

Result: 充分帧子集定义为一组紧凑视频帧，在遮蔽其他所有帧后，每个子集仍能独立保持模型预测。作者在UCF101与Something-Something v2上，使用CNN和Transformer架构评估算法，发现能找到较紧凑且有效的解释帧集。依赖更多种充分帧子集的模型往往整体表现更好，作者认为这可能反映鲁棒性，但将其表述为与既有研究一致的观察。提取的子集还用于分析模型失败、查看模型强调的视频内容及识别未来训练改进机会。

Insight: 充分子集关注“哪些帧已足以维持模型输出”，不同于分摊每帧贡献的传统归因。摘要未给出紧凑度、稳定性或准确率数字；子集数量与模型整体性能的关系是观察关联，不能直接视为因果。

Abstract: Understanding when during a video a model makes its critical observations is essential for interpreting video-understanding systems. However, existing attribution methods often spread evidence diffusely across near-duplicate frames. We introduce Sufficient Frame Subsets: compact sets of frames in which all other frames are masked, yet each subset independently preserves the model’s prediction. We evaluate our method on two complementary benchmarks, UCF101 and Something-Something v2, using both CNN- and transformer-based architectures. Our results show that the proposed algorithm identifies highly compact and effective frame subsets for explaining model decisions. We also find that models capable of relying on a greater number of sufficient frame subsets tend to perform better, possibly because they are more robust, consistent with prior research. Finally, the extracted frame subsets help us analyze model failure cases, reveal which aspects of a video the model has learned to emphasize, and identify opportunities for improving future training procedures.

[139] Think Visually Across Space and Time: Grounded Reasoning for Video Segmentation cs.CVPDF

Junwei Wen, Deshui Miao, Guangming Lu, Ming-Hsuan Yang, Xin Li

TL;DR: TCoVP将视频推理分割中的文本推理转为按时间组织的实体引用和空间证据链，帮助模型确定目标、事件时刻及其像素区域。

Details

Motivation: 推理视频分割要求模型根据隐含的自然语言指代识别目标，并预测整段视频的像素掩码轨迹。纯文本思维链虽能解释目标选择，却不显式连接跨帧视觉引用和证据，产生语言推断与视频证据之间的“视频引用差距”。

Result: 作者提出时序视频原语链TCoVP，每个原语将实体引用与时间支持及带时间戳的空间证据关联起来，并在语言推理中交错组织这些信息，以判断哪个实体相关、事件何时成立及证据出现位置。目标原语再转成边界框提示和时间支持供分割模块预测掩码。作者构建TCoVP-SFT-15k和TCoVP-RL-13k，通过监督微调和链对齐强化学习训练该表示。在五个视频推理分割基准和三个模型骨干上的实验显示方法表现强。

Insight: 把实体、时间和空间证据显式绑定到推理步骤，可缩小语言指代与视频掩码定位之间的间隙。摘要未给出五个基准的具体分数、基线列表及链式证据质量的单独指标。

Abstract: Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer. Text-only Chain-of-Thought (CoT) can explain a target choice without making the underlying visual references explicit, leaving a \emph{video reference gap} between linguistic inference and the evidence distributed across frames. We introduce \textbf{Temporal Chain of Video Primitives (TCoVP)}, a spatiotemporally grounded form of Video CoT that organizes query-relevant observations into a chronological reasoning chain. Each primitive associates an entity reference with temporal support and timestamped spatial evidence, and the chain interleaves these references with language to guide reasoning about which entity is relevant, when the queried event holds, and where its evidence appears. The selected target primitives then provide box prompts and temporal support for video mask prediction. We construct TCoVP-SFT-15k and TCoVP-RL-13k to learn this representation through supervised fine-tuning and chain-aligned reinforcement learning. Extensive experiments across five VRS benchmarks and three model backbones demonstrate the strong performance of TCoVP, highlighting the effectiveness of organizing spatial and temporal evidence within the reasoning process. All models and code will be publicly released. Project page: \url{https://victor-wjw.github.io/TCoVP/}

[140] Reflection-Guided Dynamics Rectification for Text-to-Image Generation cs.CV | cs.AIPDF

Shuo Xu, Hanrui Ying, Debin Liu, Jianjun Li, Zhiyuan Ma

TL;DR: RGDR利用已生成图像的反思反馈，在采样轨迹级别调整后续文生图速度预测，提升固定额外生成预算下的最终检查清单得分。

Details

Motivation: 测试时优化需要在有限额外计算预算内细粒度修正已有图像生成结果。既有方法可能发现残余错误，却未必能确保生成过程按反馈处理偏差；问题可能来自初始噪声、文本提示或后续轨迹执行。

Result: 反思引导动力学修正RGDR将完成图像的反思转化为采样过程中的纠正干预，可根据当前采样状态影响每一步，并在候选结果与参考之间进行判别。在额外图像生成预算匹配时，相较以参考轨迹起步，RGDR的最终检查清单得分增益分别为TIR提示改写的1.51倍、Dynamic CFG的1.81倍。

Insight: 把反思用于修正生成动力学而非仅改写初始提示，可能更有效处理轨迹中途出现的偏差。摘要未说明检查清单、生成器、样本量和绝对得分，提升倍率限于报告预算及基线。

Abstract: Fine-grained rectification of existing generations under a limited additional budget remains a key problem in test-time optimization (TTO) for image generation. As for prior approaches, they could identify residual errors, but could not directly ensure that the model acts on the feedback during reasoning. Such errors stem from multiple sources, such as the initial noise or text prompt, and ultimately induce an execution gap of failing to handle potential deviations along the trajectory generating. However, classic models generally cannot remedy this gap. To narrow it, we propose Reflection-Guided Dynamics Rectification (RGDR), which turns reflections on completed outputs into corrective interventions on the process of sampling dynamics at the trajectory level. Specifically, RGDR could impact every sampling step by adapting to instantaneous intervention. Moreover, RGDR generates flexible candidates while adjudicating between them and the reference. As experiments show, rectification from RGDR starts from the existing base trajectory and reflection, and RGDR is substantially more effective than former trajectory exploration. Under matched budgets for additional image generation, RGDR’s gain in the cross-model average final checklist score over the reference is $1.51\times$ that of TIR prompt refinement and $1.81\times$ that of Dynamic CFG.

[141] Image Dehazing as In-Context Tabular Inference cs.CVPDF

Le-Anh Tran

TL;DR: DehazeTab将单图像去雾表述为基于图像块表示的上下文表格回归，用预训练表格基础模型估计大气透射率而无需任务微调。

Details

Motivation: 传统去雾依赖手工物理先验或专用深度模型，通常需要精心设计架构或大量训练。表格基础模型可以从上下文样本直接推断连续关系，可能提供另一种无需针对任务训练模型的方案。

Result: DehazeTab从雾图提取重叠图像块，将其转换成表格样本，特征可用去雾手工描述子、预训练视觉嵌入或二者混合；对应透射率图块汇总为标量回归目标。预训练表格基础模型用带标签参考图像块进行上下文回归，预测未见图像的透射率，不需微调。随后聚合预测构成透射率图，经引导滤波细化，并代入大气散射模型恢复无雾图像。基准实验显示去雾质量有竞争力；物理启发的手工特征可超过预训练视觉嵌入，混合表示提供互补收益。研究还分析特征、图块选择和参考上下文配置的影响。

Insight: 将低层视觉任务转为结构化上下文回归，可复用预训练表格模型而不进行图像任务微调。摘要未给出具体基准分数、模型与每图运行成本，手工特征超过视觉嵌入的观察限于本任务设置。

Abstract: Image dehazing is traditionally addressed through handcrafted physical priors or task-specific deep learning models, often requiring carefully designed architectures or extensive model training. Recent advances in tabular foundation models (TFMs), however, offer an alternative learning paradigm in which continuous relationships can be inferred directly from contextual examples without task-specific model training. Motivated by this capability, this paper investigates whether single-image dehazing can be effectively reformulated as an in-context tabular regression problem and introduces DehazeTab, a framework that estimates atmospheric transmission from structured image-patch representations. Specifically, overlapping patches extracted from hazy images are transformed into tabular samples using handcrafted haze-related descriptors, pretrained visual embeddings, or their hybrid combinations, while corresponding transmission patches are summarized as scalar regression targets. A pretrained TFM then performs in-context regression using labeled reference patches to predict transmission values for unseen images without any further fine-tuning. These predictions are spatially aggregated into a transmission map, refined through guided filtering, and incorporated into the atmospheric scattering model to recover haze-free images. Experiments on benchmark datasets demonstrate competitive restoration performance and reveal that physically motivated handcrafted features can outperform pretrained visual embeddings, while hybrid representations offer complementary benefits. Further analyses examine the effects of feature representations, patch selection, and contextual reference configurations on restoration quality. Beyond image dehazing, the findings highlight the potential of TFMs as an alternative paradigm for addressing low-level vision tasks through structured and in-context inference.

[142] PLANE: Plug-and-Play Fine-Grained Visual Experts for Hallucination Mitigation in Large Vision-Language Models cs.CV | cs.AI | cs.CLPDF

Jianghao Yin, Lang Yu, Qianxi Yu, Qin Chen, Jie Zhou

TL;DR: PLANE为CLIP视觉编码器加入对象、属性和关系三类细粒度LoRA专家，并通过指令感知门控融合视觉特征以减少VLM幻觉。

Details

Motivation: 视觉语言模型可能虚构图像中不存在的对象或错误描述属性与空间关系。CLIP式编码器常聚焦主要物体而忽略细节，被认为是产生幻觉的原因之一。

Result: PLANE包含在CLIP上运行的多方面图像编码LoRA专家及指令感知门控网络。作者构建细粒度图文配对数据，按对象、属性和对象关系描述图像，并分别训练三个LoRA专家。另人工标注GIFT幻觉评测基准，设计为不依赖GPT-4的评估流程。流行幻觉基准和GIFT实验显示，该方法可在不同类型和规模模型中降低幻觉，同时只引入很少推理延迟；相较先进基线也取得更好表现。

Insight: 将视觉细节拆成对象、属性和关系专家并按指令融合，有望弥补单一CLIP表征对细节关注不足。摘要未提供具体幻觉率、基准规模和不同模型分项结果，低延迟没有量化。

Abstract: Large Vision-Language Models (LVLMs) commonly suffer from hallucinations, which appear as fabricating non-existent objects or outputting incorrect attributes and spatial relationships in images. Previous works empirically indicate that CLIP-style vision encoders tend to focus on primary objects while ignoring image details, which is one of the causes of hallucinations. To address this issue, we propose plug-and-play fine-grained visual experts (PLANE) comprising two components: LoRA visual experts operating on CLIP for multi-aspect image encoding, and an instruction-aware gate network for adaptive image feature fusion. During training, we construct a fine-grained image-caption pairs dataset, where images are described from three aspects: objects, attributes, and relations between objects. Image-caption pairs from different aspects are used to train the three fine-grained LoRA experts respectively. Aiming at comprehensively evaluate our method’s capabilities, we manually annotate a fine-grained hallucination evaluation benchmark for LVLMs which is GPT4-independent during evaluation phase (GIFT). We conduct experiments both on popular hallucination evaluation benchmarks and GIFT. The results show that our method effectively reduces hallucinations across models of different types and sizes while introducing minimal inference latency. Compared to advanced baselines, our model also demonstrates superior performance.

Yulin Dai, Yuqian Fu, Tianwen Qian, Zhen Xing, Bojia Zi

TL;DR: Map2Vid将具身导航视频生成定义为鸟瞰图到视频合成，用BEV地图表达人类易理解的场景布局和预期路线。

Details

Motivation: 视频生成模型画质进步明显，但现有控制方式不适合具身导航：文字提示空间歧义较大，低层相机轨迹又不直观且难由用户编辑。

Result: Map2Vid提出从BEV地图生成导航视频的任务和评测基准，地图显式表示全局场景布局与预定路径，使控制更易由人指定。配套数据集覆盖61个室内场景中的10万多条导航序列，每条包括第一人称视频、BEV地图、相机轨迹和语言指令。统一框架融合多种控制信号，并采用课程训练逐渐减弱非地图输入，使推理时既支持完整条件，也可仅用地图输入。多个基础生成模型上的实验显示，使用BEV地图进行可控导航视频生成是可行方向。

Insight: 地图控制将全局路径意图转成空间结构化条件，可能比文字或低层相机参数更容易使用。摘要未报告基准成绩、数据划分和仅地图与完整条件之间的性能差异。

Abstract: Controllable video generation has recently achieved impressive visual quality, yet existing control paradigms remain poorly suited to embodied navigation. Text prompts are spatially ambiguous, while camera-trajectory control relies on low-level motion specifications that are unintuitive for human users and difficult to manipulate. In this work, we introduce Map2Vid, a novel task that reformulates embodied navigation video generation as a BEV-map-to-video synthesis problem, together with a large-scale benchmark for systematic evaluation. By using BEV maps, which provide a visual representation of global scene layout and intended navigation routes, Map2Vid enables human-friendly navigation video generation. To support this task, we construct a dedicated dataset with over 100K navigation sequences across 61 indoor scenes, each paired with egocentric videos, BEV maps, camera trajectories and language instructions. We further develop a unified framework that integrates multiple control signals, together with a curriculum training strategy that progressively weakens non-map inputs, enabling flexible inference under both full conditioning and map-only input. Extensive experiments across multiple foundation generation models demonstrate that controllable navigation video generation via BEV map is a viable and promising direction.

[144] GeoV2V: Geometry-Grounded Video Diffusion Model for Driving Scene Generation cs.CVPDF

Yuchen Xi, Yuhan Guo, Chenwei Hou, Zixu Liu, Jin Fang

TL;DR: GeoV2V将LiDAR和深度几何先验结合整段视频条件化扩散，用于生成大幅轨迹变化下时序连贯且几何一致的驾驶视频。

Details

Motivation: 自动驾驶仿真需要在轨迹变化（如变道）下合成新视角视频。重建方法在大视角变化时视觉质量下降，而先重建再恢复虽然单帧逼真，却常出现几何和时间不一致。

Result: GeoV2V是几何约束视频到视频扩散框架，在Wan 2.1骨干上结合LiDAR与深度几何先验以及完整视频条件。为解决真实数据缺少同步多轨迹监督的问题，作者构建Para4D合成数据集，在动态场景中捕捉同步多车道驾驶视频，明确监督变道生成并为几何一致视角合成提供训练先验。在Waymo、nuScenes和Para4D上的实验显示，方法在视觉质量、几何一致性及复杂光照下的大轨迹偏移鲁棒性方面达到先进表现。

Insight: 几何先验与同步多车道训练样本针对新视角驾驶视频的空间及时间失真。摘要未给出定量分数、各数据集细节和真实闭环驾驶测试，先进表现限于列出的生成评测。

Abstract: Novel view synthesis under trajectory changes is essential for autonomous driving simulation, yet existing methods struggle to generate consistent videos when extrapolating to unseen viewpoints such as lane shifts. Reconstruction-based approaches maintain geometric consistency but degrade in visual quality under large viewpoint changes, while reconstruct-then-restore methods often produce visually plausible frames yet suffer from geometric and temporal inconsistencies. We present GeoV2V, a geometry-grounded video-to-video diffusion framework that synthesizes temporally coherent driving videos along novel trajectories by integrating LiDAR and depth-derived geometric priors with full-video conditioning in the Wan 2.1 backbone. A key challenge in learning cross-lane transformations is the lack of synchronized multi-trajectory supervision in real-world datasets. To address this, we also introduce Para4D, a synthetic dataset capturing synchronized multi-lane driving videos within dynamic scenes, which provides explicit supervision for lane-shift generation and serves as a strong training prior for geometry-consistent view synthesis. Experiments on Waymo, nuScenes, and Para4D demonstrate state-of-the-art performance in visual quality, geometric consistency, and robustness to complex lighting dynamics under large trajectory shifts. Project page: https://xiyuche.github.io/GeoV2V/

[145] Visual Scaffolding: Toward Native Multimodal Reasoning cs.CVPDF

Run Shao, Zhaoyang Zhang, Ziyu Li, Linrui Xu, Ji Qi

TL;DR: Visual Scaffolding允许多模态模型在推理中主动检索原生视觉特征，将其作为可复用中间状态插入思维链，而不必把所有视觉差异转成文字。

Details

Motivation: 链式思维通过显式中间结果支持后续推理，但许多视觉区别难以用简短语言准确表达，现有多模态模型缺少可与语言并列复用的视觉推理支架。

Result: 作者将原生多模态推理定义为视觉表示与语言共同承载可复用中间状态。模型可在任一推理步骤触发检索动作，轻量检索器选择原生视觉特征并直接纳入不断演进的思维链，供后续推理使用。NMR-Bench包含Burden Counting和Path Tracing两个受控视觉任务。在Qwen3-VL-4B-Thinking上增加8,800万参数（约2%）后，两任务准确率分别为66.43%和72.50%，超过被测GPT-5.5及Gemini 3.1 Pro基线。干预实验改变所选视觉内容会影响准确率，并能在固定文本条件下改变后续视觉选择。

Insight: 视觉特征可像语言中间步骤一样成为推理材料，并可支持后续重用；固定文本仍受改变视觉选择影响，说明视觉状态并非仅作一次性输入。摘要未给出基线具体数值、检索延迟和更多任务覆盖。

Abstract: The success of chain-of-thought (CoT) in complex reasoning can be attributed to a textual scaffolding effect: explicit intermediate results can be examined, reused, and composed to support subsequent reasoning, much as scaffolding supports a building under construction. However, mainstream multimodal LLMs lack an equivalent scaffold for visual distinctions that are difficult to express in concise verbal descriptions. We therefore ask: Can native visual representations become first-class citizens alongside language, providing a visual scaffold for reasoning? We define native multimodal reasoning as reasoning in which native visual representations and language jointly carry reusable intermediate states. As a concrete implementation, the visual scaffolding mechanism we proposed lets the MLLM autonomously initiate a retrieval action at any reasoning step. A lightweight retriever selects native visual features, which are then incorporated directly into the evolving CoT and remain available to subsequent steps. To evaluate this mechanism under controlled visual demands, we introduce NMR-Bench, comprising Burden Counting and Path Tracing. With 88M added parameters (approximately 2%) on Qwen3-VL-4B-Thinking, our model achieves 66.43% and 72.50% accuracy, respectively, surpassing the evaluated GPT-5.5 and Gemini 3.1 Pro baselines. Controlled interventions show that changing selected visual content affects task accuracy and can redirect subsequent visual selection even when intervening text is held fixed. These findings show that native visual states can function as intermediate reasoning material and point toward reusable scaffolds for additional modalities.

[146] Persistent Representational Misalignment Induces Vision-Language Model Failures cs.CV | cs.LGPDF

Yonatan Gideoni, Yoav Gelberg, Tim G. J. Rudner, Yarin Gal

TL;DR: 研究发现，视觉语言模型失败常与视觉和语言表征难以对齐有关；这种数据导致的线性不可对齐不会仅靠扩大模型规模或提高对齐复杂度消失，并与多种泛化失败相关。

Details

Motivation: 一些视觉语言模型会在对人类简单的任务上出错。论文据此检验一种解释：模型是否难以在视觉与语言表征之间转换，以及这种错位是否会如“柏拉图表征假说”的强版本所预期那样随规模增长而消失。

Result: 作者先发现现有视觉表征与语言表征无法线性对齐，并将这种错位归因于数据本身，因此预期它不会因模型规模增大或对齐方法更复杂而消失。进一步实验显示，错位会导致可预测的泛化失败，并与已有失败模式相关；表征对齐程度更高的视觉语言骨干模型相应地出现较少失败。

Insight: 论文反对“足够大的模型自然会形成统一表征”这一强假设，提示视觉语言系统的部分错误来自模态表征之间的结构性差异。提升对齐程度可能减少失败，但摘要没有给出各任务、模型规模或具体错误类型的定量细节。

Abstract: Vision-language models (VLMs) fail at some tasks that are simple for humans—but why? We hypothesize that many such failures stem from VLMs’ difficulties translating between vision and language representations, a misalignment the Platonic Representation Hypothesis predicts should disappear with scale. This paper is composed of two parts: in the first, we show that existing vision and language representations are linearly unalignable. This misalignment is an inherent property of the data and hence expected to persist regardless of either scale or the alignment’s complexity. In the second, we demonstrate how this misalignment both leads to a predictably induced generalization failure and correlates with existing failure modes, where more representationally aligned VLM backbones accordingly exhibit fewer failures. Overall, these results imply representations will not become aligned simply given sufficient scale, providing evidence against a strong version of the Platonic Representation Hypothesis.

Liudi Yang, Ziyuan Liu

TL;DR: GeoART 是一种无需训练的测试时自我精炼方法：它在视频扩散去噪过程中生成并筛选候选状态，以几何先验增强相机控制，在不改动预训练模型的情况下改善相机可控性、多视角一致性和视觉质量。

Details

Motivation: 相机条件视频扩散模型对相机指令的控制常不准确，生成视频也可能几何不一致，因为去噪过程中相机约束和估计几何施加得较弱。已有修复往往需要高成本的几何感知微调或额外三维监督。

Result: GeoART 在选定去噪步骤执行 Predict-and-Perturb 循环，生成多个候选并用几何先验评估，而不直接修改去噪轨迹。置信度门控会抑制不可靠的几何证据；经排序校准的概率选择综合分数大小和相对排名来决定下一精炼状态。实验显示，相比已有采样和自我精炼方法，它在相机可控性、多视角一致性和视觉质量上均有稳定提升，且无需训练或修改预训练视频扩散模型。

Insight: 测试时利用几何反馈可以在不重新训练模型的前提下改善相机条件生成；门控和排序校准用于降低不可靠几何评分造成的误选风险。摘要没有报告提升幅度、适用模型范围或计算开销，因此这些方面仍无法据此判断。

Abstract: Camera-conditioned video diffusion models often suffer from inaccurate camera control and geometric inconsisten- cies because camera constraints and estimated geometry are only weakly enforced during denoising. Existing reme- dies often require costly geometry-aware fine-tuning or ad- ditional 3D supervision. We introduce GeoART, a training- free self-refinement framework that improves camera con- trollability through geometry-aware test-time scaling. At selected denoising steps, GeoART generates multiple can- didates within the Predict-and-Perturb cycle and eval- uates them using a geometry prior, without directly modify- ing the denoising trajectory. A confidence gate suppresses unreliable geometric evidence, while rank-calibrated prob- abilistic selection combines score magnitude and relative rank to choose the next refinement state. Experiments demonstrate consistent improvements in camera controlla- bility, multi-view consistency, and visual quality over exist- ing sampling and self-refinement methods, without training or modifying the pretrained video diffusion model

[148] Holo4D: Holistic 4D Reconstruction as Geometric Control for Video Diffusion cs.CVPDF

Yushi Lan, Zeren Jiang, Koichi Namekata, Xingang Pan, Chuanxia Zheng

TL;DR: Holo4D 将前馈式 4D 重建得到的场景轨迹作为整体几何与外观支架，引导预训练视频扩散模型沿新相机轨迹重拍单目视频，同时由生成先验补全新显露区域。

Details

Motivation: 从新相机轨迹合成单目视频，既要保留观测到的场景几何和运动，也要补全视角变化后才可见的内容。前馈 4D 重建器能恢复观测场景结构，视频扩散模型则擅长生成未观测内容，论文希望把二者结合。

Result: Holo4D 将重建出的 3D 轨迹投影到目标视角，并编码几何信息和源视频外观，用作预训练视频扩散模型的条件；源视频和目标相机也共同参与生成。重建观测约束有依据的内容，缺少观测支持的区域交由生成先验补全。实验显示，基于轨迹的条件有助于相机控制和新视角合成；该框架可替换不同 4D 重建器，且无需重新训练视频生成器。

Insight: 轨迹支架把重建与生成分工衔接起来，并允许重建模块迭代升级而不必重训视频生成模型。摘要未提供量化指标、基线提升幅度或失败案例，因而无法判断不同重建器和大幅视角变化下的效果边界。

Abstract: Synthesizing a monocular video from a new camera trajectory requires preserving the observed scene geometry and motion while completing regions that become newly visible. Feed-forward 4D reconstructors recover the structure of the observed scene, whereas video diffusion models provide a generative prior for completing unseen content. We introduce Holo4D, a video reshooting framework that uses reconstructed 3D trajectories as a holistic source-view 4D scaffold. We project these trajectories into the requested views and encode their geometry and source appearance to guide a pretrained video diffusion model. Together with the source video and target cameras, the scaffold grounds generation in reconstructed observations while leaving unsupported regions to the generative prior. Experiments demonstrate the value of trajectory-based conditioning for camera control and novel-view synthesis. The framework accommodates different 4D reconstructors without retraining the video generator, allowing it to benefit from future advances in source reconstruction. More details can be found in our project page.

[149] The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability? cs.CV | cs.LGPDF

Ronen Schaffer, Hila Manor, Tomer Michaeli

TL;DR: 系统实验和理论分析表明，ViT 单个 patch token 的可解释解码并非 Transformer 架构固有属性；不同模型可能呈现语义对齐、反向解释或与图像语义无关的结果，且会随优化设置变化。

Details

Motivation: 视觉 Transformer 分类器中，将单个 patch token 解码为图像已成为常用诊断手段，因为输出往往容易被人理解。论文检验这种 patch 解码可解释性是否能跨架构、规模和训练目标普遍成立，还是只由特定训练偏置带来。

Result: 作者在多种 ViT 架构、规模与训练目标上结合受控实验和真实模型进行研究，发现有些模型的 patch 解码与语义对齐且可解释，有些则产生反向的解释图，另一些与图像语义完全不相关。这些行为并非架构决定，会随优化景观改变。理论上，作者构造了固定的玩具 Transformer 架构，使其能以多种方式实现最优分类规则，而这些实现具有不同的 patch 解码模式。

Insight: 不能仅凭 Transformer 架构就预先假设 token 解码能解释模型决策；可解释性依赖问题设置和训练过程。patch 解码可作为特定模型的诊断信号，但需要实证验证其与语义及决策的关系，不能视为普遍保证。

Abstract: Transformers are widely assumed to exhibit token-decoding explainability. In particular, in vision transformer (ViT) classifiers, decoding individual patch tokens has become a common diagnostic tool due to its tendency to yield human-interpretable outputs. In light of the increasing popularity of this approach, an important question is whether token-decoding explainability is an inherent property of transformers that should be expected to hold across architectures, scales, and training objectives, or merely an artifact of some implicit bias. In this paper we systematically study token-decoding explainability across diverse ViT settings, spanning architectures, scales, and training objectives, using both controlled experiments and real-world models. We find that while some models indeed exhibit aligned and interpretable token decodings, others exhibit either inverted explainability maps or a complete lack of correlation with the image semantics. As we show, these tendencies toward different explainability behaviors are not intrinsic to the architecture and can vary arbitrarily with changes in the optimization landscape. We provide theoretical support for these observations by demonstrating a setting where a single fixed toy transformer architecture can provably implement the optimal classification rule in multiple ways, which differ in their patch-decoding explainability patterns. Our results show that token-decoding explainability is not an inherent property of trained transformers, but rather an outcome contingent on the problem setting and thus should not be a-priori assumed to hold. Code can be found in https://ronensc.github.io/Explainability-Illusion/ .

[150] MotionShift: Environment-Compatible Motion Transfer via Planner-Evaluator Synergistic Adaptation cs.CVPDF

Yanchen Liu, Junyao Gao, Yanan Sun, Wensheng Chen, Guangming Lu

TL;DR: MotionShift 用大语言模型规划运动轨迹、视觉语言模型结合目标场景评估并修正轨迹，以免动作迁移只追求复刻参考视频而违反目标环境的物理兼容性；该训练无关流程再用轨迹条件化 I2V 模型生成视频。

Details

Motivation: 现有动作迁移方法擅长把参考视频中的运动精确复制到目标场景，却常忽略动作是否符合目标环境和物理规律。论文希望在保持高保真运动迁移的同时，让动作适应多样且复杂的目标场景。

Result: MotionShift 采用规划器—评估器迭代适配：LLM 利用语义推理生成运动轨迹，VLM 凭视觉感知和场景理解检查并修正轨迹。适配后的轨迹作为图像到视频模型的生成条件，输出目标场景视频。实验显示，该方法可在多种复杂场景中实现高保真且物理兼容的运动迁移，并能接入现有轨迹条件化 I2V 模型。

Insight: 该方法让运动轨迹在渲染前先根据环境语义和视觉证据调整，而不是直接照搬参考动作；规划与评估协作无需训练。摘要没有说明物理兼容性的定量定义、用户或自动评测细节，也没有报告各类场景的提升幅度。

Abstract: Existing motion transfer methods excel at exact, rigid motion replication from reference videos to target scenes, whereas they often fail to account for the compatibility between the transferred motion and the target scene environment in accordance with physical laws. In this work, we propose MotionShift, a novel motion transfer framework that achieves high-fidelity motion transfer while ensuring physical compatibility with the target scene. In particular, we design a planner–evaluator iterative motion adaptation mechanism, in which a Large Language Model (LLM) acts as the planner to generate motion trajectories by leveraging its strong semantic reasoning capabilities, while a Vision-Language Model (VLM) serves as the evaluator to rectify the motion based on its superior visual perception and scene understanding. Through this planner–evaluator synergistic reasoning process, the motion is adapted to the target scene in a training-free manner. The adapted motion is then used as a generative condition for an image-to-video (I2V) model to render videos in the target scene. Extensive experiments demonstrate that MotionShift achieves high-fidelity and physically compatible motion transfer across diverse and complex scenes. Moreover, we show that MotionShift can be readily integrated into existing trajectory-conditioned I2V models, endowing them with realistic motion transfer capabilities.

[151] SENRI: Understanding the Space-Event Nexus in Real-Life Videos cs.CVPDF

Kehan Lan, Kaining Ying, Henghui Ding, Kaipeng Zhang

TL;DR: SENRI 将长时连续第一视角视频中的空间—事件联系作为评测任务，涵盖由事件线索恢复空间信息和由空间区域/轨迹检索事件；其基准显示现有长视频方法仍有明显困难，SENRI-Agent 通过目标引导的局部视觉核查提升既有记忆系统表现。

Details

Motivation: 人的记忆会把事件与发生地点联系起来，长视频助手也需要记住并依据空间线索回忆事件。现有长视频理解评测和记忆方法未充分检验连续真实视频中的这种双向空间—事件关联。

Result: SENRI 定义 Event→Space 与 Space→Event 两个方向的六类任务，并构建 SENRI-Bench：对 10 段连续自录视频（约 25.3 小时）人工标注的 500 道选择题。直接感知、证据检索和持久记忆三类范式在任务上都仍具挑战。论文指出现有记忆方法的两种代表性失效，并提出 SENRI-Agent：用目标引导视觉核查生成局部检查任务，由并行验证器执行，再把报告整合进工作记忆。该代理无需改动原有存储记忆即可提升既有基线。

Insight: 长视频理解除了记住事件内容，还需把空间区域、运动轨迹、事件和时间证据联系起来；局部视觉复核可补足持久记忆的证据缺口。摘要未列出具体准确率和两种失效模式的细节。

Abstract: Human memories are organized in a way that connects events with their corresponding spatial contexts. An AI assistant should also be able to memorize and recall the nexus of events and spaces. Thus we introduce SENRI, a task that studies this space–event connection in long, continuous real-life egocentric video. It contains six task types in two directions: Event$\rightarrow$Space, which recovers spatial knowledge from event cues, and Space$\rightarrow$Event, which retrieves events from regions and trajectories. To evaluate this, we build SENRI-Bench, containing 500 manually annotated multiple-choice questions over 10 uninterrupted self-recorded videos with around 25.3 hours of content. Also, we evaluate three typical long video understanding paradigms: direct perception, evidence seeking, and persistent memory. The task remains challenging across all three paradigms, highlighting substantial room for improvement. Then we show two representative failure modes of existing memory methods and address these evidence gaps with SENRI-Agent, a controller that uses Target-Guided Visual Check (TGVC) to propose local visual check tasks for parallel verifiers and integrate the reports into a working memory to help with reasoning. Built upon existing memory systems, SENRI-Agent improves the performance of existing baselines without modifying their stored memories. We hope SENRI will provide valuable insights into how real-life egocentric understanding should be evaluated, reveal the gaps in existing approaches, and offer a possible direction for improvement.

[152] MPIG: Multi-angle Polarization Image Generation from a Single Polarization Observation cs.CVPDF

Jue Wang, Ruilin Huang, Liquan Shen, Zhen Tian, Xia Hua

TL;DR: MPIG 从单张 0°偏振观测预测缺失的 45°、90°和 135°视图，以跨角特征交互、偏振约束监督和低秩解码兼顾场景结构与角度变化；在新建的真实四角数据集上取得较高重建指标，并保留部分下游任务信息。

Details

Motivation: 完整偏振成像通常需要专用硬件和多个角度测量，增加系统复杂度并限制部署。论文研究能否仅凭一张 0°偏振图像恢复其他三个偏振角度的观测。

Result: MPIG 采用 Swin Transformer，将跨角特征交互与残差解码结合，预测随角度变化的强度，同时保持共享场景结构；Stokes 差分和偏振强度感知监督用于恢复偏振调制、方向和幅值，低秩分解降低解码成本。作者构建包含 5,474 组高分辨率真实四角图像的 HighPol。实验报告 PSNR 29.729 dB、SSIM 0.9505、AoLP MAngE 21.514°，模型含 85.86M 参数、计算量 1.126T FLOPs；玻璃相关任务和语义分割验证显示合成图像保留了任务相关信息。

Insight: 单角观测结合学习到的场景先验可用于补全多角偏振信息，且恢复质量需要同时看图像指标和下游用途。摘要未给出上述指标的对照基线数值，也未说明单角输入在特殊材料或场景下的失效边界。

Abstract: Polarization imaging provides informative cues for material analysis, scene understanding, and visual perception. However, acquiring complete polarization observations typically requires specialized hardware and multi-angle measurements, thereby increasing system complexity and limiting practical deployment. In this work, we address the problem of predicting the missing $45^\circ$, $90^\circ$, and $135^\circ$ polarization views from a single $0^\circ$ observation by leveraging a learned scene prior. To this end, we propose MPIG (Multi-angle Polarization Image Generation), a Swin Transformer-based framework that couples cross-angle feature interaction with residual decoding to predict angle-dependent intensity variations while preserving shared scene structure. Specifically, measured Stokes differences and polarization-strength-aware supervision guide the recovery of polarization modulation, orientation, and magnitude, while low-rank factorization further reduces the decoding cost. We also construct HighPol, a large-scale real-world polarization dataset comprising 5,474 groups of high-resolution four-angle polarization images collected from diverse indoor and outdoor scenes. Experimental results demonstrate improved image and polarization accuracy, achieving 29.729 dB PSNR, 0.9505 SSIM, and $21.514^\circ$ AoLP MAngE while requiring 85.86 M parameters and 1.126 T FLOPs. Furthermore, downstream validation on glass and semantic-segmentation tasks suggests that the generated polarization images can preserve task-relevant information, demonstrating the effectiveness of MPIG and the practical utility of predicted polarization observations for downstream vision applications.

[153] CoRe-WM: Conditioned Response-State World Modeling for Longitudinal Treatment-Response Prediction in Breast DCE-MRI cs.CVPDF

Lin Zhao, Bowen Chen, Qiaohui Gao, Shaowen Wan, Xiang Li

TL;DR: CoRe-WM 将乳腺癌新辅助治疗后的纵向 pCR 预测建模为条件化影像状态转移：在 I-SPY2 各前瞻决策点取得所评方法中最高区分能力，并在 I-SPY1 展示跨队列迁移能力。

Details

Motivation: 新辅助化疗后的病理完全缓解（pCR）是侵袭性乳腺癌的重要预后指标。只用治疗前影像会忽略患者逐步变化的响应；直接以结局监督多访视分类器又会把表示学习与临床终点耦合。

Result: CoRe-WM 学习从已观测影像序列预测下一影像状态的潜在转移模型，并通过加性偏移和逐特征调制纳入患者临床特征与治疗方案。因子化响应状态分支将治疗响应动态与影像外观分开，编码响应主要体现于病灶形态的先验。预训练后冻结全部参数，以线性读出在各临床决策点预测 pCR。五折交叉验证的 I-SPY2 评估中，CoRe-WM 在每个前瞻决策点均取得最高区分能力；在 I-SPY1 的跨队列迁移也较稳健。消融结果指出患者条件化和因子化响应路径是主要贡献来源。

Insight: 先学习患者与治疗条件下的影像状态变化，再用简单线性读出预测结局，可将纵向表示学习与结局拟合分离。摘要未给出具体 AUC 或队列规模，因此不能量化其相对优势及适用人群边界。

Abstract: Pathologic complete response (pCR) after neoadjuvant chemotherapy is a key prognostic indicator in aggressive breast cancer, yet existing prediction approaches either rely on pretreatment imaging alone, missing the patient’s evolving response, or optimize multi-visit classifiers under direct outcome supervision, coupling representation learning with the clinical endpoint. We formulate longitudinal pCR prediction as a world-modeling problem and propose Conditioned Response-state World Modeling (CoRe-WM), which learns a latent state-transition model that predicts the next imaging state from the observed sequence. The transition function is conditioned on patient clinical profile and assigned therapy through additive shift and feature-wise modulation, enabling patient- and treatment-specific dynamics. A Factorized Response State pathway further decouples treatment-response dynamics from imaging appearance by encoding the structural prior that response manifests primarily in lesion morphology. After pretraining, all parameters are frozen and a linear readout predicts pCR at each clinical decision point. We evaluate CoRe-WM on I-SPY2 under five-fold cross-validation and assess cross-cohort transfer on I-SPY1. CoRe-WM achieves the highest discrimination among all evaluated methods at every prospective decision point, and demonstrates robust cross-cohort transfer. Ablation studies identify patient-profile conditioning and the factorized response pathway as the principal contributors. These results establish world-model-based representation learning as a promising framework for outcome prediction in oncology and beyond.

[154] MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes cs.CVPDF

Yifei Wang, Wei Tian, Wenhua Bao, Jia Zhou, Hanshi Wang

TL;DR: MCSplat 通过跨相机外观规范化，避免曝光、ISP 和光照差异被错误吸收到高斯几何中，并用 Fisher 信息估计的置信度调节局部几何自监督；实验显示驾驶场景重建、跨视角几何一致性和跨数据集泛化均改善。

Details

Motivation: 从稀疏多视角图像前馈重建驾驶场景时，跨相机曝光、ISP 和光照变化可能被误学成几何差异；深度不连续处的法线约束也可能提供不可靠局部监督。论文针对这两类误差提出一致性重建方法。

Result: MCSplat 基于深度引导的多相机对应关系建立规范外观空间，以分离几何学习和相机特定外观；多尺度双边网格 Transformer 从规范渲染恢复原相机外观，使光度监督不污染高斯几何。方法还根据光度敏感度估计 Fisher 信息不确定性，并据此自适应加权面向表面的自监督。实验报告重建质量和跨驾驶数据集泛化提升；几何评测显示跨视角深度一致性及局部法线一致性改善，定性结果中几何伪影减少、重叠视角结构更连贯。

Insight: 多视角重建应显式区分外观变化和几何差异，并按局部几何监督的可靠性调整约束强度。摘要没有给出数据集名称、具体指标和数值增益，无法判断在不同传感器或极端光照变化下的表现。

Abstract: Driving-scene reconstruction via feed-forward Gaussian splatting from sparse multi-view images remains challenging due to two factors. First, cross-camera photometric inconsistency, caused by variations in exposure, ISP, and illumination, can be incorrectly absorbed into Gaussian geometry. Second, unreliable local geometric supervision arises from normal-based constraints near depth discontinuities. To address these issues, we present \textbf{MCSplat}, a multi-view consistent feed-forward Gaussian splatting framework for driving scenes. MCSplat constructs a canonical appearance space from depth-guided multi-camera correspondences to decouple geometry learning from camera-specific appearance. A multi-scale Bilateral Grid Transformer restores raw camera appearance from canonical renderings, enabling photometric supervision without contaminating Gaussian geometry. To improve local geometric reliability, we estimate Fisher-information-based uncertainty from photometric sensitivity and use it to adaptively weight surface-oriented self-supervision. Extensive experiments demonstrate improved reconstruction quality and generalization across driving datasets. Geometry-oriented evaluations further reveal enhanced cross-view depth consistency and local surface-normal agreement, while qualitative results show fewer geometric artifacts and more coherent structures across overlapping views.

[155] Local Evidence, Global Failure: Diagnosing Speech-Vision Temporal Composition in Audio-Visual Language Models cs.CV | cs.AI | cs.MMPDF

Jaehong Myung, Seunghun Oh, Unsang Park

TL;DR: TED 将音视频模型的长视频判断拆为基于语音时间戳的局部视觉核验和相对语音时刻的证据聚合；在四个冻结 AV-LLM 的 SVHall 预测任务上，准确率从完整视频提问时的 29.3–33.7% 提升到 79.3–82.2%。

Details

Motivation: 音视频大模型可能在完整视频的语音—视觉判断中失败，即使相关视觉证据在局部片段里可以恢复。论文希望区分“局部证据可见”与“模型能否在整段视频中正确组合证据”，并诊断失败原因。

Result: 免训练诊断框架 Temporal Evidence Decomposition（TED）先用自动语音识别获得转录文本和话语区间，再查询不含音频的局部窗口：时间对齐与预测任务使用转录文本，跨模态时间绑定使用候选答案，最后相对话语时间聚合证据。四个冻结 AV-LLM 在 SVHall 的时间预测准确率由完整视频提问的 29.3–33.7% 提升至 TED 下的 79.3–82.2%。替换查询文本会降低恢复效果；固定可用时间区域后，所观察到的时间锚点仍有帮助。物理隔离和时间捷径控制用于检验其他解释。

Insight: 长视频判错可能并非缺少局部视觉线索，而是没有把叙述内容、话语发生时间和对应视觉证据明确拆开并正确组合。结果针对 SVHall 时间预测任务；摘要未证明对其他音视频任务有同等幅度提升。

Abstract: Audio-visual large language models (AV-LLMs) can fail at full-video speech-vision decisions even when useful visual evidence is locally recoverable. We introduce Temporal Evidence Decomposition (TED), a training-free diagnostic framework that separates local visual verification from speech-relative aggregation. Automatic speech recognition supplies a transcript and utterance interval. TED queries audio-free windows with the transcript for Temporal Alignment and Forecasting, or answer candidates for Cross-Modal Temporal Binding, then aggregates evidence relative to that interval. Across four frozen AV-LLMs on SVHalluc, Temporal Forecasting accuracy rises from 29.3-33.7% under full-video questioning to 79.3-82.2% with TED. Query replacement reduces recovery; the observed anchor remains beneficial when temporal-region availability is held fixed. Physical-isolation and timing-shortcut controls test alternative explanations. The findings distinguish locally recoverable evidence from successful full-video decisions, motivating explicit separation of narration content, utterance timing, and time-indexed visual evidence.

[156] Painting Change Map as guidance for generating painting processes cs.CVPDF

Markus Pobitzer, Chang Liu, Niccolò Biondi, Nicu Sebe

TL;DR: Painting Change Map 用千步连续尺度上的时空变化图同时表达绘画过程“何时、何处”发生变化，为绘画生成和反向擦除提供局部笔触级控制；基准实验优于现有控制基线，在 oracle 条件下超过全部基线。

Details

Motivation: 绘画生成视频已从静态图像扩展到模拟绘制或擦除过程，但现有控制要么依赖刚性的边缘、深度等静态几何条件，要么只用全局进度标量，无法指定局部笔触变化及发生时机，限制了用户对过程的控制。

Result: 论文提出 Painting Change Map，在贯穿 1,000 个时间步的连续过程上追踪归一化结构变化，并把局部视觉差异编码为 RGB 表示，用于前向绘制和反向擦除的细粒度引导。作者报告其在广泛基准比较中超过现有控制基线，并达到与当前先进绘画方法相当或更优的结果；在 oracle 条件下显著超过所有基线。

Insight: 要控制生成过程，需要同时描述动作发生的位置和时间；全局进度与静态边缘等单一条件难以提供笔触级的时空约束。摘要未说明 oracle 条件的可用信息、各指标数值和推理成本，解读最优结果时需注意该条件限制。

Abstract: Digital art generation has undergone a massive paradigm shift, moving from static image synthesis to dynamic video generation that can mimic the human painting process. While existing models can synthesize visually stunning time-lapses of paintings being created or “unpainted,” they largely treat the creative process as a black box. Current conditioning mechanisms are either too rigid, relying on static geometric cues like Canny edges or depth maps, or too holistic, utilizing simple scalar progress variables that offer no localized control. Consequently, users lack the agency to steer how a painting evolves over time. In this work, we argue that true control over the video generation of artwork requires a joint understanding of two fundamental dimensions: where an artist acts, and when those actions take place. To achieve this, we introduce the Painting Change Map, a novel spatio-temporal control mechanism that explicitly tracks normalized structural changes across a 1,000-timestep continuum, a resolution sufficient to capture fine-grained stroke-level changes while remaining computationally tractable. By mapping localized visual differences into an encoded RGB space, our representation enables fine-grained guidance for both the forward painting and reverse unpainting processes. Extensive benchmarking demonstrates the efficacy of our approach. We outperform existing control baselines and achieve competitive or superior results compared to state-of-the-art painting methods. Furthermore, under oracle conditions, our method surpasses all baselines by a significant margin. Code is available at https://github.com/Markus-Pobitzer/painting-change-map.

[157] A 3D-Driven Framework for Moving-Camera Perspective-to-360° Video Generation cs.CVPDF

Michael Wakeham, Mrinmoy Kumar Das, Wenqi Wang, Ayush Tewari, Boqing Gong

TL;DR: 该 3D 驱动框架将运动相机的透视视频转换为 360°视频：显式重建观测内容，用预训练全景图像模型补全未观测区域，再将生成内容锚定到共享三维场景；在 Argus 静态场景子集上无需额外全景视频训练即取得先进结果。

Details

Motivation: 从运动相机透视视频生成 360°视频，既要保留输入随时间揭示的观测，也要补全画面外区域，并保证相机变化时场景一致。把所有需求交给单一条件视频模型通常需要联合全景视频数据训练。

Result: 作者采用模块化三维流程，将观测内容显式重建为 3D，并从选定的全景渲染结果出发，用预训练全景图像模型生成未观测内容；随后把生成内容提升并固定到共享 3D 场景表示中，保证不同视角下的空间锚定。无需额外全景视频训练，该方法在 Argus 评测集静态场景子集的输入视角保真度、感知质量和几何一致性上达到当时最先进水平。

Insight: 将观测重建、未知区域补全和跨视角一致性分配给专门组件，可能是直接视频生成之外的可行路线。当前摘要只报告静态场景子集，未给出具体数值，也不能据此推断动态场景表现。

Abstract: Turning a moving-camera perspective video into a 360° video requires more than just generating content outside the input frame: observations revealed over time must be preserved, unseen regions completed, and the resulting scene kept consistent as the camera viewpoint changes. Existing methods assign these different requirements to a single conditional video model, requiring them to be jointly learned from panoramic video data. We instead present a modular, 3D-driven framework that distributes these three responsibilities across specialized components, all connected through a shared scene representation. Input observations are reconstructed explicitly in 3D, while a pretrained panoramic image model completes unobserved content from a selected panoramic render. This generated content is lifted into the shared 3D scene, where it remains spatially anchored across camera viewpoints. Without requiring any additional panoramic-video training, our framework achieves state-of-the-art input-view fidelity, perceptual quality, and geometric consistency on a static-scene subset of the Argus evaluation set. These results support our decomposed, 3D-driven formulation as a promising alternative to direct video generation in this setting.

[158] MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion cs.CVPDF

Houyuan Chen, Fanzhou Wang, Xiangyu Fan, Zesong Qiu, Yukun Wei

TL;DR: MT Forcing 通过单轮分布匹配蒸馏和教师强制式历史转移监督，缓解自回归扩散统一多模态模型在多轮图像编辑中的画质退化、细节伪影和跨轮内容漂移。

Details

Motivation: 多轮图像编辑允许用户按连续指令交互修改图像，但现有方法往往几轮后就出现伪影累积、细节变差，以及未要求修改的区域或先前编辑发生意外变化。论文关注如何同时保持每轮视觉质量和跨轮内容一致性。

Result: MT Forcing 是面向自回归扩散统一多模态模型的训练范式。单轮分布匹配蒸馏（Single-Turn DMD）让学生模型的自回放由教师对源图像一次性施加累计编辑来监督，以提供质量指导并避免递归错误累积。Teacher-Forcing Guidance 在预热期间和蒸馏中作为正则项，沿参考轨迹监督依赖历史的状态转移，补充逐轮分布匹配未能保证的跨轮一致性。实验显示，该方法显著增强多轮编辑能力，使视觉质量更稳定、跨轮一致性更强。

Insight: 多轮编辑既要匹配各步输出质量，也要显式监督“给定编辑历史后的状态变化”，仅逐轮优化图像分布不足以防止漂移。摘要未提供具体基准、提升幅度或训练成本。

Abstract: Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users. However, existing methods often exhibit noticeable drift after only a few editing turns, manifesting in two aspects: 1) unstable visual quality, with accumulating artifacts and degraded details; and 2) inconsistent image content, with unintended changes to preserved regions and prior edits. To address these challenges, we propose MT Forcing, a novel training paradigm for multi-turn image editing with unified multimodal models (UMMs) based on autoregressive diffusion. Specifically, we propose Single-Turn Distribution Matching Distillation (Single-Turn DMD), which supervises student self-rollouts with a teacher that applies cumulative edits to the source image in a single turn, providing visual-quality guidance without recursive error accumulation. Additionally, we introduce Teacher-Forcing Guidance, which supervises history-dependent transitions on reference trajectories during warmup and as a regularizer during distillation, encouraging cross-turn content consistency beyond what per-turn distribution matching alone can ensure. Extensive experiments demonstrate that our approach significantly improves the multi-turn image editing capabilities of UMMs, yielding more stable visual quality and stronger cross-turn consistency.

[159] UniState: Unifying Memory and Intervention for Controllable World Simulation cs.CVPDF

Jaeseok Jeong, Kyungmook Choi, Sohyun Chung, Youngsik Yun, Youngjung Uh

TL;DR: UniState 用统一世界状态表示空间记忆与用户干预，让用户分别指定物体运动和观察相机轨迹，并训练视频模型遵循这两类控制、同时生成合理的额外运动；实验显示联合物体和相机控制更准确，场景变化还能在后续 rollout 中保持。

Details

Motivation: 世界模型提供了可探索环境，但在场景中进行干预后，还要从用户自由选择的视角观察结果仍很困难。控制描述也容易混淆物体真正移动与仅仅改变相机视点。

Result: UniState 将空间记忆和用户干预统一到世界状态中：用户可直接在空间记忆里指定物体运动，同时独立选择观察相机轨迹，从而明确区分场景变化和视角变化。干预无需逐项规定所有后续运动，视频模型仍可在受控物体中合成附加动作及其与周围环境的合理交互。训练和采样策略用于同时遵循物体运动和相机轨迹并保留生成响应的能力。实验显示，相较已有方法，联合物体与相机控制更准确；定性结果还显示干预造成的场景改变会持续到后续 rollout。

Insight: 将干预目标、场景记忆和观察视角纳入同一状态表示，可减少控制语义歧义，同时允许模型补全用户未明确规定的后果。摘要没有给出量化结果或跨 rollout 持久性的评测细节。

Abstract: World models offer increasingly rich environments for exploration, yet intervening in a world while observing the consequences from freely chosen viewpoints remains challenging. We tackle these challenges by introducing UniState, a framework for controllable world simulation built on a unified world-state representation. This representation unifies spatial memory and user interventions. Users specify object motion directly within spatial memory, while independently choosing the camera trajectory from which the resulting world state is observed. This unification makes the relationship between the scene and the requested changes explicit, removing ambiguity between object motion and viewpoint change from the control specification. Crucially, specifying an intervention does not require prescribing every resulting movement: the video model can synthesize additional motion within controlled objects and plausible interactions with their surroundings. We develop training and sampling strategies that help generated videos follow the specified object motion and camera trajectory while preserving this capacity to generate responses beyond the prescribed transformations. Experiments demonstrate more accurate joint object and camera control than the existing methods, alongside qualitative evidence that intervention-induced scene changes persist across rollouts.

[160] From Label Priors to Task Evidence: Efficient Long-Video Frame Selection via Bayesian GRPO cs.CVPDF

Shuochen Chang, Bingjie Gao, Qingyang Liu, Qianli Ma, Yibo Miao

TL;DR: 该方法将长视频帧选择视为潜变量推断，先用监督学习初始化轻量策略，再以冻结多模态大模型对答案 token 的似然作为任务证据，通过贝叶斯 GRPO 优化；在三个长视频基准上兼顾竞争力准确率与轻量推理效率。

Details

Motivation: 多模态大模型处理长视频受二次注意力复杂度和上下文长度限制，因此常需先筛帧。现有检索方法依赖静态关键帧标注，但这些标签可能与下游复杂推理真正需要的信息不一致。

Result: 方法先用监督数据训练轻量策略以学习一般视觉显著性，再把冻结的多模态大模型作为任务环境，以其答案似然通过贝叶斯 GRPO 优化选择策略。对真实答案 token 的负对数似然作为稠密任务奖励，匹配先前标注先验的 KL 项则维持语义一致性，使选帧依据从静态标签先验转向任务证据。Video-MME、MLVU 和 LongVideoBench 实验显示，选择器在保持轻量方法推理效率的同时，准确率与基于多模态模型评分器的方法有竞争力。选择器在冻结大模型环境中训练一次后，可直接迁移给多个下游多模态模型，无需再训练。

Insight: 帧筛选目标应对齐下游问答是否能从所选帧中回答，而非仅拟合通用关键帧标签。迁移能力使昂贵的训练期模型评分与部署期轻量选择分离；摘要未给出各基准的具体指标或效率数字。

Abstract: Multimodal Large Language Models (MLLMs) face significant scalability challenges in long-video understanding due to quadratic attention complexity and context window constraints. Existing retrieval-based approaches typically rely on supervised learning from static keyframe annotations. However, these label priors often fail to align with the actual information required by downstream MLLMs for complex reasoning, creating a misalignment between the training target and inference needs. To bridge this gap, we propose a probabilistic framework that reformulates frame selection as latent-variable inference. We first initialize a lightweight policy using supervised learning to capture general visual saliency. Then, we treat a frozen MLLM as a task environment and use its answer likelihood as dense task evidence to refine the policy via Bayesian Group Relative Policy Optimization (GRPO). Within this formulation, the negative log-likelihood (NLL) over ground-truth answer tokens provides a principled dense reward, while a prior-matching KL term preserves semantic consistency with the elicited label prior. This process transitions the selection principle from static label priors to task evidence. Extensive experiments on Video-MME, MLVU, and LongVideoBench demonstrate that our selector achieves competitive accuracy with MLLM-based scorers while retaining the inference efficiency of lightweight methods. The selector is trained once with a frozen MLLM environment and transfers directly to multiple downstream MLLMs without retraining. Code is available at \url{github.com/Miraclecsc/Frame_Selection_GRPO}.

[161] DragonMoCap: Monocular Video-based Motion Capture for Diverse Dragon Topologies cs.CVPDF

Zhongfei Qing, Xiaoxiang Zhao, Ling-Hao Chen, Zhongang Cai, Wenjia Wang

TL;DR: DragonMoCap 将目标龙角色骨架视为动态输入，通过语言关联的关节查询 token 表示不同拓扑，并用非参数查询检索实现未见骨架的零样本适配；在多种龙骨架上优于 HMR 类基线，并已接入专业动画流程。

Details

Motivation: 龙类角色的单目视频动作捕捉目前常依赖手工关键帧，原因之一是缺少高质量数据集和参数化模型。不同龙骨架的关节数量、命名和拓扑差异很大，使固定模板回归器难以直接适用。

Result: DragonMoCap 将目标 rig 作为输入而非固定先验，并限定覆盖制作中最常见的拓扑类型。其语言关联关节查询 token 同时追求同一 rig 内关节可区分、不同 rig 间表示相似；非参数查询 token 检索机制则支持对未见 rig 零样本适配，无需重新训练。多种龙拓扑评测中，该框架稳定优于 HMR 风格基线，并被整合进专业动画工作流。

Insight: 把骨架结构作为条件输入并以语义查询共享关节信息，可缓解拓扑变化导致的模板失配。作者强调目标范围是常见制作拓扑；摘要未说明未见 rig 的数量、捕捉精度指标或实时延迟。

Abstract: This work explores monocular motion capture for dragon-type characters, which currently relies on manual keyframing due to the lack of high-quality datasets and parametric models. The primary obstacle lies in the extreme topological diversity across dragon rigs, where varying joint counts and naming conventions render traditional fixed-template regressors inapplicable. To address this, we introduce DragonMoCap, a novel framework that treats the target rig as a dynamic input rather than a fixed prior. By deliberately scoping the framework to the most prevalent production topologies, we ensure high-fidelity results across diverse morphologies. The core of our approach is the design of language-grounded joint query tokens, which establish a dual-property representation: ensuring intra-rig discriminability while maintaining cross-rig similarity. We further propose a non-parametric query-token retrieval mechanism to enable zero-shot adaptation to unseen rigs without retraining. Extensive evaluations demonstrate that DragonMoCap consistently outperforms HMR-style baselines across diverse dragon topologies. The practical utility of our framework is further evidenced by its integration into professional animation workflows, highlighting its potential for real-time industrial applications.

[162] UBD-Bench: Benchmarking and Analyzing Unimodal Branch Degradation in Multimodal Perception for Autonomous Driving cs.CVPDF

Ang Liu, Xiang Gu, Ruikai Li, Qiyue Wang, Yizhe Fan

TL;DR: UBD-Bench 用配对冻结分支探针揭示自动驾驶多模态联合训练可能损害单模态表示：在四项任务和十个模型中，相机分支持续退化，且在融合前的 Camera BEV 表示最明显；针对退化表示蒸馏可同时改善分支与整模表现。

Details

Motivation: 多传感器融合可以提升自动驾驶整体预测，但输出层指标无法揭示每个模态分支的表征质量，联合训练甚至可能让某个单模态分支退化。论文提出系统测量这种单模态分支退化现象。

Result: UBD-Bench 使用配对冻结分支探针，在匹配解码条件下比较同一模态分别单独训练和多模态联合训练后的任务相关表示。四项自动驾驶感知任务、十个模型的结果显示，相机分支在联合训练后持续退化；LiDAR 分支变化较弱且更依赖模型。分层探针发现退化从视觉骨干延伸到 Camera-to-BEV 路径，在融合前的 Camera BEV 表示最严重。基于诊断，作者在退化表示处进行单模态到多模态蒸馏，四项任务上的相机探针表现和完整模型表现均稳定提高。

Insight: 多模态融合的整体收益可能与某一传感器分支变弱同时发生，因此评估不能只看融合输出，还应追踪模态专属表示及融合前路径。摘要未报告各任务具体退化幅度或蒸馏成本。

Abstract: Multimodal perception improves autonomous-driving performance by combining complementary sensor information. However, stronger multimodal predictions do not necessarily imply stronger representations in each modality-specific branch. Such branch-level weaknesses can remain hidden in conventional output-level evaluation. We introduce UBD-Bench, a systematic benchmark for studying unimodal branch degradation (UBD) in multimodal perception. UBD-Bench uses a Paired Frozen-Branch Probe to compare the task-relevant representation quality of the same modality after independent unimodal training and multimodal joint training under matched decoding conditions. Across four autonomous-driving perception tasks and 10 models, we find that Camera branches consistently degrade after multimodal joint training, whereas LiDAR branches exhibit weaker and more model-dependent changes, revealing a pronounced modality asymmetry. Hierarchical probing further shows that Camera-branch degradation extends from the visual backbone through the Camera-to-BEV pathway and is most severe at the pre-fusion Camera BEV representation. Guided by this diagnosis, we introduce unimodal-to-multimodal distillation at the degraded representation. This intervention consistently improves both Camera-branch Probe performance and full-model performance across all four tasks. These results show that multimodal gains can coexist with degraded unimodal representations, highlighting branch-level representation quality as an important dimension for evaluating and improving multimodal perception.

[163] Multimodal-Conditioned Latent Diffusion Models for Spatiotemporal Brain MRI Forecasting cs.CV | cs.AIPDF

Antonio Scardace, Silvia Ingala, Sebastiano Battiato, Daniele Ravì

TL;DR: MC-LDM 训练时结合 MRI、tau-PET、临床痴呆评分和 APOE4 多模态生物标志物，推理时仅需基线 MRI、APOE4 与患者元数据预测个体脑 MRI 进展；在 ADNI 独立测试集上，区域体积平均 MAE 比次优基线低 18%。

Details

Motivation: 阿尔茨海默病的三维脑结构变化预测对于疾病进展建模很重要。单模态 MRI 预测已有进展，但 tau-PET 和临床等互补生物标志物未被充分利用；另一方面，真实临床部署时又不一定能获得全部检查。

Result: MC-LDM 使用 LoRA 微调潜在扩散模型以提高数据效率，在训练时融合 MRI、tau-PET、临床痴呆评分（CDR）和 APOE4，建模异质病程。辅助预测器根据 MRI 特征和患者协变量估计推理阶段不可见的多模态标志物，从而在推理时只需基线 MRI、APOE4 和患者元数据。模型在 ADNI 的 852 名受试者、1,355 次访视上训练；在包含 664 次访视的独立测试集上，区域体积平均 MAE 比次优基线降低 18%。

Insight: 训练期丰富的多模态监督可通过辅助预测器迁移到临床可用的较少输入，帮助结构 MRI 预测捕捉单靠解剖外观难以体现的病程差异。摘要未说明各脑区误差、预测时间跨度和外部队列验证。

Abstract: Spatiotemporal forecasting of 3D anatomical changes is a critical generative vision challenge for modeling Alzheimer’s Disease (AD) progression. While brain MRI forecasting from single-modality data has advanced, other key complementary multimodal biomarkers remain underutilized. In this work, we propose the Multimodal-Conditioned Latent Diffusion Model (MC-LDM), a novel framework for patient-specific brain MRI synthesis. MC-LDM leverages multimodal biomarkers—MRI, tau-PET, Clinical Dementia Rating (CDR), and APOE4—during training, but relies only on baseline MRI, APOE4, and patient metadata at inference, reflecting realistic clinical availability. Our contributions are threefold: (1) a Latent Diffusion Model (LDM) fine-tuned via Low-Rank Adaptation (LoRA) for data-efficient learning; (2) a multimodal conditioning that integrates molecular and clinical signals to model heterogeneous disease trajectories; and (3) auxiliary predictors estimating unobserved multimodal biomarkers from MRI features and patient covariates to condition the diffusion model, removing the need for these assessments at inference. Trained on 1,355 visits from 852 subjects in the Alzheimer’s Disease Neuroimaging Initiative (ADNI), MC-LDM reduces average regional volumetric MAE by 18% over the second-best baseline on an independent test set of 664 visits. To our knowledge, MC-LDM represents the first diffusion-based framework to leverage multimodal priors for MRI forecasting. This conceptual advancement enables patient-specific trajectory generation that captures aspects of AD heterogeneity not observable from structural anatomy alone, bridging the gap between rich multimodal supervision and the strict constraints of routine clinical deployment. Code is available at: https://github.com/antonioscardace/MC-LDM.

Mingcong Ding, Babette Bühler, Tim Fütterer, Ulrich Trautwein, Enkelejda Kasneci

TL;DR: PRIMA 在冻结视觉语言模型中注入人体姿态提供的身体结构先验，用于识别细粒度社交身体行为；姿态只用于训练，不需推理输入，并可迁移到未见行为类别、数据集及部分其他 VLM 骨干。

Details

Motivation: 自我整理、姿势变化和倾听反馈等细微身体行为有助于社会信号分析，但重要行为会随社交语境和类别体系改变，为每种新分类重新训练模型成本较高。虽然 VLM 能用文本提示适配新类别，却常对身体结构不敏感，难以区分细微姿态差异。

Result: PRIMA 是轻量动作适配器，以训练时的人体姿态监督为冻结 VLM 注入身体结构先验，推理时仍只需 RGB。它将模型专属投影头与共享的身体构型知识核心分开，迁移到新 VLM 时只需重训投影头。作者在微动作数据集上训练并通过零样本提示评测；方法提升训练域识别表现，在冻结共享核心时迁移至不同规模的其他 VLM 骨干，并泛化到社会心理学和日常活动两个未见数据集。下游研究中，改进的行为预测与合作场景中人类对融洽度和支配性的评分相关性更高。

Insight: 姿态可作为训练监督注入身体结构知识，而不增加部署时的姿态输入，使视觉模型能跨行为词表和部分骨干复用先验。证据限于测试的模型与室内数据集；摘要未给出识别指标的具体增幅。

Abstract: Recognizing fine-grained bodily behaviors, such as self-grooming, postural shifts, and backchanneling, is a useful intermediate step for computational social signal analysis. But the behaviors that matter vary across social contexts, and retraining a model for each new taxonomy is expensive. Vision-language models (VLMs) can adapt to new categories through text prompts, yet they lack sensitivity to body structure and often confuse subtle postural distinctions. We propose PRIMA (Pose Refined Inductive Motion Adapter), a lightweight module that injects a body-structure prior into frozen VLMs using human pose as a training-time supervisory signal, requiring no pose input at inference. PRIMA separates a model-specific projection head from a shared core that encodes transferable body-configuration knowledge; adapting to a new VLM requires retraining only the head. We train PRIMA on a micro-action dataset and evaluate entirely through zero-shot prompting. The adapter improves recognition on its training domain, transfers to additional VLM backbones at different scales with the core frozen, and generalizes to two unseen target datasets from social psychology and daily activity recognition. In a supporting downstream study, the improved behavior predictions also yield higher correlations with human ratings of rapport and dominance in a collaboration context. Within the scope of the tested models and indoor datasets, the results indicate that pose-guided training can produce body-structure priors that generalize across unseen behavior vocabularies and can be partially reused across VLM backbones while preserving RGB-only inference.

[165] GroundedVDR: Grounding Video-to-Web Evidence Dependencies in Video Deep Research cs.CVPDF

Xiyuan Liang, Yu Wang, Zuhao Yang, Ziting Wang, Duo Li

TL;DR: GroundedVDR 是视频深度研究基准，分别检查答案是否依据视频片段和网络证据，并检查最终答案正确性；八个多模态智能体中最佳答案准确率为 76%，但没有系统能为超过 8% 的问题找全所需网页证据。

Details

Motivation: 现有视频深度研究基准主要评估工具使用轨迹和最终回答，却未核实模型是否真正使用了视频的时间信息和检索到的网页证据。单靠正确答案无法分辨模型是完成了跨模态证据整合，还是依赖参数记忆或单一信息源。

Result: GroundedVDR 结合人工审核与多模型受限访问探测，排除仅凭参数知识或单一来源即可回答的问题；每题带有结构化视频到网页证据标注，连接相关视频片段及其支持网页；评估时分别评分视频依据、网页证据依据和最终答案正确性。八个多模态智能体系统中最佳答案准确率为 76.0%，但没有一个系统能为超过 8% 的问题恢复完整的必需网页证据。

Insight: 深度研究评测应检验模型是否真正找到了并使用跨视频与网页的证据链，而不只看答案对错。结果显示答案准确与证据完整之间存在明显差距；摘要未报告各系统详细的分项分数。

Abstract: Video large multimodal models have improved at identifying cross-frame visual evidence and conducting open-web retrieval, as reflected in recent video deep research (VDR) benchmarks. However, existing VDR benchmarks primarily evaluate tool-use trajectories and final responses without verifying whether a model actually uses the temporal video information or the retrieved web evidence. To fill this gap, we introduce \method{}, an evidence-grounded benchmark with three features. First, it combines human review with restricted-access probing across multiple models to exclude questions answerable from parametric knowledge or a single information source. Second, each question carries structured video-to-web evidence annotations that link the relevant video segments to their supporting web evidence. Third, its evaluation protocol separately scores video grounding, web evidence grounding, and final-answer correctness. Across eight multimodal agent systems, the best reaches 76.0% answer accuracy, yet none recovers the complete required web evidence for more than 8% of questions. Our benchmark is publicly available at \url{https://huggingface.co/datasets/hh721/vdr_temp}.

[166] From Seeing to Reporting: A Theory of Reliable Visual Set Generation in Vision-Language Models cs.CV | cs.CL | cs.LGPDF

Junzhe Shi, Yuan Gan, Shida Jiang

TL;DR: From Seeing to Reporting（FSR）将视觉报告视为从模型回答到记录、集合和任务答案的完整推理链，并提出冻结 VLM 的推理时修正方法；合成场景逐类别计数误差最多降低 98.8%，ScreenSpot-Pro 点击定位漏检率由两阶段 GPT-6.1 缩放的 13.67% 降至 9.33%。

Details

Motivation: 视觉模型要计数、检测、解析屏幕或选择 GUI 点击目标，报告结果本身就是视觉任务的一部分。模型可能列出与场景完全匹配的全部对象，却在同一答案中把计数加错；仅改善感知或格式不能保证最终集合和答案正确。

Result: FSR 把视觉报告建模为响应、解码记录、构造集合及任务答案的连续链条，并以任务兼容匹配将各环节联系起来。理论关系用于分析删除记录、平均坐标和切换读数方式的影响，并把多次调用中择一答案的错误分解为候选缺失、代表丢失和错误选择。由此得到不需目标任务标签、仅依赖任务规格与固定配置的计数、检测、文档布局、屏幕解析和 GUI 定位流程。在 50 个密集合成场景中，单次报告与解码协议使 GPT-6.1 和 Claude Sonnet 5.5 的逐类计数误差分别减少 85.2% 和 98.8%。在 600 个 ScreenSpot-Pro 样本上，三模型定位层漏检率为 9.33%，低于 GPT-6.1 两阶段缩放的 13.67%；三次缩放后聚类投票为 12.67%，相近成本下三模型直接投票为 12.33%。

Insight: 错误可能发生在模型生成之后的解码、去重、集合构造和最终选择阶段，因此需要按任务语义检查整条报告链。GUI 结果表明，用其他模型提供候选、再进行任务兼容选择可减少漏检；该方法不等同于单纯多数投票。

Abstract: Vision-language models (VLMs) are asked to report what they see: to count objects, box them, parse a screen or pick a click target (GUI grounding). We argue that reporting is part of the visual task, not merely its formatting. In a dense scene with 173 shapes, GPT-6.1 lists 173 records that match every shape, yet the counts it states in the same response sum to 188. From Seeing to Reporting (FSR) is a theory of visual reporting and a set of inference-time procedures that improve the reports of frozen VLMs. It models reporting as a chain from responses to decoded records, constructed sets and task answers, linked by task-compatible matching. The chain gives exact relations for deleting records, averaging their coordinates and switching readouts, and it splits the error of one answer chosen among several calls into missing candidates, lost representatives and wrong choices. These relations guide procedures for counting, detection, document layout, screen parsing and GUI grounding that use the task specification and a shipped configuration, without requiring labels from the target task. On 50 dense synthetic scenes, a one-call reporting-and-decoding protocol cuts per-type count error by 85.2% for GPT-6.1 and 98.8% for Claude Sonnet 5.5. On 600 ScreenSpot-Pro items, a GUI-grounding tier that combines three models lowers the miss rate from 13.67% for GPT-6.1’s two-stage zoom to 9.33%, against 12.67% for three such zooms combined by a cluster vote at about the same cost. A plain vote over the same three models misses 12.33%: the other models supply the candidates, and selecting among them turns those candidates into fewer misses. We release the protocols, the stored responses and the analysis code.

[167] Modaliff: A Unified Token Diffusion Model for Any-to-Any Generation cs.CV | cs.AIPDF

Duc Kieu, Kien Do, Minh Khoa Le, Tung Kieu, Dang Nguyen

TL;DR: Modaliff 是统一 token 扩散模型，可从常见的双模态配对数据学习文本、图像和音频之间的任意方向生成，同时保留各模态自身潜序列；多数生成任务上在 any-to-any 模型中达到最高保真度，并使用更少参数和训练数据。

Details

Motivation: 单一模型若要灵活实现文本、图像、音频之间的任意转换，现有方案往往需要完全对齐的多模态样本组，或使用会压制模态特有结构的共享表示。论文寻求一种能使用更普遍双模态配对数据的统一生成方式。

Result: Modaliff 将条件和目标组织为统一 token 序列，用逐 token 属性区分 token 身份，并用角色化位置嵌入在不同序列布局中保留各模态内部结构。模型保留模态各自的潜在序列，不需要共享跨模态潜空间或按任务更改架构，单个去噪器即可处理多类生成任务。多条件生成可在推理时组合单条件预测，或用共同描述多个输入的字幕进行轻量微调。跨文本、图像和音频实验中，Modaliff 在多数生成任务上达到 any-to-any 模型中的最佳保真度，同时参数量和训练数据明显更少。

Insight: 统一序列建模不必抹去模态特有表示，也不一定要收集完整对齐的多模态元组；双模态数据和模态内结构编码即可支持广泛转换。摘要未给出各任务具体保真指标及资源节省幅度。

Abstract: Any-to-any generation seeks a single model that translates flexibly among text, image, and audio, yet existing approaches either require fully aligned multimodal tuples or a shared representation that suppresses modality-specific structure. We introduce Modaliff, a novel token-based diffusion model that learns from widely available bimodal pairs while preserving each modality in its native latent sequence. Modaliff assembles conditions and targets into a unified token sequence, with per-token attributes to disambiguate token identities and role-based positional embedding to preserve each modality’s internal structure across sequence layouts. This design enables a single denoiser to support diverse generation tasks without task-specific architectural changes or a shared cross-modal latent space. Beyond one-to-one generation, Modaliff supports multiple conditions either by combining single-condition predictions at inference or by lightweight finetuning on captions that jointly describe multiple inputs. Experiments across text, image, and audio show that Modaliff achieves the best fidelity among any-to-any models on most generation tasks while using substantially fewer parameters and training data.

[168] Residential Construction Image Dataset Using CSI Divisions for Deep Learning-based Object Detection cs.CV | cs.AIPDF

Fahad Iqbal, Shayan Mirzabeigi

TL;DR: ResCD-CSI 是按美国建筑规格学会 CSI 分部体系整理的住宅施工现场目标检测数据集，包含美国在建住宅现场高分辨率图像中的 16,365 个边界框标注对象，覆盖 14 个主要 CSI 分部。

Details

Motivation: 深度学习用于施工行业目标识别受限于缺少结构化、公开且符合行业标准的数据集，使建筑构件、材料和设备的视觉识别难以形成可比较、可复用的研究资源。

Result: 作者发布 Residential Construction Dataset Using CSI Divisions（ResCD-CSI），图像采自美国正在施工的住宅工地，涵盖多样建筑构件、材料和设备，并按照 Construction Specifications Institute 的 CSI 分部结构组织。数据集中共有 16,365 个对象，分属 14 个主要分部，均以边界框精细标注。该设计旨在提供标准化目标检测资源，连接施工实践与计算机视觉研究。

Insight: 把标签体系对齐建筑行业通用的 CSI 分类，有助于施工管理研究直接使用标准化对象类别。摘要介绍了数据规模和组织方式，但没有报告基于该数据集训练模型的检测性能，也未说明许可和采集地域以外的泛化情况。

Abstract: Accurate recognition of construction site elements is essential for the application of deep learning (DL) in the construction industry. Despite rapid advances in computer vision, the adoption of DL in construction has been limited by the lack of structured, publicly available datasets aligned with established industry standards. To address this gap, we introduce the Residential Construction Dataset Using CSI Divisions (ResCD-CSI) image dataset, a manually labeled image benchmark organized according to the divisional structure defined by the Construction Specifications Institute (CSI). The dataset comprises high-resolution images collected from active residential construction sites across the United States, capturing diverse building components, materials, and equipment under real-world conditions. In total, 16,365 objects spanning 14 major CSI divisions were precisely annotated using bounding boxes, ensuring consistency and reliability for object detection tasks. By aligning with the CSI divisional framework, ResCD-CSI establishes a standardized resource that bridges construction practice with computer vision research and supports the development of Artificial Intelligence-driven solutions for residential construction monitoring.

[169] DiffGround: Semantic Difference-Guided Zero-Shot Video Question Grounding cs.CVPDF

Cheng-Han Lee, Amir Salimi, Hsiang-Wei Huang, Che-Chun Su, Min Sun

TL;DR: DiffGround 通过从问题生成答案相关动作线索、选取关键帧并比较相邻 SigLIP2 语义嵌入来零样本定位完整视频动作片段；在 NExT-GQA 和 STAR 上超过已有零样本与监督式视频问题定位方法。

Details

Motivation: 视频问题定位既要答对问题，也要精确找到答案对应的时间片段。监督方法需要昂贵的时间标注；既有零样本方法多为视频问答设计，没有针对时间定位优化。

Result: DiffGround 基于冻结多模态大模型，先预测与答案相关的动作线索，把推理型问题转成明确的动作检索目标；再由视觉语言模型依据该线索选择有信息量的关键帧。Semantic Diff Refinement 无需训练，通过计算相邻 SigLIP2 嵌入的语义差异，从稀疏定位结果恢复完整动作片段。在 NExT-GQA 和 STAR 上，实验显示该方法持续超过已有零样本和监督方法，达到论文所报告的视频问题定位新最佳水平。

Insight: 将问题中的语义需求改写成动作线索，再用时间相邻特征变化细化边界，有助于把零样本问答扩展到片段定位。摘要未给出具体定位指标、改进幅度和跨数据集迁移结果。

Abstract: Video Question Grounding (VideoQG) requires both accurate question answering and precise temporal localization. Existing supervised methods rely on costly temporal annotations, while existing zero-shot approaches are primarily designed for Video Question Answering (VQA) and are not optimized for temporal grounding. We propose DiffGround, an action-driven zero-shot framework built upon frozen Multimodal Large Language Models (MLLMs). Our framework first predicts an answer-relevant action cue that transforms the input question into an explicit semantic target for grounding. Action Cue converts a reasoningoriented question into an action-centric retrieval query. Guided by the action cue, a Vision-Language Model (VLM) efficiently selects informative keyframes for reasoning. Finally, we introduce Semantic Diff Refinement, a training-free temporal refinement method that exploits semantic differences between neighboring SigLIP2 embeddings to recover complete action segments from sparse predictions. Extensive experiments on NExT-GQA and STAR show that DiffGround consistently outperforms existing zero-shot and supervised methods, establishing a new state of the art for Video Question Grounding.

[170] ViMoLa: Discrete Motion Tokens for Language-Guided Video Generation cs.CVPDF

Bishoy Galoaa, Sarah Ostadabbas

TL;DR: ViMoLa 将文本预测的离散运动 token 注入冻结图像到视频模型，改善语言引导视频生成的动作方向控制：在 46.2 万视频上方向准确率由 0.287 升至 0.442（随机水平 0.250），生成质量相近，但复合动作和 token—图像块对应仍有限。

Details

Motivation: 语言引导视频模型常能生成正确场景内容，却遗漏或误解提示指定的运动。论文希望先从语言估计动作先验，再将其注入冻结的视频生成骨干，显式提升运动控制。

Result: ViMoLa 用空间向量量化自编码器将稠密轨迹压缩为 64 个离散 token，码本大小为 512；一个 2,800 万参数 Transformer 从文本预测这些 token，并通过零门控交叉注意力适配器调节冻结 Wan 2.2 的每个 DiT 块。模型在 Something-Something v2 和 Kinetics-400 的 46.2 万视频上训练，方向准确率从 0.287 提升到 0.442，随机水平为 0.250，测得生成质量相近。跨提示打乱 token 后准确率为 0.283，支持运动 token 与提示匹配的重要性。低秩适配器达到 0.428，额外前向 FLOPs 为 1.45%；复用 tokenizer 和语言模型并为 CogVideoX 新训适配器时，准确率从 0.301 升至 0.398。oracle token 可达 0.531，说明仍有控制空间；复合运动增益较小，且适配器未显式保持 token 与图像块的对应。

Insight: 显式预测离散轨迹先验能帮助视频生成器更准确执行提示中的动作方向，而不仅是复现对象和场景。结果也表明运动控制尚未解决：oracle 与实际差距、复合动作较小收益以及缺少 token—patch 对齐都是局限。

Abstract: Language-guided video generation often captures scene content while missing the requested motion. We present ViMoLa (Video Motion Language), which predicts a motion prior from language and injects it into a frozen image-to-video backbone. A spatial vector-quantized autoencoder compresses dense trajectories into 64 tokens from a 512-entry codebook; a 28M-parameter transformer predicts the tokens from text; and zero-gated cross-attention adapters condition every frozen Wan~~2.2 DiT block through forward hooks. Trained on 462K videos from Something-Something~~v2 and Kinetics-400, ViMoLa raises directional accuracy from 0.287 to 0.442 (chance 0.250), with similar measured generative quality. Across-prompt shuffled codes achieve 0.283, supporting the importance of prompt-matched token content. A low-rank adapter reaches 0.428 at 1.45% additional forward-pass FLOPs, and reusing the tokenizer and language model with newly trained CogVideoX adapters raises accuracy from 0.301 to 0.398. Control remains imperfect: oracle codes reach 0.531, compound-motion gains are smaller, and the adapter does not explicitly preserve token-to-patch assignments.

[171] CheXFuse: Guided Differential Joint Attention Fusion for Robust Multimodal Chest X-Ray Classification in Data-Scarce Medical Settings cs.CV | cs.LGPDF

Muntaqim Ahmed Raju, Priyanka Siddappa, Md Shifat Haider Al Amin, Ruizhe Ma, Md Ripon Ahammed

TL;DR: CheXFuse 用引导差分联合注意力融合胸片图像和报告文本，降低弱模态冗余信号对融合的干扰；在 IU Chest X-ray 二分类中，原始报告条件下达到 95.73% 准确率和 95.79% F1，遮蔽报告后指标仍接近。

Details

Motivation: 多模态融合中，不同模态的预测能力可能差异明显；若不加区分地进行跨模态交互，较弱模态会引入冗余或不稳定信号，甚至使结果不如最强的单模态。论文针对胸片图像与文本报告设计更有选择性的融合。

Result: CheXFuse 采用 Guided Differential Joint Attention（GuDJA），将图像和文本 token 放入联合序列，以学习到的 token 引导突出信息，并通过差分注意力削弱较无用的响应。作者在 Indiana University（IU）胸片数据集上评估正常/异常二分类，比较原始报告和文本信息被遮蔽的压力测试设置。原始报告下，准确率、F1、召回率和 95% 特异度下敏感度分别为 95.73%、95.79%、96.54% 和 95.28%；遮蔽报告下分别为 95.58%、95.60%、95.60% 和 94.34%。在所评基线和消融中，GuDJA 的平均总体表现最强。

Insight: 当模态信息量不对称时，融合机制应选择性利用有效 token，而不是假设图像和文本同样可靠。遮蔽报告条件下性能接近原始报告结果，但证据限于单一数据集的二分类任务；摘要未给出外部临床验证。

Abstract: A central challenge in multimodal learning is how to fuse modalities that differ substantially in predictive strength. When one modality provides a much stronger representation than another, indiscriminate cross-modal interaction can introduce weak, redundant, or unstable signals, limiting or even degrading performance relative to the strongest unimodal representation. To address this problem, we propose CheXFuse, a multimodal fusion framework based on Guided Differential Joint Attention (GuDJA). CheXFuse combines image and text token representations within a joint multimodal sequence, where learned token guidance emphasizes informative representations and differential attention reduces less useful attention responses. We evaluate CheXFuse on the Indiana University (IU) Chest X-ray dataset, using binary (normal vs. abnormal) classification as the downstream task. Experiments are conducted under both original-report and masked-report settings, with the latter serving as a controlled stress test under reduced textual informativeness. CheXFuse with GuDJA achieves the strongest mean overall performance across the evaluated baselines and ablations, reaching 95.73% accuracy, 95.79% F1-score, 96.54% recall, and 95.28% sensitivity at 95% specificity (Sens@95Spec) with original reports, and 95.58%, 95.60%, 95.60%, and 94.34%, respectively, under masked reports. These results suggest that GuDJA provides a promising fusion strategy under asymmetric or partially reduced modality informativeness.

[172] Recurrent Vision Transformers for Event-Based Spacecraft Pose Estimation Under Slow Relative Motions cs.CVPDF

Sofia McLeod, Mohsi Jawaid, Carl Seubert, Tat-Jun Chin

TL;DR: 该方法用循环视觉 Transformer 的记忆从稀疏事件流累积长期信息，并从完整轨迹学习时间注意范围，以改善低相对速度航天器的事件相机位姿估计；真实慢速交会数据上，精度和稳定性优于逐帧或固定时间窗方法。

Details

Motivation: 事件相机能应对轨道环境中的强烈明暗变化，但追踪航天器相对位姿时，若目标与追踪器运动很慢，事件率会显著下降，数据噪声占比上升；低照度会进一步加重问题。真实交会与近距离操作还可能低至 1 mm/s，因此慢速运动下可靠估计是实际部署的关键。

Result: 作者将循环视觉 Transformer 用于事件式航天器位姿估计，通过循环记忆整合稀疏事件流中的长期信息，并从完整运动轨迹学习时间注意范围，而非预先固定窗口。基于 RVT 的热图回归器预测 2D–3D 对应关系，再用于位姿估计。真实低速相对运动事件数据上的验证显示，该方法相比逐事件帧方法和固定时间窗方法具有更高的位姿估计精度与稳定性。

Insight: 在事件稀疏、信噪比较低的慢速场景下，跨时间累积记忆可能比逐帧处理或固定窗口更可靠。摘要未提供精度数字、目标距离范围或不同照明条件下的定量结果。

Abstract: Vision-based object pose estimation is vital for conducting spacecraft rendezvous and proximity operations (RPO). Recent research has demonstrated the robustness of event cameras against the harsh lighting conditions that often occur in orbit, thanks to the extremely high dynamic range of event sensors. However, event-based spacecraft pose estimation (EBSPE) suffers when the relative motion between the chaser and target is slow, since the event rate drops significantly, leading to data with low signal-to-noise ratio. Unfortunately, for safety reasons, real RPO trajectories often involve slow relative motions (e.g., down to 1 mm/s), and operating under low illumination exacerbates the problem. Addressing slow relative motions is thus crucial to raise the practicality of EBSPE. To this end, we propose using recurrent vision transformers (RVT) for EBSPE. The key idea is to exploit recurrent memory to capture longer-term information across sparse event streams, and to learn the temporal attention span from full trajectories rather than fixing the time window. We developed an RVT-based heatmap regressor to produce 2D-3D correspondences for pose estimation, which was validated on real event data collected under slow relative motions representative of RPO trajectories. Results indicate the superior pose estimation accuracy and stability of our RVT-based method, relative to previous per-event-frame-based or fixed-time-window EBSPE methods.

[173] Confidence-Guided Pseudo-Label Refinement for Unsupervised Cross-Domain Person Re-Identification cs.CVPDF

Sundas Iqbal, Danish Ali, Qing Tian, Yuhui Zheng, Jianping Gou

TL;DR: CGPLR 根据聚类中心兼容性、邻域一致性、与竞争簇间隔及跨轮稳定性评估伪标签可信度，再将样本选择性保留、重分配或暂缓强监督；在 Market-1501 与 MSMT17 两个迁移方向上取得较高 Rank-1 和 mAP。

Details

Motivation: 无监督跨域行人重识别会因域偏移产生噪声伪标签，而迭代适配又可能不断强化错误。依赖少量可靠性线索或全局阈值的聚类方法可能把模糊样本也当作确定身份标签监督。

Result: CGPLR 联合评估样本与簇中心的兼容性、局部邻域一致性、最强竞争簇间隔以及跨轮分配稳定性，并据此设置簇自适应上下阈值，做三路决策：可信标签保留、几何和邻域证据充分时重新分配、极度模糊样本暂不施加身份监督。学习目标分别采用置信度加权监督、谨慎加权的纠正标签监督，以及对暂缓样本进行无标签表示一致性训练。Market-1501→MSMT17 上达到 76.4% Rank-1、47.6% mAP；反向迁移 MSMT17→Market-1501 上达到 95.1% Rank-1、87.7% mAP。

Insight: 伪标签利用不应只有接受或拒绝两种选择；将可纠正、可靠和高度模糊样本区分处理，能减少错误身份监督的累积。摘要未列出各基线对照值，也没有证明其在行人以外任务的泛化。

Abstract: Unsupervised cross-domain person re-identification remains challenging because domain shifts generate noisy pseudo labels that are reinforced during iterative adaptation. Existing clustering-based methods often rely on limited reliability cues or global confidence criteria and may impose hard identity supervision on ambiguous target samples. To address these limitations, we propose CGPLR, a confidence-guided pseudo-label refinement framework for unsupervised domain-adaptive person re-identification. CGPLR estimates sample-level reliability by jointly modeling centroid compatibility, local neighborhood consistency, separation from the strongest competing cluster and temporal assignment stability. The resulting confidence distribution is used to derive cluster-adaptive upper and lower thresholds for a three-way selective decision. Reliable assignments are preserved, recoverable samples are reassigned only when an alternative cluster receives sufficient geometric and neighborhood support and highly ambiguous samples are deferred from hard identity supervision. The learning objective adapts to these decisions through confidence-weighted supervision for preserved samples, cautious corrected-label weighting for reassigned samples and label-free representation consistency for deferred samples. CGPLR achieves 76.4% Rank-1 accuracy and 47.6% mean average precision (mAP) on Market-1501 to MSMT17, and 95.1% Rank-1 accuracy and 87.7% mAP on MSMT17 to Market-1501, demonstrating effective reliability-aware selective adaptation. The code is available at https://github.com/Sundas058/CGPLR.

[174] Clean Forcing: Drift-Resistant Autoregressive Video Diffusion with a Frozen Base cs.CV | cs.LGPDF

Wenqing Wang, Joonghyuk Shin, Jonathan Tremblay, Chan Hee Song, Yun Fu

TL;DR: Clean Forcing 将自回归视频扩散的长程漂移视为近乎确定的速度偏差，用 5.9M 参数 LoRA 在冻结基础模型上学习校正，并通过闭环自回放训练；生成 50 秒视频时，漂移指标从 +11.3 降至无真实视频训练的 +1.65，使用 40 段真实视频时降至 −0.16。

Details

Motivation: 流式自回归视频扩散模型以先前生成片段为上下文，训练时见到的干净上下文与推理时自身有误差的上下文不一致，误差会累积并在数秒内导致质量崩溃。已有方法需要大规模重训、推理时修正效果不稳定，或针对采样和缓存误差训练校正器。

Result: 作者发现固定噪声状态、只改变历史上下文时，反事实速度偏差在不同噪声种子间约 95% 是系统性的。Clean Forcing 用 5.9M 参数 LoRA 校正器在冻结基础模型上回归该偏差，再通过 DAgger 式数据聚合和漂移收缩目标让校正器接触自身闭环 rollout。训练历史可来自基础模型的一次性生成，因此无需真实视频；合并校正器后推理不增加成本。50 秒文生视频中，漂移 Δ 从 +11.3 降至 +1.65（不使用真实视频），或用 40 个真实片段降至 −0.16；这优于先前报告的 +3.57，所用真实数据少逾 150 倍。实验还报告审美质量和数据效率优于先前方法。

Insight: 长视频漂移可被当作可学习的系统误差，并不一定要用昂贵长视频或大规模基础模型重训来纠正。摘要没有定义 Δ-drift 的具体量纲，也未给出真实视频训练成本和各场景质量指标。

Abstract: Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated. However, they suffer from exposure bias, where a model trained on clean context must continue from its own imperfect outputs, and errors compound until quality collapses within seconds (drift). Existing methods either require cluster-scale base retraining, apply training-free inference corrections with inconsistent results, or train frozen-base correctors for sampling and cache errors. We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is ~95% systematic across noise seeds. Based on this observation, we introduce Clean Forcing, which trains a 5.9M-parameter LoRA corrector on the frozen base model to regress this gap, with a closed-loop stage that adds DAgger-style aggregation and a drift-contraction objective to expose the corrector to its own rollouts. Because the clean histories can be the base model’s own single-shot generations, training needs no real videos, and the merged corrector adds no inference cost. In 50 s text-to-video generation, Clean Forcing reduces Δ-drift from +11.3 to +1.65 without real video and to -0.16 with 40 real clips, improving on the best published result of +3.57 with over 150x less data. Extensive experiments demonstrate that Clean Forcing outperforms prior methods in aesthetic quality and data efficiency with reduced drift for long-horizon video generation.

[175] FluidGen: Physics-Aware Fluid Video Generation via Motion-Guided Post-Training cs.CV | cs.GRPDF

Yang Zheng, Zizhang Li, Boyang Deng, Jan Ackermann, Lei Shu

TL;DR: FluidGen 用模拟数据监督视频模型同步预测 RGB 帧和流体速度场，并将速度场用作偏好优化奖励，以增强烟、火和爆炸碎屑等流体画面的物理运动结构；作者报告其流体视频比强视频基线和优化方法更真实。

Details

Motivation: 通用视频生成模型虽然能合成视觉上吸引人的运动，却常无法生成连贯涡旋等流体结构，烟、火和爆炸碎屑会呈现单调、不符合真实流体物理的动态。经典求解器物理精度高，但依赖详细 3D 场景和精细参数，不适合野外图像及生成应用。

Result: FluidGen 将数据驱动的流体先验注入预训练视频模型，训练数据结合带配对速度场真值的高保真 Houdini 模拟和增加外观多样性的真实流体视频。双流架构在生成 RGB 的同时预测每帧速度场，并用模拟真值监督速度；这一辅助信号正则化运动，使其更符合物理一致性，同时提供可解释的物理量，进一步作为速度感知直接偏好优化的奖励。实验显示，相较强视频基线和基于优化的方法，FluidGen 生成的流体动态序列更真实。

Insight: 显式速度场既能为视频生成提供物理监督，也能为偏好优化提供可解释的运动信号，弥补纯外观训练缺少流体结构约束的问题。摘要未给出定量指标、模型规模或不同流体现象的分项结果。

Abstract: Video generation models produce visually compelling motion but systematically fail on fluid phenomena, including smoke, fire, and explosion debris, where outputs lack coherent vortical structure and collapse to bland dynamics that violate the physics of real fluids. Classical solvers achieve high physical fidelity but require detailed 3D scenes and carefully tuned parameters, making them impractical for in-the-wild imagery and video generation. We address this gap by injecting a data-driven fluid prior into a pre-trained video model, trained on a mixture of high-fidelity Houdini simulation data with paired velocity-field ground truth and curated in-the-wild fluid video data for appearance diversity. The core of our approach is a dual stream architecture that predicts per-frame velocity fields alongside RGB, supervised by simulation ground truth. This auxiliary signal regularizes generation toward physically consistent motion and exposes an interpretable physical quantity that we further leverage as a reward for velocity-aware direct preference optimization. Our model produces substantially more realistic fluid dynamics sequences than strong video baselines and optimization-based methods. Project page: https://y-zheng18.github.io/FluidGen/

[176] PoSTCache: From Forecasting to Postdictive Smoothing for Revisable Diffusion Caching cs.CVPDF

Liwei Lan, Hui Chen, Jungong Han, Guiguang Ding

TL;DR: PoSTCache 将扩散缓存从不可回退的前向近似改为可修订轨迹推断：后续完整计算提供误差信号，反向修正前序缓存预测；PVR 据预期轨迹不确定性而非局部误差选择计算锚点，在图像和视频生成中改善质量—效率权衡。

Details

Motivation: 扩散和流匹配模型通过反复采样获得高质量生成，但计算成本高。现有免训练缓存方法会把跳过计算时的近似立即写入轨迹，后续完整计算产生的反馈无法回头纠正这些误差。

Result: PoSTCache 先用现有缓存策略为跳过的步骤生成前向提案；在较后的完整计算步骤中，将刷新预测与前向提案的差异作为校正信号，向后传播以修正并重新整合此前预测。作者指出完整计算锚点的选择也应考虑反馈影响，因此提出 Posterior Variance Reduction（PVR）：从离线去噪轨迹估计不同时间步依赖关系，按对替代轨迹不确定性的预期降低量选锚点，而非只看局部近似误差。图像和视频生成模型实验显示，该框架持续改善相较代表性缓存方法的质量—效率权衡。

Insight: 生成轨迹中的缓存误差可以利用未来的完整计算反馈回溯修正，因此缓存不必是不可逆的前向近似。PVR 重视时间步之间的依赖而非单步误差；摘要未给出具体加速比例和质量指标。

Abstract: Diffusion and flow matching models achieve impressive visual generation quality but remain computationally expensive due to iterative sampling. Caching-based acceleration provides a practical training-free solution by reusing or predicting intermediate features. However, existing methods largely enforce a forward-only information flow: approximations produced for skipped model evaluations are immediately committed to the sampling trajectory and cannot benefit from feedback generated by later full-computation steps. We propose \textbf{PoSTCache} (\emph{Postdictive Smoothing Trajectory Cache}), a training-free and plug-and-play framework that reformulates diffusion caching as revisable trajectory inference. PoSTCache first uses an existing caching strategy to construct forward proposals for skipped steps. At a later full-computation step, the discrepancy between the refreshed prediction and its forward proposal yields a correction signal, which PoSTCache propagates backward to refine and reintegrate the preceding predictions. Viewing full-computation steps as sources of corrective feedback also changes how they should be selected. We therefore introduce the Posterior Variance Reduction (PVR) strategy, which estimates cross-timestep dependencies from offline denoising trajectories and selects anchors according to their expected reduction in surrogate trajectory uncertainty rather than local approximation error alone. Experiments on representative image and video generation models show that PoSTCache consistently improves the quality–efficiency trade-off over representative caching methods.

[177] Frechet Forcing: Training Long-Horizon Streaming World Models from Short Clips cs.CVPDF

Yian Zhao, Yu Yan, Hongcan Guo, Ruochong Zheng, Jian Zhang

TL;DR: Fréchet Forcing 仅用短真实片段训练流式世界模型与其多步采样器，以生成历史上的场景保真、逼真度和动作条件为约束，不依赖教师、在线评论器或长视频；MiniWorld-1B 用至多 36 GPU 小时微调后可稳定生成超过 4,000 帧并遵循机器人操作和相机探索动作。

Details

Motivation: 流式视频世界模型逐段预测未来并以自身历史为条件，若只在短片上训练，超过训练时长后误差会反复进入上下文而累积。常见分布匹配蒸馏需要教师，但外部视频教师不理解动作，自身预训练模型又有同样漂移问题。

Result: 作者观察到长 rollout 不必逐帧匹配某个教师，只需延续自身场景、保持真实感并遵循控制动作。Fréchet Forcing 联合微调模型和部署时使用的多步采样器，施加三种约束：在以真实数据结尾的自生成历史上使用流匹配损失；用以各短片真值为中心的潜变量统计 Fréchet 距离保持场景保真；在冻结感知空间中匹配真实帧统计以维持逼真度。三个目标均由短真实片段构造，不需教师、在线评论器或长视频。MiniWorld-1B（预训练片段 253 帧）用至多 36 GPU 小时在 381 帧片段上微调后，可稳定生成超过 4,000 帧，并在机器人操作和相机控制探索中持续遵循动作。

Insight: 长时世界模型训练可用场景统计和真实帧统计约束自生成 rollout，而无需与记录的长未来逐帧对齐；这让短片数据也能监督远超训练片段长度的行为。摘要未报告不同场景的定量真实度或其他模型规模验证。

Abstract: Streaming video world models generate the future chunk by chunk, conditioned on their own past and a stream of actions. Trained on short clips, they drift once a rollout exceeds the training horizon, as prediction errors re-enter the context and compound. A common remedy is to train the model on its own rollouts via distribution matching distillation (DMD). For action-conditioned world models, the teacher becomes a bottleneck: an external video model is agnostic to the actions, whereas the model’s own pretrained network drifts in the same manner as the student. We observe that a long rollout need not match a teacher, provided it remains a realistic continuation of its own scene that follows the commanded actions. Realism and scene fidelity can be measured against real-data statistics, which remain valid at any horizon, while action following is preserved by the model’s conditional generation loss. We propose Fréchet Forcing, which fine-tunes a world model together with its deployed multi-step sampler under three constraints. Generative anchoring applies the flow-matching loss after a self-generated history ending on real data. Context fidelity, a Fréchet distance (FD) on latent statistics of the free rollout centred on each clip’s own ground truth, keeps every rollout faithful to its scene. Reality matching, an FD to real frames in a frozen perceptual space, keeps the generated frames realistic. Because all three targets are derived from short real clips without comparing a rollout against a long recorded future, training requires no teacher, no online critic and no long videos. Fine-tuning MiniWorld-1B, pretrained on 253-frame clips, for at most 36 GPU-hours on 381-frame clips yields rollouts that remain stable beyond 4,000 frames and continue to follow their actions in both robot manipulation and camera-controlled exploration.

[178] Think Densely, Act Sparsely: Latent Expert Cognitive Chains for Vision-Language-Action Autonomous Driving cs.CV | cs.ROPDF

Jie Wang, Guang Li, Zhijian Huang, Jinlong Li, Chenxu Dang

TL;DR: LECDrive 通过递进的潜在专家认知链，让自动驾驶 VLA 在潜空间吸收视觉、语义和几何密集表征，再输出稀疏指令或轨迹；在多个驾驶基准的目标感知、状态预测和轨迹规划任务上持续改善基础 VLM。

Details

Motivation: 安全驾驶需要密集理解视觉、语义和几何信息，但现有视觉—语言—动作模型常以语言指令或轨迹等稀疏监督训练，直接把密集视觉观察映射到稀疏输出会迫使模型压缩乃至丢弃关键场景信息。论文认为问题不只是模型容量不足，还缺少促使潜空间形成稠密世界表示的机制。

Result: LECDrive（Latent Expert Cognitive Chains for VLA Driving）加入紧凑的任务专属潜在 token，作为载体吸收基础视觉模型知识：DINOv3 的视觉表示、SAM3 的语义结构和 DepthAnything3 的空间几何。其相互依赖的认知链按课程学习逐步提供从表示、语义到几何的监督，同时保持推理效率。多个自动驾驶基准实验中，LECDrive 在目标感知、状态预测和轨迹规划等任务上持续提升基础 VLM 的表现。

Insight: 先在潜空间组织密集场景认知，再生成稀疏驾驶决策，可减轻直接端到端压缩观察造成的信息丢失。摘要未提供各基准指标、推理开销或组件消融数值。

Abstract: Autonomous driving fundamentally demands dense understanding of visual, semantic, and geometric information, yet existing Vision-Language-Action (VLA) models are typically trained with sparse supervision such as language instructions or trajectory signals. This direct mapping from dense visual observations to sparse representations forces the model to compress or even discard critical scene information essential for safe driving. We argue that this limitation does not stem from insufficient model capacity, but from the lack of explicit mechanisms that encourage the formation of dense world representations in the latent space. In this spirit, we propose a novel paradigm for driving intelligence, Think Densely, Act Sparsely, and instantiate it with an efficient and interpretable framework, LECDrive (Latent Expert Cognitive Chains for Vision-Language-Action autonomous Driving). The core idea is to mimic the hierarchical perception process of humans by embedding a progressive chain of latent expert cognition within the VLA model. Concretely, we inject a compact set of task-specific latent tokens as carriers, enabling the model to internalize dense knowledge from foundational vision models, including visual representations (DINOv3), semantic structures (SAM3), and spatial geometry (DepthAnything3), while maintaining efficient inference. This interdependent cognitive chain follows a curriculum learning principle, where supervision is progressively structured from representation to semantics and geometry. Extensive experiments demonstrate that LECDrive consistently improves the performance of base VLMs across multiple autonomous driving benchmarks, covering key tasks such as object perception, state prediction, and trajectory planning. These findings suggest that constructing dense world models through structured latent expert chains is a promising direction for VLA-based autonomous driving.

[179] Functional Cross-layer Experts Routing for One-step Diffusion Distillation cs.CVPDF

Zile Huang, Ser-nam Lim

TL;DR: DeMoE 为一步扩散蒸馏保留始终激活的共享专家，并在图像 FFN 中路由条件化残差专家；在 FLUX.2-4B 上仅激活 80.5% FFN 宽度即得 GenEval 0.8220，整体匹配 50 步教师且在复杂组合与计数上超过教师。

Details

Motivation: 一步蒸馏把多步扩散压缩为一次前向计算，但单次计算仍需同时保留通用生成先验并处理提示中特定细节。论文研究在唯一一次前向中如何分配内部容量，指出只有专家相互干扰超过共享迁移收益时，表示解耦才有帮助。

Result: DeMoE 在图像 FFN 中组合持久共享专家与按条件路由的专家切片，并采用带已执行路由监督的一步分布匹配目标训练。在 FLUX.2-4B 一步生成中，模型只激活 80.5% 图像 FFN 宽度，GenEval 得分为 0.8220，明显优于稠密一步基线；总体表现与 50 步教师相当，并在复杂构图和计数任务上超过教师。在 COCO-10K 上，它在单步模型中达到当时最优的保真度和多样性。

Insight: 一步生成不必在完全共享和传统孤立 MoE 间二选一；持续共享主干保留通用先验，条件路由残差可把额外计算集中到提示细节。摘要未报告对不同教师或模型规模的系统结果。

Abstract: One-step distillation compresses diffusion trajectories into a single forward evaluation, yet how internal capacity should be structured within that remaining pass remains largely unexplored. In this setting, identical parameters must simultaneously preserve general generative priors and resolve prompt-specific details without iterative refinement. We formalize this challenge as functional allocation. Our analysis reveals that decoupling representations is beneficial only when expert interference outweighs shared transfer, favoring an always-active shared backbone augmented with conditionally routed residuals over standard isolated MoE architectures. We propose DeMoE, which pairs a persistent shared expert with conditionally routed expert slices across image FFNs, trained under a one-step distribution-matching objective with executed-route supervision. On one-step FLUX.2-4B generation, DeMoE achieves a GenEval score of 0.8220 while activating only 80.5% of the image-FFN width, substantially outperforming dense one-step baselines, matching the 50-step teacher overall, and surpassing it on complex composition and counting. Across COCO-10K, DeMoE establishes state-of-the-art fidelity and diversity among single-step models, demonstrating that structured internal computation reallocation provides a superior compression paradigm for diffusion models. Code is available at https://github.com/toffeecat/demoe-ttt-router.

[180] VHA: VQA-based Hallucination Assessment for Generative Image Super-Resolution cs.CVPDF

Donghun Ryou, Inju Ha, Jaeyoo Park, Bohyung Han

TL;DR: VHA 用双向视觉问答比较超分图像与真值图像的语义及低层证据，判断新增细节是否有依据、原信息是否被删除或篡改；配套轻量预测器 VHA+ 直接估分，二者在人工构建的 VHA-bench 上与人类幻觉判断高度相关。

Details

Motivation: 生成式图像超分辨率能产生逼真的高分辨率细节，却可能添加参考图中没有的内容，或改变原有细节。常用超分指标衡量视觉质量，却难以判断生成细节是否仍有真值图像支持。

Result: VQA-based Hallucination Assessment（VHA）让视觉语言模型从真值图像和超分图像分别构造并回答问题，再比较两个方向的分数差异，以发现真值信息缺失、被改动或无根据新增的内容。为减少昂贵的 VLM 推理，作者还训练 VHA+，直接从图像对预测 VHA 分数。通过人工研究构建的 VHA-bench 评估显示，VHA 和 VHA+ 都与人类对超分幻觉的判断高度相关。

Insight: 超分评估需要检查语义和低层视觉证据是否由输入支持，而不只看输出是否逼真；双向提问可以同时发现遗漏/修改和凭空添加。摘要未报告人类相关性的具体数值、VHA+ 的速度优势或对不同退化类型的表现。

Abstract: Generative image super-resolution produces visually realistic high-resolution images, but may also introduce hallucinations by adding unsupported content or altering existing details. Because existing super-resolution metrics fail to measure whether generated details remain grounded in the reference image, we present a hallucination assessment technique based on bidirectional visual question answering (VQA), which we term VQA-based Hallucination Assessment (VHA). The proposed approach uses a vision-language model (VLM) to determine whether the semantic content and low-level visual evidence in a super-resolved image are supported by its ground-truth counterpart. To this end, it constructs question–answer pairs from each image and evaluates them on both images. Discrepancies between the resulting scores in both directions reveal missing or altered ground-truth information and unsupported additions. However, obtaining these scores requires computationally expensive VLM inference. We therefore develop VHA+, a learned, lightweight model that directly predicts VHA scores from ground-truth and super-resolved image pairs. We construct a hallucination benchmark, VHA-bench, through a human study and show that both VHA and VHA+ correlate highly with human judgments of hallucination. The code, benchmark, and model weights are available at https://github.com/dhryougit/Hallucination-Assessment

[181] UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models cs.CVPDF

Jianing Wang, Youwen Jian, Derong Wang, Jinyang Yu, Qingyi Si

TL;DR: UniJev 用共享决策头联合训练文本、图像和视频的候选答案评分器；研究发现监督微调可提升三种模态准确率，但现有 RL 校准算法效果不稳定。独立校准目标 Cora 在 2B 和 9B 模型上降低图像与视频 ECE，同时保持准确率。

Details

Motivation: System One 决策模型直接输出供软件使用的概率判断，因此置信度可靠性与答对同样重要。但文本、图像和视频的联合决策训练尚少，监督微调与面向校准的强化学习各自贡献也难以区分。

Result: UniJev 以共享头对候选答案打分，并在文本、图像和视频上联合监督训练。与骨干模型直接读数相比，监督决策模型在三种模态上都提升准确率；现有 RLCD 算法对准确率影响较小，对视觉置信度校准的效果则因训练方案而不一致。作者因此提出单独的置信度目标 Cora（Confidence–Outcome Reliability Alignment）：按预测置信度分箱，并将各预测置信度优化到对应区间的经验准确率。2B 和 9B 规模实验中，Cora 在不经过强化学习阶段、不损害准确率的情况下，降低图像和视频的期望校准误差（ECE）。

Insight: 准确率提升和置信度校准是不同目标，联合多模态监督不会自动保证视觉概率可靠；单独对齐置信度与经验结果可避免依赖不稳定的 RL 校准。摘要未给出 ECE 降低的具体值或文本模态校准结果。

Abstract: System One models produce probabilistic judgments that software can consume directly, making reliable confidence as important as answer accuracy. Yet joint training across text, images, and video remains underexplored, and the contributions of supervised fine-tuning (SFT) and reinforcement learning for calibrated decisions (RLCD) are difficult to separate. We introduce UniJev, a trimodal framework that scores candidate answers with a shared head under joint supervised training across text, images, and video. Supervised decision models improve accuracy over their backbone readouts across all three modalities, while the existing RLCD algorithms change accuracy little and have inconsistent effects on visual calibration across recipes. These findings motivate a separate supervised objective for confidence calibration. We introduce Cora (Confidence–Outcome Reliability Alignment), which groups current predictions by confidence and optimizes each prediction’s confidence toward the empirical accuracy of its bin. At both 2B and 9B, Cora lowers image and video expected calibration error (ECE) while preserving accuracy, without a reinforcement stage. We release our code at https://github.com/wjnwjnwj/UniJev.

[182] Do We Need Learned Targets? Classical Gradients as Supervision for 2D Gaussian Initialization cs.CVPDF

Kyungjune Baek

TL;DR: 论文用无需离线逐图优化的 Sobel 梯度作为二维高斯初始化监督信号：解析式初始化平均 0.32 秒，神经初始化可在线训练并扩展至 1 万张 LSDIR 图像；在 Manga109 和 Kvasir-SEG 等域偏移设置中超过 Instant-GI。

Details

Motivation: 基于学习的高斯图像初始化方法可能需要对每张图像先做昂贵优化来生成监督目标。论文检验能否以经典图像梯度替代 Instant GaussianImage 使用的 Pseudo-PPM 目标，降低数据准备和训练成本。

Result: 框架用归一化 Sobel 图构建两条路线。免训练解析初始化结合概率增强、并行网格采样和几何处理，平均 0.32 秒生成初始参数。神经初始化在 DIV2K 上用在线 Sobel 监督训练，免去离线目标优化，并可扩展到 1 万张 LSDIR 图像而无需额外预处理。两种路线在 Kodak 和 DIV2K 上的长预算重建表现有竞争力。域偏移下，DIV2K 训练的神经初始化从 500 次拟合迭代起超过 Instant-GI，在 5K 迭代时高 1.40 dB；用 LSDIR 训练后差距增至 1.85 dB。免训练解析初始化在 Kvasir-SEG 的所有报告预算下都超过 Instant-GI。受控流程时间对比估计 DIV2K 管线缩短 4.9 倍。

Insight: 简单且便宜的解析梯度可替代昂贵的逐图优化监督目标，改善初始化扩展性，并可能增强域外重建。摘要没有给出各数据集完整重建指标或所有预算下的数值。

Abstract: Learning-based Gaussian image initializers can require costly supervision obtained by per-image optimization. We investigate whether the Pseudo-PPM targets used by Instant GaussianImage can be replaced by classical image gradients. Our framework uses normalized Sobel maps in two tracks. A training-free analytic initializer combines probability boosting, parallel grid-based sampling, and geometric processing, producing initial parameters in 0.32 seconds on average. A neural initializer is trained on DIV2K with online Sobel supervision, which removes offline target optimization and allows training to scale to 10,000 LSDIR images without additional preprocessing. Both tracks achieve competitive long-budget reconstruction on Kodak and DIV2K. Under domain shift, the DIV2K-trained neural initializer surpasses Instant-GI on Manga109 from 500 fitting iterations onward (by 1.40 dB at 5K iterations), and training on LSDIR increases this margin to 1.85 dB. On Kvasir-SEG, the training-free analytic initializer exceeds Instant-GI at every reported budget. A controlled pipeline-time comparison yields an estimated $4.9\times$ reduction for DIV2K. These results establish inexpensive analytic targets as a practical alternative to optimization-derived supervision for scalable Gaussian initialization.

[183] ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models cs.CVPDF

Zhiyang Deng, Yuzhe Wu, Zijun Lin, Yeying Jin

TL;DR: ActionRoPE 把平面相机游戏中的动作转换为世界坐标旋转位置，使同一世界位置跨时间拥有一致 token 坐标，并按已观测与新揭示区域安排去噪进度；相较四种控制机制，动作跟随更准确，重访漂移比最强基线降低 66%。

Details

Motivation: 游戏世界模型虽能实时交互生成，但动作控制通常只是特征条件，让模型自行从数据学习动作如何移动画面。对俯视或等距等平面相机世界，角色运动会引起几何明确的相机平移，因此直接使用世界坐标可能更适合空间一致控制。

Result: ActionRoPE 是无需训练参数的动作接口，将动作引起的相机位移转换为世界坐标旋转位置编码，使同一世界位置的 token 跨时间沿用同一坐标。作者还区分初始观测已经覆盖、几何已知的区域与新显露、需要合成的区域，并提出坐标条件噪声调度（CNS）：按 token 设定去噪进度，让已观测区域先稳定，新内容以之为参照生成。与四种代表性动作条件机制比较，该方法更忠实执行动作，重访漂移较最强基线降低 66%。该接口迁移到 1.3B–33B 的视频骨干，并能处理分布外场景。

Insight: 对平面相机环境，动作可以表达为世界几何坐标变化，而非仅作为学习到的特征条件；显式区分已知和新生成区域也能帮助模型保持空间一致性。摘要未给出基准名称、漂移绝对值或对非平面相机的结果。

Abstract: Game world models have made rapid progress toward real-time interactive generation. Yet action control is still predominantly implemented as feature conditioning, leaving the model to learn from training data how each action should move the visual world. This formulation is poorly matched to a broad class of planar-camera worlds, such as top-down and isometric games. In these worlds, player motion induces a geometrically well-defined planar camera displacement. We observe that the key to spatially consistent control in such worlds is to represent the effect of each action directly in world coordinates. To this end, we introduce ActionRoPE, a parameter-free action interface that converts action-driven camera displacement into world-coordinate rotary positions. As a result, tokens corresponding to the same world location retain the same positional address across time. These world coordinates further expose a spatial asymmetry in generation: regions covered by the initial observation are geometrically known, while newly revealed regions remain to be synthesized. We therefore propose Coordinate-Conditioned Noise Scheduling (CNS), which assigns token-wise denoising schedules according to this known-new structure. It allows observed regions to stabilize earlier, while new content is generated with reference to them. Compared with four representative action-conditioning mechanisms, our approach follows the commanded motion more faithfully and reduces the revisit drift by 66% over the strongest baseline. The same interface transfers across video backbones from 1.3B to 33B parameters and to out-of-distribution scenes. Project Page: https://teawhite.cn/ActionRoPE/

[184] Rethinking Sparse-View 3D Gaussian Splatting: A Few-Shot Learning Perspective on Overfitting Prevention cs.CVPDF

Lejia Ye, Bin Wu, Jingjing Gao, Yiyang Wu, Xiaohu Zhang

TL;DR: 论文将稀疏视角 3D 高斯泼溅重构视为少样本学习和过拟合问题，通过光流引导伪视角扩充监督，并用运动感知 dropout 与早停正则化；在 SIGGRAPH Asia 2025 稀疏视角体积视频挑战赛中获得第一名。

Details

Motivation: 稀疏视角新视角合成对 AR/VR 和实时通信很重要，但有限观测下 3DGS 易过拟合、几何退化。既有工作多分别处理数据质量或训练正则，缺少统一且可解释的改进框架。

Result: 作者将每个高斯基元类比为神经元，以少样本学习视角统一使用深度学习正则。监督增强方面，通过光流引导伪视角合成与相机—场景联合优化，从稀缺观测构造更有效的训练信号；训练正则方面，引入运动感知 dropout 和早停，以抑制过拟合并维持动态场景稳定。在 SIGGRAPH Asia 2025 体积视频挑战的稀疏视角数据集上，方法显著超过当时的优化式和前馈式基线，并获得稀疏视角赛道第一名。

Insight: 稀疏视角问题既需从少量视角补出更有效监督，也需训练期抑制拟合噪声；少样本学习提供了整合这两类策略的视角。摘要未给出具体 PSNR 或与各基线的差值。

Abstract: Sparse-view novel view synthesis (NVS) is essential for AR/VR and real-time communication applications, yet remains challenging when reconstructing 3D scenes from limited viewpoints. While 3D Gaussian Splatting (3DGS) excels in dense-view scenarios, it suffers from overfitting and geometry degradation under sparse inputs. Current approaches often tackle these problems from a local perspective, addressing either data quality or training regularization individually, lacking a unified and interpretable framework for systematic improvement. We introduce a new perspective that reframes sparse-view 3DGS from a few-shot learning viewpoint, treating each Gaussian primitive as a neuron and leveraging deep learning regularization techniques. Under this unified framework, our approach generates more effective supervision from scarce observations through two strategies: (1) data-related supervision enrichment via optical flow-guided pseudo-view synthesis and joint camera-scene optimization; and (2) training-time regularization via motion-aware dropout and early stopping, which suppress overfitting while preserving dynamic-scene stability. Extensive experiments on the SIGGRAPH Asia 2025 volumetric video challenge sparse-view dataset demonstrate significant improvements over state-of-the-art optimization-based and feed-forward baselines. Our system won first place in the sparse-view track of the SIGGRAPH Asia 2025 challenge.

[185] Multi-scale Fusion of Fast Vision Mamba and Foundation Model Features for Patch-Level Gleason Grading in Computational Pathology cs.CVPDF

Zihe Zhang, Weimiao Yu, Piumi Don Simonge

TL;DR: 该病理分类器融合冻结病理基础模型与轻量 FastVim 多尺度特征，经 Mamba 编码空间上下文后预测前列腺 Gleason 等级；PANDA 五折评估的宏 F1 为 0.894、Gleason 5 F1 为 0.802，较最佳单尺度模型分别相对提升 12% 和 21%。

Details

Motivation: 前列腺全切片图像 Gleason 分级受细微形态差异、罕见类型导致的类别不平衡以及跨尺度诊断特征影响。病理基础模型常以单尺度提取 patch 表征，缺少更广空间上下文；在数据和计算受限条件下融合多尺度、互补模型仍具挑战。

Result: 方法将冻结的病理基础模型特征与轻量 FastVim 特征结合；FastVim 在 ImageNet 预训练并从基础模型蒸馏，可直接读取原始图像裁块。每个尺度使用 Mamba 对两类特征编码线性时间空间上下文，再由联合训练的预测头融合各尺度表示。在 PANDA 数据集的五折交叉验证中，宏平均 F1 为 0.894，Gleason 5 类 F1 为 0.802，相比最佳单尺度模型分别相对提高 12% 和 21%。对五种基础模型、五种分类器和三个尺度的控制实验显示收益可跨编码器复现，且较小尺度提升更大。

Insight: 结合大模型语义特征与轻量原始图像特征，并显式建模多尺度上下文，可改善罕见等级和需要更广视野的分级。摘要未报告跨医院外部验证、绝对基线指标或运行成本。

Abstract: Automated Gleason grading of prostate whole slide images (WSIs) enables faster and more consistent cancer risk stratification. However, it remains a challenging task due to the subtle morphological differences between grades, the severe class imbalance arising from rare patterns, and the diagnostic features that span multiple spatial scales. Pathology foundation models (FMs) extract strong patch representations but often operate at a single-scale, which limits spatial context and disproportionately affects classes that depend on broader context. The challenge of fusing multiple scales and complementary models, with minimal data and computational resources, remains unresolved. We propose a multi-scale classifier built on state-space models. Our method combines two complementary feature sources: a frozen foundation model, and a lightweight FastVim, pre-trained on ImageNet and distilled from the foundation model, that reads the raw image crops. At each scale, a Mamba model encodes the features from both sources with linear-time spatial context, and a jointly trained head fuses the encoded representations into one prediction. On the PANDA dataset, our method reaches macro F1 of 0.894 and Gleason 5 F1 of 0.802 under 5-fold cross-validation, a 12% and 21% relative improvement over the best single-scale model. A controlled study over 5 foundation models, 5 base classifiers and 3 scales confirms that the gain holds across encoders, with larger improvements observed at smaller scales.

[186] SAPO: Self-Supervised Attention Policy Optimization for Improving Temporal Video Grounding in Vision-Language Models cs.CVPDF

Zixuan Chen, Depeng Wang, Wei Chen, Ke Xu, Ya Guo

TL;DR: SAPO 利用 VLM 内部时间注意力作为无标签时序定位监督，以 GRPO 优化视频问题定位；研究发现注意区间在约 66% 的有效样本上优于模型输出区间，所提方法的相对 mIoU 提升在 Qwen2.5-VL-3B 上达 33.7%、Time-R1-7B 上达 5.4%，且保留通用视频理解能力。

Details

Motivation: 视频 VLM 语义理解较强，但精确时间定位仍不可靠。以绝对时间对齐帧的先前结果暗示，模型可能已在内部表示了时间位置，只是在输出时间戳时丢失精度；昂贵的逐帧时间标注限制进一步训练。

Result: 作者在四个 VLM 和两个定位数据集上发现，内部注意区间在约 66% 的有效样本中优于预测区间，将此称为 temporal generation gap。SAPO 用模型内部时间注意力作为自监督信号，通过组相对策略优化缩小这一差距，并用无需真值标注的准则筛选高价值视频—问题样本。多个初始能力不同的模型和数据集上，时序定位均有提升；相对 mIoU 增益在 Qwen2.5-VL-3B 上达 33.7%，在较强的 Time-R1-7B 上达 5.4%。提升可跨数据集迁移，同时保持一般视频理解能力。

Insight: 模型内部注意可能比显式生成的时间戳更准确，利用两者差异可在不依赖时间边界标签的情况下训练定位。所报增益因模型强弱不同，摘要未说明各基准绝对 mIoU 和推理开销。

Abstract: Video vision-language models (VLMs) have strong semantic understanding, yet remain unreliable at precise temporal grounding. Existing gains from explicitly aligning video frames with absolute time suggest that VLMs may already encode accurate temporal locations internally but lose precision when expressing them as timestamps. Our analysis across four VLMs and two grounding datasets confirms this hypothesis: attention intervals outperform predicted intervals on approximately 66% of valid examples. We term this discrepancy the temporal generation gap. Because precise temporal annotations are expensive, this finding motivates closing the gap to improve temporal grounding using only unlabeled video-query pairs. We propose Self-Supervised Attention Policy Optimization (SAPO), which uses internal temporal attention as self-supervision for Group Relative Policy Optimization and applies a ground-truth-free criterion to select high-value samples. SAPO consistently improves temporal grounding across multiple datasets and models spanning weak to strong initial grounding ability. Relative mIoU gains reach 33.7% on Qwen2.5-VL-3B and 5.4% on the strong Time-R1-7B. The improvements transfer across datasets while preserving general video understanding.

[187] MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports cs.CVPDF

Hongrui Wu, Kaiwen Tuo, Congcong Wang, Gen Li, Shuai Dong

TL;DR: MIRA 把文本与视觉证据贯穿多步网页研究和报告写作，并以需求合同连接证据收集、报告提交、反馈修订和最终交付；基于 Qwen3.6-35B-A3B，在七个多模态搜索与深度研究基准上达到开源智能体领先表现，且生成证据可追溯的定制图文报告。

Details

Motivation: 深度研究智能体虽能搜索网络并综合证据，却不适合按用户要求生成可控的多模态报告：推理时常忽略搜索中遇到的图片，生成时则趋向固定格式，难以控制报告详略、图文组合和证据依据。

Result: MIRA-Harness 将用户要求转成生成合同，连接证据收集、报告提交、基于反馈的修订和最终交付四阶段，使视觉与文本证据贯穿研究和写作。MIRA-Data 构建高质量多跳多模态研究轨迹，再后训练基础模型执行这一流程。使用 Qwen3.6-35B-A3B 时，MIRA 在七个既有多模态搜索和深度研究智能体基准上达到开源智能体中的最佳表现，并能生成符合用户要求、包含可追溯文本和视觉证据的图文报告。

Insight: 多模态深度研究需要把检索到的图像证据纳入推理和写作，并把用户要求落实为可迭代检查的报告流程。摘要未提供具体七项基准分数、报告质量的人评数据或相对基线增幅。

Abstract: Deep research agents are increasingly capable of searching the open web and synthesizing evidence, yet they remain poorly suited to controllable report generation in multimodal deep research. This task requires agents to handle interleaved textual and visual inputs, gather multimodal evidence through multi-step web search, and deliver comprehensive illustrated reports under user-specified requirements. Existing systems face two limitations. During reasoning, they often underuse images discovered along the search trajectory; during generation, they tend to produce fixed-form answers with limited control over report granularity, multimodal composition, and evidence grounding. We introduce MIRA, a Multimodal Illustrated Deep Research Agent for controllable reports. MIRA integrates textual and visual evidence throughout research and writing. At its core, MIRA-Harness translates user requirements into generation contracts and connects four stages: evidence collection, report submission, feedback-driven revision, and final delivery. Building on this harness, MIRA-Data constructs high-quality multi-hop multimodal research trajectories, and we post-train foundation models to operate within the workflow. Using Qwen3.6-35B-A3B, MIRA achieves state-of-the-art performance among open agents on 7 representative, established multimodal-search and deep-research-agent benchmarks, while producing user-aligned illustrated reports with traceable textual and visual evidence.

[188] CondenseVLA: Learnable History Condensation for Efficient Multi-Frame VLA cs.CV | cs.ROPDF

Feiyang Hong, Yujie Wei, Bo Zhao, Xiu Su, Hongxun Yao

TL;DR: CondenseVLA 将 VLM 缓存的多帧视觉特征用可学习查询压缩成固定 slot token，再用于动作解码；Simpler-WidowX 成功率从 58.7% 提至 69.8%，LIBERO 达 97.2%，真实任务达 52.8%，但依赖多次观察的 RMBench 任务仅 21.6%。

Details

Motivation: 机器人操作依赖短暂接触、遮挡和物体中间状态等单帧歧义线索，需要整合一段时间的观察。单纯扩大历史窗口成本高、冗余大且采样不随任务调整，往往收益有限。

Result: CondenseVLA 在解码器侧加入可学习历史压缩模块，将视觉语言骨干缓存的逐帧特征通过可学习查询和多头交叉注意力压缩为固定数量 slot token。辅助正则惩罚 slot 注意图重叠并去相关 slot 表示，以鼓励各 slot 关注互补信息、减少冗余。实验中 Simpler-WidowX 成功率由 58.7% 增至 69.8%（提高 11.1 个百分点），LIBERO 达 97.2%，真实世界任务达 52.8%；记忆依赖型 RMBench 任务为 21.6%。

Insight: 针对近期视觉历史进行选择性压缩，可避免扩大窗口引入大量冗余，同时保留动作解码需要的短时线索；困难的多观察任务也暴露固定有限窗口的边界。摘要未报告各任务的基线数值或推理成本。

Abstract: Robot manipulation often depends on short-term cues - such as brief contacts, temporary occlusions, and intermediate object states - that are ambiguous in a single frame and can only be resolved by integrating observations over time. However, injecting historical information solely by enlarging the history window is computationally costly and frequently yields limited gains due to substantial redundancy and task-agnostic, inflexible sampling of observations. We propose CondenseVLA, a multi-frame VLA policy with a learnable decoder-side history condensation module for using short-term visual context in action decoding. Given a sequence of cached per-frame features from the VLM backbone, CondenseVLA distills observation history into a fixed set of slot tokens using learnable queries and multi-head cross-attention. Simple auxiliary regularization penalizes overlap among slot attention maps and decorrelates slot embeddings, encouraging complementary attention and reducing slot redundancy. Across the evaluated tasks, CondenseVLA improves Simpler-WidowX success from 58.7% to 69.8% (+11.1 percentage points), achieves 97.2% on LIBERO, and reaches 52.8% on real-world tasks. It achieves 21.6% success on memory-dependent RMBench tasks. These results support selective condensation of bounded recent visual evidence for action decoding, while harder multi-observation tasks expose the limits of a finite history window. Project page: https://github.com/FeiyangHong/CondenseVLA

Zixuan Chen, Depeng Wang, Hao Lin, Ke Xu, Ya Guo

TL;DR: OmniGround 通过 TACG 将模型内部时间注意与预测时间戳及真值对齐，并用 58 万样本训练数据强化全模态定位；在 LongVALE 和 UnAV-100 分别达到 37.5 和 35.7 mIoU，同时保持一般音视频理解和纯视觉定位的竞争力。

Details

Motivation: 全模态大语言模型能联合理解视频、音频和文本，但时间定位仍较弱。论文发现模型内部注意力比最终输出时间戳更准确地定位目标时刻，认为这是内部时间证据到显式输出之间的生成差距。

Result: Temporal Attention-Consistency Grounding（TACG）后训练方法把内部注意力作为时间锚点，同时将其与模型预测和真值标注对齐。作者从密集音视频标注重建时间问答，并整理通用全模态指令数据，构建 58 万样本训练集；训练 OmniGround 时不额外注入时间线索。模型在 LongVALE 上达到 37.5 mIoU，在 UnAV-100 上达到 35.7 mIoU，均为论文报告的全模态时间定位先进结果，并在一般音视频理解及纯视觉定位上保持竞争力。

Insight: 将内部定位线索直接约束到时间戳输出，可缩小“模型注意到了哪里”与“模型报告了何时”的差距。摘要未给出与基线的具体增幅或对训练集各部分的消融结果。

Abstract: Omni-modal large language models can jointly understand video, audio, and text, yet they remain weak at temporal grounding, a capability essential for real-world audio-visual applications. To understand this limitation, we look beyond final predictions and examine the model’s internal temporal evidence. We identify a temporal generation gap: the internal attention of the model localizes the target moment more accurately than the timestamps it generates. To close this gap, we propose TACG (Temporal Attention-Consistency Grounding), a post-training method that treats internal attention as a reliable temporal pivot and aligns it with both the model prediction and the ground-truth annotation through attention-consistency training. We further build a 580K-sample training dataset by reconstructing temporal QA from dense audio-visual annotations and organizing general omni-modal instruction data. Using TACG and this training dataset, we train OmniGround, a general-purpose omni-modal model with stronger temporal grounding capability, without injecting any additional temporal cues. OmniGround reaches state-of-the-art omni-modal temporal grounding, including 37.5 mIoU on LongVALE and 35.7 mIoU on UnAV-100, while staying competitive on both general audio-visual understanding and vision-only grounding.

[190] HORIA: 4D Human and Object Reconstruction from a Monocular Video with Interaction-Aware Re-Rendering cs.CVPDF

Hyeongjin Nam, Taeryung Lee, Daniel Sungho Jung, Kyoung Mu Lee

TL;DR: HORIA 先用运动感知字幕把单目视频拆为时间定位明确、语义不同的人—物交互阶段，再从新视角重渲染初始 4D 重建并用文生视频扩散先验细化，以弥补遮挡造成的单视角证据不足；作者报告多种交互下重建连贯且达到先进水平。

Details

Motivation: 从单目视频联合重建人体与物体的 4D 交互时，视频中多个交互阶段的语义混杂，难以追踪其时间变化；人—物遮挡又使单视角观察不完整，容易产生几何不一致结果。

Result: HORIA 由两部分构成：用运动感知字幕把视频分解为时间定位明确、语义不同的交互阶段，并为每阶段生成专属描述；随后以这些描述为条件执行交互感知重渲染。具体地，先把初始重建的人体和物体从新视点渲染，再用文生视频扩散先验细化渲染结果，为单视角线索补充新视角监督。作者报告该框架在多样交互场景中产生时间连贯且准确的 4D 重建，并达到论文所述的先进表现。

Insight: 分阶段语义描述帮助跟踪交互的时间结构，新视角重渲染则为遮挡区域增加单视角输入之外的监督。摘要未提供具体数据集、指标、相较基线的提升幅度或生成先验引入的伪影分析。

Abstract: Joint 4D reconstruction of a human and object from a monocular video is essential for understanding how people interact with their environment. While recent methods have made noticeable progress, capturing such human-object interactions remains difficult due to two main challenges. First, a video typically contains composite semantic context spanning multiple interaction phases, making it difficult to track the temporal changes of the interaction in the reconstruction. Second, under human-object occlusion, a single-view observation contains incomplete visual information about the interaction, resulting in geometrically inconsistent reconstructions. To address these challenges, we introduce HORIA, built upon two core designs. First, we decompose the entire input video into temporally grounded and semantically distinct interaction phases with motion-aware captioning, obtaining a specific text description for each phase. Second, conditioned on these text descriptions of the interaction, we propose an interaction-aware re-rendering that compensates for the limited single-view observation. Specifically, we render the initially reconstructed 4D human and object from novel viewpoints and refine the renderings with a text-to-video diffusion prior, which provides auxiliary novel-view supervision beyond single-view cues. As a result, our framework produces temporally coherent and accurate 4D reconstructions under diverse interaction scenarios, achieving state-of-the-art performance.

[191] Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation cs.CV | cs.CLPDF

Kesav Kumar Jayakumar, Karthigeyan Thilak

TL;DR: POPE 系统审计发现，VLM 幻觉评测的 yes/no 词表读数选择会单独改变 F1：LLaVA-1.5-7B 的空格前缀 token 读数在对抗、热门和随机划分分别为 0.8221、0.8498、0.8713；跨 tokenizer 复用错误 token ID 甚至可使两个模型 F1 降至 0。

Details

Motivation: 作者尝试复现 POPE 上一项幻觉缓解结果时，发现 LLaVA-1.5-7B 的对抗划分基线比准备对比的基线高 1.21–8.67 分，随后检查评测代码并发现二 token yes/no logit 读数可能不匹配模型实际生成的词表项，导致评测协议本身混淆结果。

Result: 两 token 读数比较 ID 3582（yes）和 1217（no），但 9,000 个跨三划分的贪心解码问题中，这两个 ID 一次也未生成；模型实际输出带前导空格的 Yes/No（ID 3869、1939），原因是提示后缀后的 SentencePiece 空格字节。覆盖空格前缀表面形式的八 token 查找得出 LLaVA-1.5-7B 在对抗、热门和随机划分 F1 分别为 0.8221、0.8498、0.8713；对抗划分相差 6.13 个百分点，超过作者调查的六篇论文中五篇报告的对抗增益。论文说明由于相关论文未披露读数方式，且其参考运行使用 4-bit NF4 量化，在读数和精度未匹配前无法判断既有增益能否复现。固定基线上测试的九种推理期修正均未超过该读数基线。3,000 题探针发现错误预测比正确预测吸引更多图像注意（Cohen’s d=-0.46，p=1.4×10⁻²⁴），与基于 grounding 的修正假设相反。另对三种 tokenizer 家族的四个 VLM 做四种读数审计：直接照搬 LLaVA ID 会让两个词表不相交模型 F1 降至 0.00，而按 tokenizer 得出的读数保持在 0.80–0.92。

Insight: POPE 分数不仅取决于模型，也会被 yes/no tokenization 和读数协议显著影响；模型间比较必须匹配 tokenizer、输出读数和量化精度。论文明确指出既有报告之间的比较仍未解决，不能将读数造成的变化直接归因于模型或缓解方法。

Abstract: We set out to reproduce a recent hallucination-mitigation result on the POPE benchmark. The LLaVA-1.5-7B baseline we measured, 82.21 F1 on the adversarial split, sat 1.21 to 8.67 points above the adversarial baselines we planned to compare against (81.00, 79.30, and 73.54). We looked at the evaluation code and found a readout choice that moves F1 on its own. A two-token logit readout decides yes versus no by comparing two vocabulary IDs, 3582 (“yes”) and 1217 (“no”). Across 9,000 greedy-decode questions spanning all three splits, neither ID is generated even once. The model emits “ Yes” (token 3869) and “ No” (token 1939), because SentencePiece prepends a space byte after the prompt suffix “ASSISTANT:”. Read with an eight-token lookup that covers the space-prefixed surface forms, LLaVA-1.5-7B scores F1 0.8221, 0.8498, and 0.8713 on the adversarial, popular, and random splits. That 6.13-point adversarial shift exceeds the adversarial gain reported by five of the six papers we surveyed. None of them states its readout, and our reference ran under 4-bit NF4 quantization, so we cannot say how much of any reported gain a matched evaluation would keep. That comparison stays unresolved until readout and precision are matched. On the fixed baseline we run nine inference-time corrections of our own (none beat it) and dissect why the adversarial split resists logit-level intervention. In a 3,000-question probe, wrong predictions draw more image attention than correct ones (Cohen’s $d=-0.46$, $p=1.4\times10^{-24}$), the opposite of what grounding-based corrections assume. A four-readout audit across four further VLMs from three tokenizer families confirms the confound is protocol-level, not model-specific: porting the LLaVA-1.5 IDs verbatim drops F1 to 0.00 on the two disjoint-vocabulary models, whereas a tokenizer-derived readout holds between 0.80 and 0.92.

[192] FrameGRPO: Incentivizing Video Reasoning via Multi-Frame Rollouts and Temporal Credit cs.CV | cs.AIPDF

Jiaxuan Sun, Biao Wang, Shuai Li

TL;DR: FrameGRPO 通过同一次最大帧解码构造均匀时间子采样视图，增加视频证据和推理轨迹多样性；其门控时间效率信用只在固定全帧参考组全部失败时奖励正确的低帧率轨迹，多个视频理解与推理基准上整体优于标准 GRPO。

Details

Motivation: 标准视频 GRPO 在一组 rollout 中只变化语言回答，却让所有轨迹看到相同抽样帧；当共享视觉证据不足或带来干扰时，组内奖励可能趋同，削弱组相对学习。

Result: FrameGRPO 无需辅助网络或修改模型架构，组合两个组件：多帧 rollout 从一次最大帧解码中均匀时间子采样，构造对齐但视觉证据不同的视图；门控时间效率信用则在固定全帧参考子集的所有轨迹都失败时，选择性奖励一条正确的低帧率轨迹，而不无条件偏好少看帧。多个视频理解和推理基准实验显示，该方法相较标准 GRPO 有稳定的整体提升。

Insight: 视频强化学习中的探索不应只改变语言轨迹，也可在对齐的时间子视图之间变化视觉证据；奖励低帧率答案需要由全帧参考失败作为条件，避免牺牲正确性换取效率。摘要未给出各基准的具体增益或计算节省。

Abstract: Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has improved video reasoning in multimodal large language models. However, standard video GRPO varies only linguistic trajectories while conditioning every rollout in a group on the same sampled frames, which can produce homogeneous rewards and weaken group-relative learning when the shared visual evidence is insufficient or distracting. We propose \textbf{FrameGRPO}, a simple and effective reinforcement-learning framework that requires neither auxiliary networks nor changes to the model architecture. Specifically, FrameGRPO combines two lightweight components: (1) \textbf{Multi-Frame Rollouts}, which use only uniform temporal subsampling of a single maximal-frame decode to construct aligned views that diversify visual evidence and reasoning trajectories; and (2) \textbf{Gated Temporal Efficiency Credit}, which selectively reinforces a correct lower-ratio trajectory when all rollouts in a fixed full-ratio reference subset fail, without imposing an unconditional preference for fewer frames. Extensive experiments across diverse video understanding and reasoning benchmarks demonstrate robust overall improvements over standard GRPO, highlighting temporal-view diversification as an effective mechanism for exploration and credit assignment in video MLLMs.

[193] TamedFlow: Trajectory-Aware Distillation\for Teacher-Assisted Diffusion Sampling cs.CV | cs.AIPDF

Shuaiqi Cheng, Jiahao Zhang, Boyuan Gu, Siyu You, Yuxi Chen

TL;DR: TamedFlow 针对压缩扩散学生再叠加自适应推理时出现的组合性能落差，使用随机 rollout 监督训练其短时自回归执行能力，并以轨迹估计器让草稿跳过教师或回退后接受一次教师纠正；ImageNet 256×256 上 FID 为 2.58，实测速度比 DDIM-50 快 2.25 倍。

Details

Motivation: 扩散 Transformer 可通过压缩降低每次模型调用成本，也可通过缓存、预测或推测执行跳过部分教师调用；但把已有自适应推理直接用于压缩学生可能损害质量。论文将此称为 Combination Gap，认为逐点蒸馏在教师状态上训练，未保证学生连续使用自身预测时仍可靠。

Result: TamedFlow 通过跨架构特征对齐、近似恒等的 Velocity Alignment Head 和随机 rollout 监督，让压缩学生适应部署时的短程连续执行。推理时，Learnable Trajectory Estimator（LTE）判断已完成草稿：接受则跳过教师，拒绝则回滚并用一次教师步骤补救。将 DiT-XL/2 蒸馏为 DiT-B/2 后，在 ImageNet 256×256 上 FID 为 2.58，实测速度相较 DDIM-50 提升 2.25 倍。消融支持轨迹监督的作用；在 ImageNet 512×512、MS-COCO 和 Latte/FaceForensics 上也进行评估。

Insight: 压缩模型必须针对自身预测被连续采纳的部署轨迹训练，单纯匹配教师中间状态不足以保证叠加跳步推理仍稳定。摘要提到更广评估，但未给各数据集结果和路由额外成本。

Abstract: Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two acceleration strategies are increasingly important: model compression makes each evaluation cheaper, while adaptive inference avoids selected teacher computation through reuse, forecasting, or speculative execution. Their combination, however, is brittle: directly applying the evaluated adaptive methods to a compressed student can degrade quality, an observed failure mode we call the \textbf{Combination Gap}. We link this gap to a mismatch between pointwise distillation and short-horizon execution: a student trained at teacher states need not remain reliable when its own predictions are chained. We propose \textbf{TamedFlow}, which trains a compressed DiT for this deployed role through cross-architecture feature alignment, a near-identity Velocity Alignment Head, and randomized rollout supervision. At inference, a lightweight Learnable Trajectory Estimator routes completed drafts: accepted drafts bypass the teacher, while rejected drafts roll back and receive one teacher rescue step. Distilling DiT-XL/2 into DiT-B/2, TamedFlow reaches FID $2.58$ at $2.25\times$ measured speedup over DDIM-50 on ImageNet $256\times256$. Ablations support trajectory supervision, and routing controls favor LTE over random decisions at nearly matched latency. Evaluations on ImageNet $512\times512$, MS-COCO, and Latte/FaceForensics extend the empirical scope.

[194] AVTime: Reinforcing Dense Audio-Visual Captions with Bidirectional Alignment cs.CVPDF

Mingyang Wu, Kaituo Feng, Bohao Li, Kaixiong Gong, Zihao Yin

TL;DR: AVTime 构建高时间分辨率的 5 万规模音视频字幕集，并提出双向时间—语义对齐奖励 BiTSA、配套双向评测基准 AVTime-Bench，以同时评价事件语义正确性和时间精确度。

Details

Motivation: 密集音视频字幕对于多模态视频理解与生成很重要，但现有资源缺少同时覆盖高时间分辨率、较长视频和时间—事件双向对齐的数据；许多 RL 方法依赖清单式奖励，基准也没有双向评估事件语义与时间定位。

Result: 作者提出 AVTime-50K，包含时间精确、语义准确且视频较长的 5 万规模字幕数据；提出带双向时间—语义约束的新奖励 BiTSA，并据此优化 AVTime 模型；同时构建 AVTime-Bench，双向评估事件语义正确性与时间准确性。

Insight: 密集字幕质量不能只看文本语义或时间边界中的一项，事件与时间需要互相校验并纳入训练奖励和基准。摘要没有给出 AVTime 的具体性能指标、基线比较或数据集视频分布。

Abstract: Dense temporal audio-visual captioning is important for multimodal video understanding and generation. However, prior work is constrained by three main limitations: (1) the lack of a dataset combining high temporal resolution, longer videos, and bidirectional alignment between time and events; (2) the reliance of most existing RL methods on checklists; and (3) the lack of a benchmark that bidirectionally evaluates event semantic correctness and temporal accuracy. To address these challenges, we propose: (1) AVTime-50K, a high-quality 50K-scale dataset with temporally precise, semantically accurate captions and longer videos; (2) AVTime, a state-of-the-art model optimized with a novel Bidirectional Temporal-Semantic Alignment (BiTSA) reward that imposes explicit bidirectional constraints between time and events; and (3) AVTime-Bench, a specialized benchmark that introduces a novel bidirectional evaluation of temporal accuracy and semantic correctness. Our code, models, and datasets are available at https://huggingface.co/collections/Apryle/avtime.

[195] Risk Decomposition for Hallucination-Aware Video Sampling cs.CVPDF

Chenyan Zheng, Fan Wu, Ruibo Li, Jiaxing Li, Minghao Yang

TL;DR: 论文把物理视频幻觉分为低密度采样风险（模型不偏好的不连续生成）和高密度漂移风险（高似然却违反物理规律），据此提出使用外部物理引导的 HA-Sampler 与基于生成器似然评分的 HA-Scorer；在 Physics-IQ 和 VideoPhy 上超过所比较的采样及评分基线。

Details

Motivation: 视频模型能生成逼真画面，却仍可能违反物理规律，限制其作为世界模拟器的可靠性。既有方法用模型内部或外部信号改善物理合理性，但这些信号分别解决什么风险、又有哪些局限尚缺统一解释。

Result: 作者定义二维物理幻觉风险：低密度采样风险是模型分配低似然的生成，例如球突然瞬移；高密度漂移风险是模型高似然但物理不一致的生成，例如球在没有额外能量时越弹越高。分析指出，模型内部信号可把采样引向已学分布的高密度区域，处理前一风险，但似然高不代表物理正确，因此后一风险仍需外部引导。HA-Sampler 在潜变量精炼中加入外部物理指导生成物理合理视频；HA-Scorer 使用连续正规化流的生成器似然，以粗粒度 ODE 积分和有限差分散度估计为生成视频评分。Physics-IQ 和 VideoPhy 实验支持该分析，并显示两方法超过所比较的先进采样与评分基线。

Insight: 提高似然只能缓解偏离模型常见动态的采样错误，无法保证高概率生成符合物理规律；需要将模型内部概率信号与外部物理约束分别用于不同风险。摘要未给出定量提升或每种风险的单独评测结果。

Abstract: Despite achieving high visual fidelity, video models still suffer from physical hallucinations, limiting their reliability as world simulators. Existing methods leverage model-internal or external signals to improve physical plausibility, but their roles and limitations require a unified explanation. We propose a two-dimensional physical-hallucination risk framework that defines low-density sampling risk and high-density drifting risk. The first risk captures generations assigned low likelihood by the video model, for example, a ball suddenly teleporting despite the model favoring continuous motion, while the second risk captures physical inconsistencies in high-likelihood generations, for example, a ball bouncing progressively higher without additional energy. Our analysis explains how model-internal signals steer sampling toward higher-density regions of the learned distribution to address low-density sampling risk. However, higher model likelihood does not guarantee physical plausibility, motivating external guidance to address high-density drifting risk. Based on this framework, we develop Hallucination-Aware Sampler (HA-Sampler), which applies external physical guidance within latent refinement to generate physically plausible videos, and HA-Scorer, which derives a likelihood-based signal from the generator itself for evaluating generated videos using continuous normalizing flow (CNF), with coarse ODE integration and finite-difference divergence estimates. Experiments on Physics-IQ and VideoPhy benchmarks support our theoretical analysis and demonstrate that our methods outperform the compared sampling and scoring SOTA baselines.

[196] DCR-TSF: Dual-Level Compositional Reasoning with Triadic Spherical Fusion for Training-Free Zero-Shot Composed Image Retrieval cs.CVPDF

Gwang-Ho Na, Ho-Joong Kim, Seong-Whan Lee

TL;DR: DCR-TSF 是训练无关的零样本组合图像检索框架，先从参考图和修改文本生成细粒度与粗粒度目标描述，再把二者与原始图文组合查询锚点在单位球面融合；在 CIRCO、CIRR 和 Fashion-IQ 上达到训练无关方法中的先进结果。

Details

Motivation: 零样本组合图像检索要在保持参考图语义的同时，按文字修改检索目标图。既有免训练方法常依赖细粒度目标描述，容易过度关注参考图特有外观而出现外观偏置；多模态大模型生成描述还可能削弱原始图文查询对最终检索表示的直接影响。

Result: Dual-level Compositional Reasoning（DCR）依据参考图和修改文本生成细、粗两种粒度描述，补充互补的目标语义。Triadic Spherical Fusion（TSF）将参考图像特征与修改文本特征球面插值，形成组合查询锚点，再在单位超球面上用加权 Karcher 均值融合锚点、细描述与粗描述，兼顾具体视觉证据、高层语义和原始查询。CIRCO、CIRR 和 Fashion-IQ 实验显示，该方法在训练无关组合检索方法中达到先进性能。

Insight: 仅用生成描述可能遗失原始图文组合意图；保留直接查询锚点并融合不同粒度语义，有助于平衡外观细节与抽象修改。摘要未给出具体检索指标、基线差距或融合权重敏感性。

Abstract: Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image that reflects an intended textual modification while maintaining semantic consistency with a reference image. Existing training-free ZS-CIR methods typically rely on fine-grained generated target descriptions, which often overemphasize reference-specific visual details and lead to appearance-biased retrieval. Moreover, since these descriptions are generated through multimodal large language models (MLLMs), the direct influence of the original image-text query can be weakened in the final retrieval representation. In this paper, we propose DCR-TSF, a training-free ZS-CIR framework that combines Dual-level Compositional Reasoning (DCR) and Triadic Spherical Fusion (TSF). DCR generates fine- and coarse-grained descriptions from the reference image and modification text, capturing complementary target-side semantics across different granularities. To preserve the original image-text query information, TSF constructs a composed-query anchor by spherically interpolating the reference image feature and modification text feature, and fuses it with the fine- and coarse-grained representations using a weighted Karcher mean on the unit hypersphere. The resulting final query representation balances detailed visual evidence, high-level semantic abstraction, and direct preservation of the original composed query. Extensive experiments on CIRCO, CIRR, and Fashion-IQ demonstrate that our method achieves state-of-the-art performance among training-free methods.

[197] DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay Removal cs.CVPDF

Deshui Miao, Yuhang Ma, Shaoyong Jia, Wenzhao Gao, Yue Zhang

TL;DR: DeOverlay 用源视频锚定的流模型恢复字幕、标题和 logo 遮挡的视频画面，并以不含遮挡位置的跨帧聚合选择局部源线索；新建的 DVTE-Bench 含 1,631 对视频，方法在五项重建与感知指标上超过四个基线，Mask LPIPS 和 Mask VFID 分别降低 58.3% 和 70.3%。

Details

Motivation: 视频叠加层移除要去掉字幕、标题和 logo，同时保留周边场景。现有时序传播或源条件生成方法在有用源信息被遮挡时，可能残留叠加物或扭曲场景结构。

Result: DeOverlay 从源潜变量出发，以固定源引导重建干净视频；共享速度网络沿成对源到干净视频路径训练，并用区域加权流匹配支持对中间状态逐步纠正。可见度归一化时序聚合会排除被遮挡位置，只平均仍可见的跨帧观察。全局源上下文和编辑区域嵌入通过加性条件补充这些局部线索，同时保留单一视频 token 序列。作者建立 DVTE-Bench，包含 1,631 对视频、八类文字覆盖，带 OCR 筛选的干净目标、逐帧像素遮罩和统一评测协议。完整发布集实验中，DeOverlay 在全部五项重建和感知指标上超过四个基线；相较最强对照，Mask LPIPS 降低 58.3%，Mask VFID 降低 70.3%，定性示例还展示了图形 logo 和贴纸去除。

Insight: 恢复遮挡场景时应只聚合真正可见的跨帧源证据，并以源潜变量锚定生成，减少模型凭先验改写场景。结果覆盖八类文字覆盖的基准；摘要未说明复杂运动或源画面长期遮挡情况下的性能。

Abstract: Video overlay removal aims to recover scenes obscured by subtitles, titles, and logos while preserving surrounding content. Existing methods use temporal propagation or source-conditioned generation, but can leave residual overlays and distort scene structure when useful source information is obscured. We propose DeOverlay, a source-anchored flow model that reconstructs the clean video from the source latent with fixed source guidance. Specifically, we train a shared velocity network along paired source-to-clean paths using region-weighted flow matching, enabling successive cor- rections from intermediate states. To select local source cues, we develop visibility-normalized temporal aggregation that excludes masked positions and averages the remaining observations. Global source context and edit-region embeddings complement these cues through additive condi- tioning, retaining a single video-token sequence. To evaluate scene recovery across diverse overlay conditions, we introduce DVTE-Bench, comprising 1,631 paired videos across eight text-overlay categories, with OCR-screened clean targets, frame-aligned pixel masks, and a common evaluation protocol. Experiments on the complete release show that DeOverlay outperforms four baselines on all five reconstruction and perceptual metrics, reducing Mask LPIPS and Mask VFID by 58.3% and 70.3%, respectively, relative to the strongest evaluated baseline. Additional qualitative examples demonstrate the removal of graphic logos and stickers. We will make the complete benchmark, evaluation code, and test outputs availabl. Project page:https://yahooo-m.github.io/DVTE-Bench/

[198] Context-Aware Topological Distillation Graph Network for Asymmetric EEG-Eye Movement Emotion Recognition cs.CV | cs.HCPDF

Peijun Shi, Xiujin Zhu, Zhenguo Lao, Chee-Onn Chow, Wei Ru Wong

TL;DR: CTD-GNet 将 EEG 表示为电极图节点，并用眼动和前序窗口调制图结构、节点及读出；双向蒸馏把融合教师知识传给单窗口 EEG 学生，在 SEED 和 SEED-IV 跨被试评估中取得最高平均准确率，并对缺失或受噪眼动保持鲁棒。

Details

Motivation: EEG 与眼动联合情绪识别可用于心理健康监测和自适应界面，但常见融合模型把两种信号当成并列特征、逐窗口分类，眼动缺失时性能下降。论文针对眼动与脑电信息贡献不对称及跨窗口上下文不足提出方法。

Result: CTD-GNet 将 EEG 特征作为电极图节点，由眼动信号和前序窗口调节节点、读出及带电极专属提示的边。多模态教师与单窗口 EEG 学生之间对图结构、节点关系和预测进行双向蒸馏，既提升教师表现，也让两图结构对齐程度提高 35 倍。在 SEED 和 SEED-IV 留一被试测试中，平均准确率分别为 84.91% 和 76.17%；相较 17 个基线取得最高均值、最低多模态跨被试波动和最高最差被试准确率。SEED-IV 各种眼动缺失率下均最准确，无眼动的 SEED 上仍达 66.1%；眼动噪声达到一个标准差时，比单窗口融合模型高 9.3 和 11.9 分。对训练中未见电影片段，SEED 上准确率为 77.1%。

Insight: 将跨窗口上下文和眼动可靠性纳入图结构，并蒸馏给只依赖单窗口 EEG 的学生，可同时提高融合效果与缺模态鲁棒性。结果来自 SEED 系列和指定噪声设置，摘要未给出实际临床监测评估。

Abstract: EEG-eye movement emotion recognition can support mental-health monitoring and adaptive interfaces, but most fusion models treat the two signals as parallel features, classify each window alone and degrade when eye data are lost. We propose the Context-Aware Topological Distillation Graph Network (CTD-GNet), in which EEG features form the nodes of an electrode graph, and eye movements and preceding windows modulate its edges, nodes and readout, the edges through electrode-specific prompts. Bidirectional distillation of the graph, node relations and predictions couples this multimodal teacher with a single-window EEG student, improving the teacher and aligning the two graphs 35-fold. In leave-one-subject-out tests on SEED and SEED-IV against 17 baselines, CTD-GNet has the highest mean accuracy, 84.91% and 76.17%, significantly above 9 and 6 of 13 eye-based baselines after Holm correction, the smallest cross-subject spread among multimodal models and the highest worst-subject accuracy. It is the most accurate model at every rate of missing eye data on SEED-IV, keeps 66.1% on SEED without eye data, more than any EEG-only model, and stays 9.3 and 11.9 points above all single-window fusion models under eye-movement noise of one standard deviation. On film clips unseen in training, it leads on SEED with 77.1%.

[199] RGB–Event Object Detection Across Operating Frequencies via Reliability-Supervised Routing and Asymmetric-Resolution Fusion cs.CVPDF

Peijun Shi, Jiannan Zhao, Xiujin Zhu, Zhenguo Lao, Chee-Onn Chow

TL;DR: ERA-Det 为 RGB—事件相机目标检测设计逐位置可靠性路由和不对称分辨率融合，在事件检测频率升高、事件变稀疏时动态调整两种传感器贡献；DSEC-Det 上 20 至 180 Hz 的精度损失由 6.3 降至 3.1 mAP，整体达到 63.2 mAP。

Details

Motivation: 自动驾驶需要比普通帧相机更快地检测目标，事件相机可在两帧之间以微秒级记录亮度变化。检测频率越高，每次决策可用的事件越少，事件相对帧图像的可靠性还会随场景位置和时间变化；现有方法往往整样本加权两种传感器，并把帧图降到事件分辨率。

Result: 可靠性监督路由为每个位置学习融合权重，并在训练期用两个单模态分类器的损失生成监督信号，将原先样本级的损失引导权重细化到局部位置。不对称分辨率融合保留原生分辨率帧图，将各层特征池化到事件网格，并复用预训练权重。在 DSEC-Det 上，从 20 Hz 提升至 180 Hz 时，路由将精度损失从 6.3 降至 3.1 mAP，所有随机种子均降低；ERA-Det 达 63.2 mAP，比 FlexEvent 高 3.7，并在每个频率领先，180 Hz 时甚至超过 FlexEvent 的 20 Hz。两个 DAVIS346 数据集还显示，夜间和低照时路由降低帧图权重，事件消失时则提高帧图权重。单 GPU 5,000 次迭代即可升级已发布 FlexEvent 权重。

Insight: 传感器可靠性需要按空间位置、时间和操作频率自适应评估，且不应因事件分辨率低而丢弃高分辨率帧信息。摘要未给出端到端检测延迟或不同天气条件的结果。

Abstract: Autonomous vehicles must detect objects faster than a frame camera delivers images. Event cameras record brightness changes with microsecond resolution, so RGB–event detectors can decide between frames, at an operating frequency, the rate of decisions, far above the frame rate. As this rate rises, each decision receives a sparser event window, and the reliability of the events relative to the frame varies over the scene and over time. Existing detectors weight the sensors per sample and encode the frame at the reduced event resolution. Reliability-supervised routing assigns a fusion weight to every location and supervises it with targets from the losses of two single-modality classifiers used only in training, moving loss-based supervision of fusion weights from whole samples to individual locations. Asymmetric-resolution fusion lets the frame carry the decision where the events fail: it encodes the frame at native resolution and pools every feature level onto the event grid, reusing all pretrained weights. On the DSEC-Det dataset, routing halves the accuracy loss from 20 to 180 Hz, from 6.3 to 3.1 mAP, and lowers it in every seed. The resulting detector, ERA-Det, sets a new state of the art at 63.2 mAP, 3.7 mAP above FlexEvent, leads at every operating frequency, and at 180 Hz is more accurate than FlexEvent at 20 Hz. On two DAVIS346 datasets, the router lowers the frame weight at night and in low light and raises it when the events vanish. ERA-Det upgrades released FlexEvent weights in 5,000 iterations on one GPU.

Tao Sun, Huanting Guo, Baao Xie, Zhou Fang, Qiuyu Chen

TL;DR: GeoNaviDiff 将 StyleSpace 的因子化编辑控制迁移为扩散模型 h 空间中的局部黎曼轨迹，用匹配的 StyleGAN 重建—编辑对构造监督，再以共享因子条件导航器和测地线生成保持真实图像保真度与编辑局部性。

Details

Motivation: StyleSpace 提供可解释的单因素语义控制，但依赖 GAN 反演；扩散模型有更强真实图像生成能力，却缺少显式因子化控制基底。论文希望把前者的解耦控制能力迁移到扩散特征，而不需要针对每个因素训练专用模型或在推理时依赖编辑参考。

Result: GeoNaviDiff 从匹配的 StyleGAN 重建—编辑图对提取球面 Log 映射切向量，再平行移动到真实图像的扩散状态，构造与重建解耦的监督。共享且受因素条件化的 Navigator 在推理时预测该方向；范数保持测地线实现与源图锚定 rollout 稳定反向过程。多数据集实验显示，相较所比较方法，GeoNaviDiff 在图像保真度、因素选择性、编辑局部性和轨迹平滑性上均更好。

Insight: 将 GAN 可解释编辑方向转成扩散空间的局部几何路径，可尝试兼得因子控制与扩散生成质量；推理时不需编辑参考或逐因素模型。摘要未给出具体数据集、指标数值或计算开销。

Abstract: Disentangled real-image manipulation requires factor-specific semantic control while preserving image fidelity and non-target content. StyleSpace $S$ provides interpretable controls but is limited by GAN inversion, whereas diffusion models offer stronger image realization without an explicit factorized control basis. We introduce GeoNaviDiff, an unsupervised cross-space framework that transfers disentangled StyleSpace interventions to local Riemannian trajectories in diffusion $h$-space. GeoNaviDiff constructs reconstruction-decoupled supervision by extracting a spherical Log-map tangent from a matched StyleGAN reconstruction-edit pair and parallel-transporting it to the real-image diffusion state. A shared, factor-conditioned Navigator predicts this tangent at inference, while norm-preserving geodesic realization and source-anchored rollout stabilize the reverse process without edited references or factor-specific models. Experiments across multiple datasets show that GeoNaviDiff outperforms the compared methods in fidelity, factor selectivity, edit locality, and trajectory smoothness.

[201] TriTD: Tri-Partite Trajectory-Distribution Distillation for Real-Time Autoregressive Video Generation cs.CVPDF

Jia Li, Xurui Peng, Haowei Zhu, Tianyu Zhao, Jiexi Wang

TL;DR: TriTD 将自回归视频扩散误差拆为基础模型误差、训练—推理分布差异和上游传播，并用下一节点轨迹蒸馏、分块渐进加噪和正交并行分布匹配联合优化；单张 H100 达 26.8 FPS，同时改善视觉质量。

Details

Motivation: 自回归视频扩散适合高效视频生成，但质量与推理效率之间仍有瓶颈。论文将 rollout 误差归因于模型本身、训练与推理噪声分布不匹配，以及前序误差传播，并推导逐步与累计误差上界以共同优化质量和速度。

Result: TriTD 含三部分：下一节点轨迹蒸馏缩短 ODE 初始化监督跨度并降低回归误差下限；单调非递减的分块加噪调度减小预期训练—推理噪声差异，同时减少冗余 KV-cache 更新；正交并行 DMD 保留原分布匹配监督，并在正交子空间增加回归约束以抑制偏轴漂移、改善训练稳定性和生成质量。实验显示，该方法视觉质量持续超过现有基线，并在单张 H100 上达到实时吞吐 26.8 FPS。

Insight: 训练轨迹、分块噪声分布和分布匹配约束需共同设计，才能同时降低自回归误差累积和推理成本；只优化单步去噪可能忽略误差在后续块中的传播。摘要未提供具体视频质量指标、分辨率或不同硬件速度对照。

Abstract: Autoregressive (AR) video diffusion has emerged as a leading paradigm for efficient video generation, yet its practical deployment remains bottlenecked by the trade-off between generation quality and inference efficiency. We attribute this bottleneck to three error sources in AR rollouts: the base model error, the train-inference distribution mismatch, and the upstream error propagation. Building on this decomposition, we derive closed-form upper bounds on the per-step and cumulative errors, providing a principled basis for jointly optimizing generation quality and inference efficiency. Guided by this analysis, we propose TriTD, a Tri-Partite Trajectory-Distribution co-optimized distillation framework with three complementary modules. First, for ODE initialization, we adopt a Next-Node Trajectory Distillation (NTD) objective, which shortens the supervision horizon and admits a lower regression-error floor. Second, we introduce Chunk-Graded Noising (CGN), a monotonically non-decreasing per-chunk noising schedule that helps narrow the expected train-inference noise mismatch and removes redundant KV-cache updates, yielding substantial speedups at no quality cost. Third, we propose Orthogonal-Parallel DMD (OP-DMD), a distribution-matching loss that retains the original DMD supervision while introducing an additional regression constraint in the orthogonal subspace to suppress off-axis drift, improving both training stability and final generation quality. Extensive experiments show that TriTD consistently surpasses existing baselines in visual quality while reaching a real-time throughput of 26.8 FPS on a single H100 GPU, delivering a balanced quality-efficiency solution for interactive streaming video generation.

[202] Wan-Alpha++: Animating a Single RGBA Asset into Controllable Transparent Video Effects cs.CVPDF

Haotian Dong, Xin Wang, Wenjing Wang, Chen Li, Jing Lyu

TL;DR: Wan-Alpha++ 在共享 RGBA 潜空间中扩展文生和图生透明视频生成，并通过透明感知双向扩散、透明引导均值采样和时间相位重参数化解决 RGB/alpha 耦合、透明边界帧导致的外观条件失效问题。

Details

Motivation: 电影、游戏和数字内容生产需要可合成的 RGBA 视频，但同时生成高质量 RGB 外观和时间连贯透明度很难。已有方法常把 RGB 与 alpha 分布耦合、难以精确控制目标资产；图像条件生成还可能因首尾完全透明而出现内容显隐异常。

Result: Wan-Alpha++ 将统一 RGBA 潜空间接入文生与图生视频扩散骨干，并从原文本条件版本扩展至包含内部参考帧的图像条件生成。方法学习可移动的 RGB—alpha 分布，同时保持预训练 RGB 先验：透明感知双向扩散目标形成可区分 RGBA 潜表示，透明引导均值采样通过移动噪声分布改善透明度生成并支持空间控制。针对图像条件效果视频，时间相位重参数化将零—正—零 alpha 循环变为正—零—正，使可见内容出现在时间两端，以适配有限 RGBA 数据下的外观条件。

Insight: 透明度的时序分布需要和 RGB 外观分开控制；重排 alpha 的时间相位可解决透明边界帧与图像条件不兼容的问题。摘要未报告具体生成指标、数据规模或与先前 Wan-Alpha 版本的定量比较。

Abstract: RGBA video generation is essential for creating compositable visual assets in filmmaking, game development, and digital content production. However, generating high-quality RGB appearance alongside temporally coherent transparency remains a significant challenge. Existing text-to-RGBA video methods often entangle RGB and alpha distributions and offer limited control over the precise appearance of a target asset. In addition, incorporating image conditioning typically suffers from the emergence-disappearance effects, which fail existing first-frame or first-and-last-frame conditions due to their fully transparent boundary frames. To address these issues, we present Wan-Alpha++, a unified framework that brings controllable RGBA video generation to both text- and image-to-video diffusion backbones through a shared RGBA latent space, extending our conference version Wan-Alpha from the text-conditioned setting to image-conditioned generation with an interior reference frame. Our method first learns shiftable RGB-alpha distributions while preserving the pretrained RGB prior. Specifically, we introduce a transparency-aware bidirectional diffusion objective to produce distinguishable RGBA latent representations, and a transparency-guided mean sampler that shifts the noise distribution to improve transparency generation and enable spatial controllability. Building upon this RGBA generation prior, we introduce temporal phase reparameterization for image-conditioned effect generation. By cyclically transforming a zero-positive-zero alpha pattern into a positive-zero-positive pattern, it relocates visible content to both temporal boundaries, facilitating effective appearance conditioning with limited RGBA data.

[203] GeoEdit: Adversarial De-Evidencing for Silent Geoprivacy Protection via Image Editing cs.CV | cs.AIPDF

Yu Yan, Shengjia Cheng, Fei Ma, Zhe Wang, Peng Chen

TL;DR: GeoEdit 对图像中高风险地理证据锚点做最小编辑，打断多模态推理模型的定位证据链，同时尽量保持整体场景语义；在 LifePhotoBench 上防护六种主流模型，图像保真度比视觉扰动基线高 0.59 以上。

Details

Motivation: 多模态推理模型能组合细微视觉线索推断照片拍摄地点，使非专业攻击者也可能进行地理位置曝光。现有防御难以兼顾照片可分享性与有效保护，也常不能充分打断嵌在场景语义中的隐式地理证据链。

Result: GeoEdit 先结合结构化知识和对抗发现的地理线索识别高风险证据锚点，再对这些位置进行最小化编辑，以干扰空间推理链而保留整体场景语义。为处理较大的编辑动作空间，方法使用可学习策略迭代优化编辑决策，并把成功轨迹蒸馏为可迁移的自然语言经验，以保持编辑一致性和效率。在 LifePhotoBench 上，GeoEdit 针对六种主流多模态推理模型提供较强保护，图像保真度比视觉扰动基线高 0.59 以上。

Insight: 静默地编辑少数关键地理证据，比全图强扰动更可能保留照片可分享性；通过迭代策略学习还能复用成功编辑经验。摘要未给出定位成功率下降幅度、保真度指标定义或对人类观察者的影响。

Abstract: Recent efforts have demonstrated the strong capabilities of Multimodal Large Reasoning Models (MLRMs) in geographic reasoning. By jointly analyzing subtle visual cues, these models can infer the locations of photos, thus enabling doxing even by non-expert adversaries. Existing defense methods struggle to balance the shareability of photos with effective protection, and often fail to sufficiently disrupt the implicit geo-evidence chains embedded in scene semantics. To this end, we propose GeoEdit, an adversarial de-evidencing framework for silent geoprivacy protection via image editing. Specifically, we apply minimal edits to high-risk evidence anchors in the image, thereby disrupting the underlying spatial reasoning chains while preserving overall scene semantics. We identify these anchors by combining structured knowledge with adversarially discovered geographic cues. To further navigate the large editing action space, we adopt a learnable policy to iteratively refine editing decisions, distilling successful trajectories into transferable natural-language insights for consistent and efficient editing. Evaluated on LifePhotoBench, GeoEdit achieves strong protection against 6 mainstream MLRMs, outperforming visual perturbation baselines by over 0.59 in image fidelity.

Fuyao Zhang, Wenwen He, Jiaming Zhang, Yurong Hao, Xuan Wang

TL;DR: FocusGUI 根据当前决策动态筛选截图、音频、视频、历史和设备状态中的关键证据，并将其压缩注入冻结解码器；四个全模态骨干的平均 Type Match 提升 33.2，Exact Match 达冻结基线的 3 倍，Type Match 比最强既有增强方法高 30.1。

Details

Motivation: GUI 智能体面对流式截图、音频、视频、交互历史和设备状态，但每一步都均匀提供完整上下文会忽视相关性随动作变化且信号稀疏的特点。前一步动作又会进入后续上下文，使分心错误累积为不可逆的轨迹漂移。

Result: FocusGUI 将需求形式化为逐步自适应、足以支持动作且不干扰原模型的 Focused Perception，并实现为轻量适配器：根据当前决策参照动态优先排序多模态通道，把相关证据压缩到紧凑潜表示，再轻量注入冻结解码器嵌入，引导注意力而不改动骨干参数空间。训练只需动作监督，不需模态相关性标签或骨干微调。四种全模态骨干上的平均结果显示，较冻结骨干 Type Match 提升 33.2，Exact Match 达其 3 倍；Type Match 较最强现有增强方法高 30.1。

Insight: 多模态 GUI 智能体应针对每个动作选择当前有效证据，而不是把所有通道长期并列塞入上下文；紧凑而非侵入式引导可避免微调整个骨干。摘要未报告不同设备、任务的分项成绩或延迟变化。

Abstract: Graphical User Interface (GUI) agents increasingly act over streaming omni-modal observations spanning screenshots, audio, video, interaction history, and device states. However, exposing this complete context uniformly at every step ignores that relevance is transient and sparse: the signal determining one action can become clutter at the next. Because executed actions feed back into the context, these errors compound into Sequential Distraction, driving the agent into irreversible trajectory drift. We formulate the underlying requirement as Focused Perception: a representation that is step-adaptive, action-sufficient, and non-intrusive. We realize this in FocusGUI, a lightweight adapter that dynamically prioritizes multimodal channels against a step-specific decision reference. Instead of uniformly processing all inputs, it compresses the most relevant evidence into a compact latent and subtly injects it into the frozen decoder’s embeddings, guiding the model’s attention without altering its native parameter space. FocusGUI trains from action supervision alone, requiring neither modality-relevance labels nor backbone fine-tuning. Across four omni-modal backbones, it raises overall Type Match by +33.2 and Exact Match by 3.0 times over the frozen backbone on average, and Type Match by +30.1 over the strongest existing enhancement method.

Tom Wehrbein, Bodo Rosenhahn

TL;DR: ExMR 在全身时序估计后额外回看局部手足证据，分别细化脚踝、肘、腕和手指关节旋转；在 ARCTIC、MOYO、RICH、MOOF 上超过先进方法，MOYO 上手部顶点误差和脚踝角误差分别相对最强竞争者降低 19% 和 18%。

Details

Motivation: 动画和动作分析需要精细的手足关节动作，但单目视频难以恢复。已有表达式视频方法从共享全身表示推断手部动作，作者此前发现这会浪费脚部局部关键点证据，局部足部细化能改善重建；本文将同一思路扩展到手部。

Result: ExMR 首先用时序估计器预测全身动作，再由 FootMR 与新增 HandMR 重新检查局部证据并预测残差关节旋转。FootMR 根据足部关键点细化踝关节旋转；HandMR 针对二维手指点缺失或不可靠的问题，结合预训练手部重建模型的图像特征与初始臂手旋转，细化肘、腕和手指。ARCTIC、MOYO、RICH 和 MOOF 实验显示，ExMR 的手足重建优于先进方法；相较最强竞争者，MOYO 上手部逐顶点误差降低 19%，脚踝关节角误差降低 18%。

Insight: 全身估计提供粗略动作后，再针对手足局部图像证据进行残差修正，可以弥补共享整体表示忽略的精细关节线索。摘要未给出其他数据集的误差数值或推理成本。

Abstract: Detailed hand and foot articulation is important for animation and motion analysis, yet remains challenging to recover from monocular video. Recent expressive video methods infer hand articulation from shared whole-body representations. Our conference work suggests that this underexploits local evidence: the Foot Motion Refinement (FootMR) module refines ankle rotations from foot keypoints, improving reconstruction even when the initial estimator receives the same keypoints. We extend this principle to hands with the Hand Motion Refinement (HandMR) module. Because 2D finger keypoints are often missing or unreliable, HandMR combines image features of a pretrained hand reconstruction model with initial arm and hand rotations to refine elbows, wrists, and fingers. We integrate FootMR and HandMR into ExMR (Extremity Motion Refinement), a system for reconstructing 3D body, hand, and foot motion from monocular video. An initial temporal estimator predicts whole-body motion, after which both refiners revisit local evidence to predict residual joint rotations. Experiments on ARCTIC, MOYO, RICH, and MOOF show that ExMR outperforms state-of-the-art methods in hand and foot reconstruction, reducing hand per-vertex error and ankle joint angle error on MOYO by 19% and 18%, respectively, relative to the strongest competitor. Code will be made available for research purposes. Project page: https://twehrbein.github.io/exmr-website/

[206] RECAP: Reciprocal Sparse Attention and Hierarchical Caching for Efficient Video Diffusion cs.CV | cs.LGPDF

Tiange Xu, Kai Liu, Linghe Kong, Yulun Zhang

TL;DR: RECAP 将稀疏注意力误差与跨去噪步缓存误差闭环联动：用缓存的稠密细化修正当前稀疏估计，再以新鲜稀疏统计跟踪输入变化并更新历史预测源；HunyuanVideo 生成 720p、5 秒视频时达到 3.05 倍加速，质量和速度均超过直接组合基线。

Details

Motivation: 视频扩散 Transformer 计算昂贵，稀疏注意力和时序缓存分别减少单步内部计算和跨去噪步重复计算，二者看似互补。但稀疏注意力预测本身会成为后续缓存复用的来源，使误差沿采样轨迹耦合累积。

Result: RECAP 是免训练方法，用在线误差反馈协调稀疏注意力与分层缓存。缓存的稠密细化负责修正当前稀疏估计中省略的交互；新计算的稀疏统计跟踪输入变化并调节修正信息随时间老化，再用更新后的预测刷新整步复用来源，让预测历史能在稠密刷新之间推进。在多个视频扩散模型上，作者与稀疏注意力、缓存方法及两者直接组合进行比较。HunyuanVideo 生成 720p、5 秒视频时，RECAP 加速 3.05 倍，且保真度和推理速度均优于所有测试的直接组合。

Insight: 稀疏化产生的预测误差与缓存重用不能独立优化；稠密校正与新鲜稀疏统计形成反馈，能够让缓存继续复用而不长期累积错误。摘要未报告其他模型的具体速度和质量指标。

Abstract: Diffusion transformers generate high-quality videos but remain computationally expensive. Sparse attention and temporal caching reduce computation within and across denoising steps, respectively, offering a natural composition. However, predictions produced with sparse attention become sources for temporal reuse, coupling their errors along the sampling trajectory. We observe that improving the accuracy of these sources recovers fidelity without changing the reuse schedule, and that sparsification errors remain directionally aligned across nearby steps, suggesting that correction information can be shared rather than recomputed. These observations motivate RECAP, a training-free method that couples sparse attention with hierarchical caching through reciprocal correction, adaptively coordinated by online error feedback. Cached dense refinement corrects omitted interactions in current sparse estimates, while fresh sparse statistics track input changes and regulate the aging correction. The resulting predictions then renew the sources for whole-step prediction reuse, allowing prediction history to advance between dense refreshes. We evaluate RECAP on multiple video diffusion models against sparse-attention methods, caching methods, and their direct compositions. On HunyuanVideo at 720p 5-second generation, RECAP achieves a 3.05x speedup and outperforms all tested direct compositions in both fidelity and inference speed.

[207] XGLASS: X-Ray, Gaze and Language Assisted Slice Sequencing for CT Volumes cs.CVPDF

Michael Glöckler, Timo Ropinski

TL;DR: XGLASS 将语言对齐的放射科医生 X 光注视先验与 CT 冠状投影对齐，为病种相关轴位切片排序；最佳病种 top-10 命中率达 75.9%，仅需查看 240 层中的 4.17%，且用 30 层候选集合使下游分类模型调用减少 87.5%、F1 提高 11.3 个百分点。

Details

Motivation: 二维医疗图像 VLM 表现较好，但直接处理数百层 CT/MRI 的三维模型受限于 3D 训练数据少和基础架构以 2D 为主。要利用强大的二维模型分析体数据，需先挑出诊断相关切片。

Result: XGLASS 将放射科医生基于语言对齐的 X 光注视图转成疾病条件重要性图，并与冠状 CT 投影在共享二维表示中空间对齐，融合注视先验和 CT 内容后预测病种专属轴位切片顺序。最佳病种 top-10 命中率为 75.9%，即只查看 240 层中的 10 层（4.17%）时，超过 75% 病例能找到诊断相关层。用于逐层异常分类时，XGLASS 选择的 30 层候选集比全层推理少调用 87.5% 模型，同时 F1 高 11.3 个百分点；相较最强二维选择器高 4.2 点。其性能仍接近原生 3D 选择器，而每项发现最多快 130.8 倍、GPU 内存少 10.9 倍。

Insight: 二维医生注视先验可辅助在三维 CT 中进行病种相关切片排序，在显著减少检查层数和推理资源的同时改善下游分类。摘要未列出各疾病的命中率分布，也未说明医生注视数据规模。

Abstract: Vision-language models (VLMs) achieve remarkable performance on 2D medical images, such as X-rays or pathology images, while 3D VLMs that directly process CT/MRI volumes lag behind due to limited 3D training data and the predominance of 2D foundation architectures. To leverage powerful 2D VLMs for analyzing medical volumes, which contain hundreds of slices, it is therefore necessary to focus only on those few slices, which are diagnostically relevant. To enable such a prioritization, we propose XGLASS, a method that transfers medical expertise from 2D X-rays to 3D CT for task-aware slice prioritization. XGLASS uses disease-conditioned importance maps derived from language-aligned radiologist gaze on X-rays and spatially aligns them with coronal CT projections, bringing gaze priors and CT content into a shared 2D representation. The fused gaze-prior/CT projection is then used to predict a disease-specific ordering of axial slices, prioritizing diagnostically relevant slices. XGLASS significantly improves budgeted retrieval over 2D and 3D-native baselines, achieving a top-10 hit rate up to 75.9% for the best-performing disease. In practice, this means that after inspecting only the top 10 slices (4.17% of a 240-slice scan), XGLASS identifies a diagnostically relevant slice in more than 75% of cases. For downstream slice-wise finding classification, an XGLASS-selected 30-slice candidate set reduces model calls by 87.5% while improving F1 by 11.3 percentage points over all-slice inference, 4.2 points over the strongest 2D selector, and remains competitive with 3D-native selectors while requiring up to 130.8 times less time and 10.9 times less GPU memory per finding, supporting its use as a lightweight preselector for volumetric CT analysis. The code and pretrained checkpoint are available on https://github.com/michaelgloeckler/XGLASS.

[208] Dynamics-Aware Sparse Attention for Efficient Autoregressive Video Diffusion cs.CVPDF

Yuanyu He, Zhuokun Chen, Yefei He, Zhiwei Tang, Jiasheng Tang

TL;DR: DASA 按视频运动稀疏性将内容划分为动态和静态区域，只保留生成片段中的动态历史特征压缩 KV 缓存，并按局部运动分配计算；VBench 上自回归视频生成加速 2.16 倍，同时保持较高视觉保真度。

Details

Motivation: 自回归扩散 Transformer 可生成长视频，但密集计算带来高算力和内存开销，且对所有时空区域均匀分配资源，未利用多数帧间内容静态、运动只发生在稀疏区域的冗余结构。

Result: Dynamics-Aware Sparse Attention（DASA）受视频编码器启发，将视频内容显式拆成动态与静态区域。为降低内存，方法选择性压缩历史 KV 缓存，仅保留生成片段中的动态特征；为降低计算，按局部运动程度分配区域计算资源；另用轻量蒸馏训练提高性能。在 VBench 上，DASA 将自回归视频生成加速 2.16 倍，同时保持较高视觉保真度。

Insight: 视频的时间冗余既可用于减少历史缓存，也可用于把计算集中到真实运动区域；单纯全局稀疏化可能无法利用这种局部动态结构。摘要未报告详细 VBench 分数、分辨率或内存节省比例。

Abstract: Autoregressive Diffusion Transformers have emerged as a powerful paradigm for long video generation. However, they are often constrained by prohibitive computational overhead and substantial memory footprints. This inefficiency arises from dense computation mechanisms that allocate uniform processing resources across all spatiotemporal regions, failing to leverage the inherent temporal redundancy of video where motion is typically confined to sparse areas. In this paper, we propose Dynamics-Aware Sparse Attention (DASA), a unified framework that shifts from dense computation to sparse, motion-driven generation to eliminate both storage and computational redundancies. Inspired by video codec principles, our approach explicitly decouples video content into dynamic and static regions. To mitigate memory bottlenecks, we selectively compress the historical KV cache by retaining only dynamic features within the generated chunk. Furthermore, to reduce computational overhead, we employ a region-level computation allocation strategy based on localized dynamics. Additionally, a dynamics-guided distillation method is introduced to enhance performance via a lightweight distillation training process. Experiments on VBench demonstrate that our approach accelerates autoregressive video generation by 2.16x, while preserving high visual fidelity.

Yiheng Dong, Yi Lin, Shilong Huang, Xiyan Yang, Xin Yang

TL;DR: TIM 将纵向放射学报告生成分成时间与病理表征解耦、以及前后时点报告互相校正两阶段；在 Longitudinal-MIMIC 上，语言和临床指标均超过既有单时点及纵向基线。

Details

Motivation: 自动放射报告生成通常只分析单一时间点，不能表达疾病在连续检查间的变化。纵向方法虽加入历史影像，却常将不同时间图像混入同一表示空间，导致语义模糊和时间推理不一致。

Result: Temporal Decoupling with Iterative Mutual-Refinement Model（TIM）包含两阶段。第一阶段进行时间解耦表示学习，将疾病特征与时间进展模式分离，并分别生成前次和当前检查的报告。第二阶段的互报告精炼机制检测旧报告中的诊断不一致，再通过对错误敏感的反馈迭代修正前次和当前报告。Longitudinal-MIMIC 实验显示，TIM 在语言和临床指标上均超过已有单图像与纵向方法，达到论文报告的新最佳结果。

Insight: 纵向报告需要区分“是什么病变”和“病变如何变化”，并允许不同检查时点的报告互相校正，而非简单拼接影像特征。摘要未给出各指标的具体数值、数据划分或迭代次数。

Abstract: Automatic radiology report generation (RRG) aims to translate medical images into diagnostic text, reducing radiologists’ workload and standardizing clinical documentation. Nonetheless, existing approaches mainly focus on single-timepoint analysis and fail to capture temporal disease evolution across longitudinal examinations. While recent longitudinal RRG (LRRG) approaches incorporate historical data, they often combine images from different time points within a single representation space, leading to blurred semantics and inconsistent temporal reasoning. In this work, we propose a Temporal Decoupling with Iterative Mutual-Refinement Model (TIM), a two-stage framework that explicitly decouples spatial pathology from temporal progression and iteratively refines reports through mutual feedback. Stage I performs temporal-decoupled representation learning, separating temporal evolution patterns from disease-specific features and generating radiology reports for both prior and current studies. Stage II introduces a mutual report refinement mechanism that identifies diagnostic inconsistencies within prior reports and iteratively rectifies both prior and current reports through error-sensitive feedback. Experiments on the Longitudinal-MIMIC dataset demonstrate that TIM surpasses existing single-image and longitudinal baselines, achieving new state-of-the-art performance across both language and clinical metrics. Code is available at https://github.com/yihengd/TIM.

[210] From Cause to Effect: Generating Intermediate Event Chains in Videos cs.CVPDF

Ridouane Ghermi, Vicky Kalogeiton, Ivan Laptev

TL;DR: 因果链生成任务要求视频模型补出原因与结果之间的中间事件，并以语义匹配和因果事实完整性双指标评估；作者发现先进 VLM 常幻觉或遗漏关键事件，再用结合格式、语义和因果问答奖励的 GRPO 微调显著改善表现。

Details

Motivation: 视频语言模型擅长识别画面，却仍难以理解叙事因果。多项选择问答可被捷径利用，开放式问答又缺少稳健自动评估，因此需要既强制模型生成因果推理、又能可靠评分的任务。

Result: 论文提出 Causal Chain Generation（CCG），要求模型生成连接视频原因与结果的中间事件序列，并配套两种指标：SODA-Embedding 衡量生成链与参考链的语义匹配，Causal-QA 检查链中是否包含重建叙事所需的关键因果事实。评测数据包括人工标注 20 部短片的 SF20K-CC，以及由音频描述构成的电影片段集 CMD-CC。分析显示先进 VLM 经常误报或遗漏关键事件。作者用 GRPO 对 VLM 微调，复合奖励结合格式、SODA-Embedding 和 Causal-QA 准确率，结果较强基线有显著改善。

Insight: 要求生成中间因果事件链比只回答结论更直接检验模型是否连接了叙事过程；语义相似与因果事实完整需要分别评估。摘要未报告提升数值，也未说明两个指标与人类因果判断的相关性。

Abstract: Video-Language Models (VLMs) excel at visual recognition but struggle with causal reasoning. Current evaluations are flawed: multiple-choice QA allows shortcuts, and open-ended QA lacks robust automatic evaluation. We introduce Causal Chain Generation (CCG), a novel task where models have to generate the sequence of intermediate events connecting a cause and an effect. This structured setup demands genuine reasoning and enables reliable automatic scoring. We accompany this task with two complementary metrics: SODA-Embedding, which measures the semantic matching of a generated chain against a reference, and Causal-QA, which probes the chain’s causal coherence by asking whether it contains the specific causal facts needed to reconstruct the narrative. We provide two test sets: SF20K-CC, a manually annotated diagnostic on 20 short films, and CMD-CC, a movie clip benchmark derived from audio descriptions. Our analysis reveals that state-of-the-art VLMs struggle with narrative causality understanding, frequently hallucinating or omitting key events. To address this, we fine-tune VLMs with GRPO using a composite reward that combines a format reward, SODA-Embedding, and Causal-QA accuracy, demonstrating substantial improvements over strong baselines.

[211] HGASurv: Hierarchical Granularity-Aware Integration for Multimodal Survival Prediction cs.CVPDF

Peike Yu, Wuchao Liu, Jihong Guan, Shuigeng Zhou

TL;DR: HGASurv 将全切片病理表示做成 patch、邻域和区域三级层次，再让基因组签名跨模态查询这些尺度，并用临床相似图传播融合图像—基因组表示；在三个 TCGA 队列上取得最强总体 C-index。

Details

Motivation: 多模态生存预测要联合分析病理全切片图像、基因组和临床变量，但各模态的空间组织与特征维度差异很大，现有方法常未充分处理这些粒度差异。

Result: HGASurv 的层次 WSI 编码器按原始坐标将 patch 特征映射到空间网格，通过并行卷积构造 patch、局部邻域和区域表征；基因组签名嵌入再通过跨模态共注意查询不同尺度表示，形成签名条件形态特征。临床引导图传播模块根据临床变量定义患者相似边，在图拓扑上用图注意传播融合后的图像—基因组节点嵌入，而不是简单拼接临床特征。三个 TCGA 队列上的总体 C-index 最强；消融支持层次 WSI 编码和临床引导传播的互补作用。

Insight: 融合机制应匹配每种模态自身的组织粒度：病理图像需保留空间层次，临床信息可通过患者关系拓扑注入。摘要未给出具体 C-index、各癌种队列名称或统计显著性。

Abstract: Objective: Multimodal survival prediction jointly analyzes histopathological whole-slide images (WSIs), genomic profiles, and clinical variables to stratify patients by risk. However, existing approaches often inadequately account for granularity differences across modalities, particularly in spatial organization and feature dimensionality. To address these challenges, we propose HGASurv, which introduces a hierarchical WSI encoding module and a clinical-guided graph propagation module for granularity-aware multimodal integration. Methods: The hierarchical WSI encoding module maps WSI patch features onto a spatial grid using their original coordinates and applies parallel convolutions to construct representations at patch, local-neighborhood, and regional scales. Genomic signature embeddings then query these scale-specific WSI representations through cross-modal co-attention, producing signature-conditioned morphological features. The clinical-guided graph propagation module uses clinical variables to define patient-similarity edges and propagates fused image-genomic node embeddings with graph attention, integrating clinical information through topology rather than feature concatenation. Results: Across three TCGA cohorts, HGASurv achieved the strongest overall C-index performance, with ablation studies supporting the complementary contributions of the hierarchical WSI encoding and clinical-guided graph propagation modules. Conclusion and Significance: By introducing hierarchical WSI encoding and clinical-guided graph propagation, HGASurv unifies multi-scale pathology representation with topology-based clinical integration in a hierarchical granularity-aware framework. This design shows that matching representation and integration mechanisms to modality-specific granularity improves multimodal survival prediction.

[212] From Legacy to Contemporary: Scalable Color Modernization for Long Video cs.CVPDF

Dongyun Kim, Jiwoo Choi, Jaeseong Lee, Seon Joo Kim

TL;DR: 该免训练视频色彩现代化框架先用文本引导图像编辑模型联合处理时间相邻关键帧，再以紧凑的镜头级颜色变换传播到整段镜头，力求更新旧片色调而保持场景内容、空间结构和跨帧稳定；在跨五十年视频评估中获得最高主观现代感。

Details

Motivation: 旧影像常有偏色、色度不足、对比度弱和色调失衡，单纯提升分辨率并不能让画面看起来现代。长视频逐镜头调色仍高度依赖人工调色师，难以扩展并保持镜头间一致性。

Result: 论文提出 long video color modernization 任务：更新旧彩色视频的色彩和色调，同时保持内容及跨帧、跨镜头连贯。免训练框架使用预训练文本引导图像编辑模型，对时间相邻关键帧网格联合现代化，生成无需配对监督的连贯参考；随后学习紧凑的镜头专属颜色变换，将调整传播到镜头内所有帧，维持空间结构和时间稳定。作者对跨五十年的多种旧视频做多维评估，结合人类主观现代感与内容保真、时间稳定、感知质量及效率指标，并提出 Source-Weighted Inter-shot Grade Consistency（SW-IGC）评估跨镜头色彩等级一致性。方法获得最高主观现代感，同时在其他指标上保持竞争力。

Insight: 长视频调色可把少量相邻关键帧编辑与镜头级颜色变换结合，减少逐帧生成带来的不一致，并用源画面加权指标检查镜头间风格连贯。摘要未报告具体数值、受试者规模或对不同编辑模型的敏感性。

Abstract: Bringing legacy video back to contemporary screens does not necessarily make it look contemporary. Earlier imaging pipelines often leave color casts, muted chroma, weak con- trast, and tonal imbalance, while correcting these character- istics still relies heavily on expert colorists and scales poorly to long, multi-shot videos. We formulate long video color modernization, a task that updates the chromatic and tonal appearance of legacy color footage while preserving scene content and coherence across frames and shots. We introduce a scalable, training-free framework that jointly modernizes grids of temporally adjacent keyframes using a pretrained text-guided image editing model, producing coherent con- temporary references without paired supervision. Compact shot-specific color transforms then propagate these adjust- ments throughout each shot, preserving spatial structure and temporal stability. We conduct a multidimensional evaluation on diverse legacy videos spanning five decades. We combine a human evaluation of perceived modernness with quantita- tive measures of content fidelity, temporal stability, percep- tual quality, and efficiency, and introduce Source-Weighted Inter-shot Grade Consistency (SW-IGC) to assess grade co- herence across shots. On our evaluation set, our frame- work achieves the highest perceived modernness while re- maining competitive across complementary measures, open- ing a path toward the systematic study of long video color modernization. Code is available at https://github.com/jiwoomood/ Video_Color_Modernization.git

[213] LaMER: Language-Conditioned Multi-Expert Routing for Weakly Supervised Referring Expression Comprehension cs.CV | cs.CL | cs.LGPDF

Yu Yu Chang, Tzu-Hsuan Weng, Youssef Mohamed Amin, Mohaned Ahmed Elsayed, Hung-Hsuan Chen

TL;DR: LaMER 在弱监督指代表达理解中，通过语言条件的跨模态路由动态组合四个冻结视觉专家，并以带负锚点增强的 cell-text 匹配稳定训练；RefCOCO、RefCOCO+ 和 RefCOCOg 上 Acc@IoU≥0.5 分别为 77.67%、62.81% 和 69.49%。

Details

Motivation: 指代表达理解要根据自然语言描述在图像中定位区域。弱监督场景中，只有图像级粗信号使区域—文本对应难以学习，且不存在一个视觉编码器能同时提供语义对齐、细粒纹理和边界信息；直接融合多个预训练编码器又会产生异质表示。

Result: LaMER 使用语言条件跨模态路由融合模块，按每条指代表达动态聚合四个冻结专家编码器，无需实例级边界框监督。Cell-Text Matching Head 配合负锚点增强，并以对比、重建、对齐和负载均衡多目标损失稳定优化。RefCOCO、RefCOCO+、RefCOCOg 上 Acc@IoU≥0.5 分别达 77.67%、62.81%、69.49%；相较 DViN 和 WeakMCN，最高分别提升 16.49 和 14.49 个百分点。消融实验支持各组件作用。

Insight: 语言条件路由能按目标描述选择有用的视觉专家，避免把语义、纹理和边界能力不加区分地混合。摘要未报告各编码器贡献细节或跨领域图像上的泛化。

Abstract: Referring Expression Comprehension (REC) localizes the image region described by a natural language expression, underpinning applications such as autonomous navigation, human-robot interaction, and medical image analysis. Under weak supervision, REC is challenged by the visual-linguistic semantic gap and the difficulty of inferring region-text correspondences from coarse image-level signals alone. Moreover, no single visual encoder captures all complementary properties—semantic alignment, fine-grained texture, and boundary awareness—while naively fusing multiple pretrained encoders yields heterogeneous, hard-to-reconcile representations. We propose LaMER, a weakly supervised REC framework that dynamically aggregates four frozen expert encoders through a language-conditioned Cross-Modal Routing Fusion module, adapting expert selection to each referring expression without instance-level bounding-box supervision. A Cell-Text Matching Head with Negative Anchor Augmentation, trained with a multi-objective loss combining contrastive, reconstruction, alignment, and load-balancing terms, further stabilizes optimization. LaMER achieves 77.67%, 62.81%, and 69.49% Acc@IoU(\geq 0.5) on RefCOCO, RefCOCO+, and RefCOCOg, surpassing DViN and WeakMCN by up to 16.49 and 14.49 points, with ablations validating each component.

[214] Policy-Adaptive Credit Assignment for Spatial Reasoning with Verifiable Rewards cs.CVPDF

Hailiang Yang, Shengchao Chen

TL;DR: CALIPER 根据组内阶段覆盖与答案正确性的关联，动态调整可验证奖励的空间推理阶段权重，并用闭环控制处理估计不可靠时的回滚、增加多样性和早停；在八个基准、三个骨干上，一次训练和闭环训练的平均表现均最高。

Details

Motivation: 图像空间推理包含感知、分解和推断等阶段，训练中限制准确率的瓶颈可能变化。只给答案奖励无法定位失败阶段；过程监督虽然提供中间反馈，固定的阶段权重却无法适应训练过程中的瓶颈变化，且当所有 rollout 奖励相同，组相对反馈缺乏区分性。

Result: 作者构建 SpatialTrace 阶段分解推理轨迹数据，并提出 CALIPER。方法用组内岭回归联合估计各阶段参考覆盖与答案正确性的关联，将其作为效用代理，逐步更新阶段奖励权重；所得过程奖励也可用于所有 rollout 组，包括全部失败的组。闭环训练控制器监测效用估计可靠性和验证表现，决定是否回滚检查点、增加 rollout 多样性或早停。八个基准、三个骨干的实验中，CALIPER 在单次训练和带闭环控制训练下均取得最高平均表现，并在不同 RL 优化器下持续超过固定奖励分配。

Insight: 过程监督权重应随当前阶段瓶颈自适应变化，但阶段效用估计本身也可能因缺少能区分答案的组而失真，需配合训练监控和回滚。摘要未给出基线具体分差或效用估计误差。

Abstract: Image-based spatial reasoning requires perception, decomposition, and inference, but the stage that limits accuracy can change during training. Answer-level rewards neither localize failures to individual stages nor provide discriminative group-relative feedback when all rollouts receive the same reward. Process supervision provides intermediate feedback, but fixed weights cannot adapt reward allocation to these changing bottlenecks. To address this gap, we propose CALIPER, an adaptive process-reward framework supported by SpatialTrace, a corpus of stage-decomposed spatial reasoning traces. CALIPER jointly estimates the associations between stage-wise reference coverage and answer correctness through within-group ridge regression, using these associations as utility proxies to gradually update reward weights. The resulting process reward supervises all rollout groups, including uniformly failing ones. Because utility estimates can become unreliable when answer-discriminative groups are scarce, a closed-loop harness monitors estimator reliability and validation performance to guide checkpoint rollback, rollout diversification, and early stopping. Extensive experiments across eight benchmarks and three backbones show that CALIPER achieves the highest mean under both one-shot and harness training, with consistent gains over fixed allocation across RL optimizers. Further analyses confirm the benefits of adaptive reward allocation and harness control, with favorable scaling in model/data size.

[215] Long-form Video-to-Text Summarization using Recursive Language Models cs.CVPDF

Alessia Intini, Lorenzo Vaiani, Enrico Chen, Luca Cagliero

TL;DR: RVLSM 用场景感知分段保持长视频摘要的时间连续性，并按需调用动作识别、OCR、ASR 和目标检测工具补充事件语义；在 VideoXum 和 MixTyVSum 上持续超过零样本 VideoLLM 基线，并与先进微调模型竞争。

Details

Motivation: VideoLLM 在短视频理解上进步显著，但长视频摘要易受长上下文性能退化影响，也难保持事件时间连续性并生成语义丰富的描述。

Result: 递归视频语言摘要模型 RVLSM 面向长视频到文本摘要，先以 Semantic Temporal Decomposition（STD）根据场景语义边界动态切分视频，使片段在语义连贯的同时保持时间顺序；再用 Adaptive Multimodal Tool Calling（AMTC）按需求调用动作识别、光学字符识别、自动语音识别和目标检测模块，提升摘要的事件语义丰富度。VideoXum 和 MixTyVSum 实验中，RVLSM 在多种模型设置与评测指标上持续优于零样本 VideoLLM 基线，并可与先进微调模型竞争。其摘要命名实体与真值摘要检测实体的匹配更好，支持事件描述更具显著性。

Insight: 将长视频分成语义连续片段并按需调用多模态工具，可减少一次性长上下文处理造成的连续性和细节损失。摘要未提供具体指标增幅、工具调用次数或不同视频长度下的退化情况。

Abstract: While VideoLLMs have made significant progress in understanding short video benchmarks, they often struggle to summarize long-form ones. This is mainly due to performance degradation caused by long-context prompts, the inherent difficulty of preserving temporal continuity and a lack of semantically rich concepts in event descriptions. To tackle these issues, this paper introduces the first recursive Video Language Summarization Model (RVLSM) designed for video processing and tailored to long-form video-to-text summarization. RVLSM recursively decomposes videos into semantically coherent segments and delegates chunk-level reasoning to VideoLLM sub-models and external tools. It employs (1) Semantic Temporal Decomposition (STD), which preserves temporal coherence and continuity through a scene-aware chunking strategy that dynamically partitions videos according to semantic boundaries, and (2) Adaptive Multimodal Tool Calling (AMTC), which ensures semantic richness of the summary via a routing mechanism that selectively invokes auxiliary modules such as action recognition, optical character recognition, automatic speech recognition, and object detection. Experiments on the VideoXum and MixTyVSum benchmarks show that RVLSM consistently improves over zero-shot VideoLLM baselines across multiple model settings and performance metrics, and is competitive with a state-of-the-art fine-tuned model for video-to-text summarization. Importantly, the named entities in the RVLSM’s summaries align more closely with those detected in the ground-truth summaries than those detected by baseline and state-of-the-art methods, confirming the saliency of the event descriptions it generates.

[216] YUTELA: A Scalable Edge Cloud AIoT System for Multi-Site Video Anomaly Detection with Scene Specific Adaptation cs.CVPDF

Shanle Yao, Babak Rahimi Ardabili, Amir Kimiyaie, Hamed Tabkhi

TL;DR: YUTELA 将每路摄像头的异常模型部署到 Jetson Thor 边缘节点，并由集中 GPU 服务器基于近期伪正常数据按需重训；四个公共地点、31 路 CCTV 连续运行 170 小时，验证其可扩展的多站点适配流程。

Details

Motivation: 长时多地点视频异常检测需要在摄像头端持续处理，同时适应不同场景。原 Ancilia 依赖现场工作站，论文希望改为边缘—服务器—云系统，并减少多摄像头数据传输和部署成本。

Result: YUTELA 用 NVIDIA Jetson Thor 替代现场工作站，改造多摄像头处理流程，包括相机批处理与再分配、共享内存、减少阶段间数据传输、任务整合和 TensorRT 推理。每个摄像头维护独立异常模型，收集到足够近期伪正常数据后由中央 GPU 重训并将权重回传。八路摄像头正常密度下，TensorRT 将吞吐由 14.17 提至 29.94 FPS；优化边缘节点在九种负载设置中有七种超过工作站版 Ancilia，极高密度时最高为其 2.03 倍，但四路或八路轻负载下延迟更高。系统随后在四个公共地点的 31 路既有 CCTV 上连续运行 170 小时。运行负载主要由场景活动和帧率决定，而非摄像头数量；地点间重训频率约差 3 倍。101 次自动触发训练仅占单 GPU 约 6.6%，平均端到端告警延迟按地点流量和异常类型为 6.8–15.4 秒。

Insight: 每路相机单独适配并通过集中节点定期重训，可覆盖异质地点；但轻负载延迟、场景活跃程度和帧率会影响边缘服务表现。摘要提供了长期部署证据，未给异常检测准确率或告警误报率。

Abstract: Building upon our previous Ancilia framework, this work presents \emph{YUTELA}, an edge–server–cloud AIoT system for long-term, multi-site video anomaly detection. YUTELA replaces workstation-based on-site processing with NVIDIA Jetson Thor edge devices and redesigns the multi-camera pipeline through camera batching and redistribution, shared memory, reduced inter-stage data transfer, task integration, and TensorRT-accelerated inference. It maintains a separate anomaly model per camera, retrained on a centralized GPU server once enough recent pseudo-normal data are collected, with updated weights returned to the edge. With eight cameras, TensorRT increases throughput from 14.17 to 29.94 FPS under normal density, and the optimized edge node exceeds the throughput of the workstation-based Ancilia system in seven of nine workload settings, reaching up to 2.03 times its throughput at extreme density. This gain comes at the cost of higher latency under lighter loads with four or eight streams. YUTELA was then operated continuously for 170 hours at four public sites using 31 existing CCTV cameras. The deployment shows that runtime load is driven by scene activity and input frame rate rather than by camera count, and that retraining frequency varies by about three times across sites. The 101 automatically triggered training events occupied only about 6.6% of a single GPU, indicating that one global node can support adaptation for tens of sites. Mean end-to-end alert latency ranged from 6.8 to 15.4~s depending on site traffic and anomaly type. These results show that YUTELA enables practical multi-camera edge processing with low-cost, camera-specific adaptation across heterogeneous sites.

[217] Holi-World: Building Holistic World Intelligence from Human-Ego Experience cs.CVPDF

Hao Li, Xiyuan Zheng, Yuanyuan Gao, Yalun Dai, Xinhao Ji

TL;DR: HOLI-WORLD 将第一视角经验采集、世界表示、自动数据构造、模型学习和独立评估串联起来，从约 100 个场景自动生成约 8 万问答对训练 4B 模型；在 ScanNet、ScanNet++、Aria Digital Twin 和自建基准上验证静态感知、动态跟踪及空间时间推理。

Details

Motivation: 真实第一视角经验包含世界结构如何组织以及交互如何改变场景的丰富信息，但要规模化学习，需把持续增长的视频转换成跨视角、跨时间关联场景结构、对象语义和交互历史的资源，同时减少昂贵人工标注。

Result: HOLI-WORLD 是连接体验采集、世界表示、自动数据构造、模型训练和独立评测的整体生态。HOLIWORLD-Foundry 结合静态与动态感知，自动生成将持久场景结构、对象状态变化、交互历史和支持观察联系起来的世界记录；HOLIWORLD-Data 将约 100 个自采场景组织为学习资源，含约 8 万问答对，用于训练 HOLIWORLD-VL-4B。独立人工标注留出录像构成 HOLIWORLD-Bench，评估标注质量及空间、时间理解。ScanNet、ScanNet++、Aria Digital Twin 和自建基准实验显示静态感知和动态跟踪表现较强；用该数据训练显著提升模型在 HOLI-WORLD-Bench 和 EgoDynamic4D 上的空间与时间推理。

Insight: 把连续第一视角经验整理为带观察依据、场景结构和交互历史的世界记录，可同时支持训练数据自动化和独立时空理解评估。摘要未给出各数据集具体分数、自动标注误差或场景外推能力。

Abstract: Human egocentric experience offers a rich foundation for learning how the world is organized and how it changes through interaction. Realizing this potential at scale requires turning growing volumes of recordings into learning resources that connect scene structure, object semantics, and interaction histories across viewpoints and time, while reducing dependence on costly manual annotation. We introduce HOLI-WORLD, an integrated ecosystem connecting experience capture, world representation, automated data construction, model learning, and independent evaluation. To sup- port learning at scale, HOLIWORLD-Foundry combines complementary static and dynamic perception to automatically construct world records that connect persistent scene structure and evolving object states with interaction histories and supporting observations. HOLIWORLD-Data organizes these records from approximately 100 self-collected scenes into grounded learning resources, including approximately 80K QA pairs used to train HOLIWORLD-VL-4B. Independent human annotations of heldout recordings form HOLIWORLD-Bench, supporting evaluation of both annotation quality and learned spatial and temporal understanding. Experiments on ScanNet, ScanNet++, Aria Digital Twin, and HOLIWORLD-Bench demonstrate strong performance in static perception and dynamic tracking, while training on HOLIWORLD-Data substantially improves HOLIWORLD-VL-4B’s spatial and temporal reasoning on HOLI- WORLD-Bench and EgoDynamic4D. HOLIWORLD provides a practical foundation for expanding learning resources from the continuous flow of human experience and advancing machine understanding of the dynamic world.

[218] PrismWAM: Decoupling State Representations in World Action Models cs.CVPDF

Tianyu Xin, Wanqi Zhong, Bowen Hu, Zhangchi Guo, Lijie Wang

TL;DR: PrismWAM 将世界动作模型状态拆成语义外观与动作驱动的物理演化，按“语义规划—深度空间物理 rollout—可执行动作”流程从单张初始 RGB 图像控制操作；无需具身动作预训练或额外数据增强，在仿真和现实操作中都表现强。

Details

Motivation: 世界动作模型常直接使用预训练视频模型潜变量表示世界，把纹理、颜色等外观与占据、姿态等物理状态混在一起。论文提出 WAM 的状态应区分语义上下文和动作作用下的物理状态，重点从“世界看起来怎样”转向“世界会如何响应动作”。

Result: PrismWAM 将具身任务拆成语义规划、物理 rollout 和可执行动作生成三阶段：只给初始 RGB 观察，先预测最终目标图像，再在深度空间预测到达目标的细粒度演化，最后从深度 rollout 生成可执行动作。该方法不使用具身动作预训练，也不做额外数据增强；仿真与真实环境的操作实验表现强，并超过多种使用大规模具身动作预训练的模型，包括 OpenWAM-α。

Insight: 把最终语义目标与实现目标所需的物理过程分开建模，可能减少视频外观潜变量对动作状态的干扰。摘要未给出具体操作成功率、任务列表或与具身预训练模型的计算成本比较。

Abstract: World Action Models (WAMs) typically represent the world through latent embeddings from pretrained video models, which entangle heterogeneous aspects of the world, from how it looks (e.g., texture and color) to how it is physically configured (e.g., occupancy and pose). This raises an underexplored question: $\textbf{what should constitute the state of a WAM?}$ We argue that the world state should separate how the world appears from how the world evolves under action: the former captures semantic context, while the latter is governed by physical state. To this end, we propose $\textbf{PrismWAM}$, which decouples physics from semantics while jointly modeling state and action. PrismWAM factorizes an embodied task into a $\mathrm{Semantic\ Planning}\rightarrow\mathrm{Physical\ Rollout}\rightarrow\mathrm{Executable\ Action}$ pipeline: given only the initial RGB observation, semantic planning learns to predict the final goal image, while physical rollout predicts the fine-grained evolution toward this goal in depth space; executable actions are then generated from the depth rollout. Without any embodied action pretraining or additional data augmentation, PrismWAM achieves strong manipulation performance in both simulation and the real world. Notably, PrismWAM even surpasses a broad range of models with large-scale embodied action pretraining (e.g., OpenWAM-$α$). More broadly, we reframe World Action Models from modeling how the world looks to how it responds to action, paving the way for action-centric world models grounded in occupancy, flow, tactile sensing, proprioception, and beyond.

[219] Drive2Gauss: FeedForward 4D Gaussian Reconstruction from Multimodal Generative Driving Videos cs.CVPDF

Tong Zhao, Zhiyuan Han, Lin Chen, Bowen Xie, Yunxi Qiao

TL;DR: Drive2Gauss 将多模态驾驶视频生成器输出的 RGB、度量深度和动态光流一次前馈转为动态 4D 高斯场景，避免每段视频单独优化；nuScenes 上未来 RGB 生成 FID 为 6.38、FVD 为 12.43，重建所生成观测的 PSNR 为 29.91 dB。

Details

Motivation: 生成可复用驾驶环境不仅需要能控制未来观察，还需要可从不同相机视点渲染的显式场景表示。现有方法通常对每段生成 RGB 视频单独运行优化来恢复场景，成本高且缺乏直接前馈重建。

Result: Drive2Gauss 使用多模态视频生成器，根据参考多视角图像和未来控制信号预测 RGB、度量深度与动态光流。高斯查询解码器以深度初始化稀疏 3D 查询，用光流建立跨时间关联，再通过时空特征聚合与迭代查询细化融合外观。场景重建一次前馈完成，无需逐场景优化。nuScenes 结果中，未来 RGB 生成 FID 为 6.38、FVD 为 12.43；重建生成观测的 PSNR 为 29.91 dB，并支持横向相机偏移下直接渲染及下游视频修复。

Insight: 同步生成深度和运动线索能为显式动态场景重建提供几何初始化与跨时关联，减少对逐场景迭代优化的依赖。摘要未给出横向视点渲染或视频修复的量化成绩。

Abstract: Generating reusable driving environments requires both controllable future observations and an explicit representation that supports rendering from different viewpoints, yet existing methods still recover explicit scenes from generated RGB videos through optimization for each sequence. To overcome these limitations, we present Drive2Gauss, a feedforward framework that converts jointly generated RGB, metric depth, and dynamic flow into dynamic 4D Gaussian scenes. A Multimodal Video Generator predicts the three modalities from reference multiview images and future control signals. Furthermore, we introduce a Gaussian Query Decoder that uses depth to initialize sparse 3D queries and flow to associate the queries across time, and integrates appearance information through spatiotemporal feature aggregation and iterative query refinement. The reconstruction stage operates in a single feedforward pass without optimization for each scene. Extensive results on nuScenes demonstrate that our approach attains an FID of 6.38 and an FVD of 12.43 for future RGB generation, a PSNR of 29.91 dB when reconstructing its generated observations, and supports both direct rendering under lateral camera shifts and downstream video fixing. Code and more visualizations are available at https://zttt04. github.io/Drive2Gauss_page/.

[220] Scaling Vision Foundation Models for Toxicologic Pathology cs.CV | eess.IVPDF

Jessica Loo, Theodoros Pissas, Byunghak Kang, Vincent A. Ella, Xiaoyan Zhang

TL;DR: TOX-FM 是面向非临床毒理病理的视觉基础模型系列，基于超过 100 万张全切片图像，覆盖六个动物物种、60 多个器官、14,000 多项研究和六十年数据；论文建立 18 项非临床下游任务基准，并纳入 9 项临床任务比较。

Details

Motivation: 病理视觉基础模型通常基于临床数据训练，其在动物毒理病理领域的适用性、领域专属训练价值及跨临床—非临床泛化能力仍不清楚。论文希望系统扩展数据和评测资源。

Result: 作者提出 TOX-FM 模型家族，训练数据包含超过 100 万张组织病理全切片图像，覆盖六种动物、60 多个器官、14,000 多项毒理研究及六十年时间跨度。研究构建 18 项非临床下游任务的综合基准，并加入既有的 9 项临床下游任务，与先进临床病理基础模型在非临床和临床两侧比较，以分析领域专属训练和跨域泛化中的机会与挑战。

Insight: 大规模、跨物种且跨时间的非临床训练资源可支撑毒理病理模型研究，并允许同时检查其临床迁移能力。摘要只概述数据和评测范围，没有给出具体任务成绩或模型间优劣。

Abstract: Here, we present TOX-FM, a family of non-clinical PFMs trained on large-scale diverse histopathology datasets of over one million whole slide images consisting of over 60 organs across six animal species from over 14,000 toxicology studies across six decades. We introduce a comprehensive benchmark of 18 non-clinical downstream tasks, incorporate an established benchmark of 9 clinical downstream tasks, and evaluate TOX-FM against state-of-the-art clinical PFMs across both non-clinical and clinical domains. Notably, we explore open questions surrounding domain-specific training and cross-domain generalization, offering insights into the opportunities and challenges of applying PFMs in the non-clinical domain.

[221] EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience cs.CVPDF

Xinhao Ji, Hui En Pang, Hao Li, Kairui Hu, Jiner Wang

TL;DR: EgoXperience 提供从自动标注、数据集、联合重建模型到基准的第一视角全栈方案；其 50 小时 EgoX-Data-6M 用于训练单次前向模型，从单目视频和首帧物体掩码联合重建相机、场景、手和物体，多个重建误差较最强基线降低 42.9%–68.9%。

Details

Motivation: 从第一视角视频重建人的活动，不仅要恢复各个 3D 对象，还要让观察者运动、周围几何和手—物交互在同一世界坐标系中保持一致。现有方法和数据集对这种联合重建及对齐监督支持有限。

Result: EgoXperience 通过采集、自动标注、统一建模与基准评测形成全栈方案。EgoX-Engine 把经几何验证的目标跟踪与基于 SLAM 的世界对齐结合，统一监督相机、场景、手和物体；由此构建 EgoX-Data-6M，含 50 小时室内外操作和行走活动以及秒级任务描述。研究从这些活动中选取一小时建立 EgoX-Bench。Ego 模型结合该数据及其他公开标注训练，输入单目第一视角视频和首帧物体掩码，一次前向就在共享世界坐标中联合度量重建相机、场景、手和物体。在场景估计及世界坐标手—物姿态恢复上达到论文报告的先进水平；相较最强基线，相机、手和深度误差分别降低 68.9%、65.6%、42.9%，物体姿态分数提高 28.0%。

Insight: 第一视角重建的关键在于相机、场景与交互对象共享几何一致的坐标监督；自动对齐可把大量连续体验转化为联合重建数据。摘要未提供各任务绝对误差或自动标注本身的误差率。

Abstract: Reconstructing human activity from egocentric video requires more than recovering individual 3D components: the observer’s motion, surrounding geometry, and hand-object interactions must fit together in a consistent world. Achieving this coherence requires both joint reconstruction and aligned supervision, yet existing methods and datasets offer limited support for either. We present EgoXperience, a full-stack paradigm that addresses these needs through data collection, automatic annotation, unified modeling, and benchmarking. Central to EgoXperience is an automatic annotation engine, EgoX-Engine, that couples geometrically verified object tracking with SLAM-based world alignment, yielding consistent supervision for the camera, scene, hands, and objects. This yields EgoX-Data-6M, a 50-hour dataset of task-driven manipulation and locomotion within and across indoor and outdoor environments, with second-level task captions. We carefully select one hour across these activity categories as EgoX-Bench to evaluate joint reconstruction. To translate this supervision into reconstruction, we introduce EgoX, trained with EgoX-Data-6M and complementary annotations from additional public datasets. Given monocular egocentric video and first-frame object masks, our model achieves joint metric reconstruction of the camera, scene, hands, and objects in a shared world coordinate frame with a single forward pass. Extensive experiments demonstrate state-of-the-art performance across scene estimation and world-space hand-object pose recovery. Against the strongest evaluated baselines, EgoX reduces camera, hand, and depth errors by 68.9%, 65.6%, and 42.9%, and improves object pose scores by 28.0%.

[222] SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes? cs.CV | cs.CLPDF

Zhaoxi Zhang, Hao Liang, Linxuan Wu, Yiming Wu, Yifu Yuan

TL;DR: SpatialScout 评测 VLM 智能体能否主动探索室内场景并构建公制 3D 对象地图；80 个场景、3,140 个目标上的最强模型平均重叠分数仅 41.71%，IoU>0.9 的目标仅 3.25%，显示精确完整建图仍是明显短板。

Details

Motivation: 空间问答成绩无法证明模型能否通过自主探索获得准确的度量地图。空间智能体需要选择视点、跨视角关联对象、估计几何并在有限交互预算内修订地图。

Result: SpatialScout 包含 80 个多样室内场景和 3,140 个目标对象。智能体获得对象清单、RGB 观察、相机内参和位姿后，可控制相机探索，并在共享坐标系估计各对象中心。评测联合考察定位准确率与目标覆盖率，使用对象尺度重叠分数惩罚漏检。前沿 VLM 实验中，最佳模型 Claude Opus 5.5 平均重叠分数为 41.71%；47.23% 目标的 IoU 超过 0.5，而仅 3.25% 超过 0.9。

Insight: 当前模型较能恢复粗略对象布局，却难构建精确、完整的度量地图；主动感知评测应同时惩罚定位偏差和目标遗漏。摘要未报告不同模型之间的完整排名或不同交互预算的成绩。

Abstract: Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through autonomous exploration. We introduce SpatialScout, a benchmark for evaluating vision-language model (VLM) agents on active object-level 3D mapping. SpatialScout comprises 80 diverse indoor scenes containing 3,140 target objects. Given an object inventory, RGB observations, camera calibration, and camera poses, agents control a camera to explore each scene and estimate object centers in a shared coordinate frame. The task requires coordinating viewpoint selection, cross-view association, geometric estimation, and map revision under bounded interaction budgets. Evaluation jointly measures localization accuracy and target coverage through object-scale overlap scores that penalize missing predictions. Experiments with frontier VLMs reveal substantial limitations: the strongest evaluated model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with 47.23% of targets exceeding an intersection-over-union (IoU) threshold of 0.5 and only 3.25% exceeding 0.9. These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps. SpatialScout provides a quantitative testbed for developing agents that build reliable 3D spatial representations through active perception. Code is available at https://github.com/JesseZZZZZ/SpatialScout.

[223] CaptionFuse: Training-Free Composed Product Retrieval Via VLM Intent Captioning and Weighted Embedding Fusion cs.CVPDF

Zhao Gao, Mina Metias, Mohamed Noordeen Alaudeen, Bassel Saleh

TL;DR: CaptionFuse 用 VLM 联合理解商品参考图和文字修改，生成补全未说明保留属性的视觉描述，再进行训练无关的语言空间与图像嵌入融合；控制实验显示属性内容而非句式决定收益，描述属性相同时 R@10 相同，较原始修改文字高 6.5 分。

Details

Motivation: 组合图像检索需要按自然语言修改参考商品图，同时保留用户没提及的属性。现有免训练方法常直接算术融合 CLIP 向量，但用户文字通常只说明要改什么，不会重述要保留的花纹、领口和版型等信息。

Result: CaptionFuse 使用视觉语言模型共同解释参考图与修改文本，生成有视觉依据的商品标题式描述，补回未明确写出的保留属性；随后在语言空间融合该描述与查询表示。方法无需任务专属训练且不依赖特定模型。受控实验比较属性一致但句式不同的祈使式和描述式字幕，两者 R@10 相同，均比原始修改文字高 6.5 分，说明增益来自补充的属性内容，而非语言风格。

Insight: 组合检索的文字输入往往稀疏，利用参考图补齐“要保留什么”可减少属性丢失；对检索而言，属性信息比改写句式更关键。摘要未报告各数据集指标或最终方法在标准测试集上的完整成绩。

Abstract: Composed Image Retrieval (CIR) retrieves products matching a reference image modified by natural language—a core capability for eCommerce search. Existing training-free methods fuse CLIP embeddings via arithmetic, but we identify a fundamental limitation: user text modifications are attribute-sparse—a user who types make it red with shorter sleeves'' specifies what to change but not what to keep (the floral pattern, the V-neck, the fit-and-flare silhouette). We introduce language-space fusion: a Vision-Language Model (VLM) jointly interprets the reference image and modification to generate a visually grounded product-title caption---e.g., Red floral fit-and-flare dress with short sleeves and V-neck’’—that fills in the missing attributes from the image. Our method, CaptionFuse, requires no task-specific training and is model-agnostic. A controlled experiment confirms that attribute content— not linguistic style—drives the gains: imperative and descriptive captions with identical attributes achieve equal R@10, while both outperform raw modifications by +6.5 points.

[224] From Short Clips to Real-Time Streams: Identity-Aware Supervision and Context-Efficient Generation for Video Face Swapping cs.CVPDF

Zekai Luo, Yuwei Liang, Ruiqi Ma, Haoyu Huang, Xuehao Wang

TL;DR: StreamingSwap 以噪声水平条件的身份与重建联合损失强化身份迁移，并用人脸区域生成与短上下文训练降低长视频成本；在单张 RTX 4090 上达到 20 FPS，视频最长 10 分钟仍稳定生成。

Details

Motivation: 参考引导的人脸交换需把参考身份迁移到目标视频，同时保留其他视觉属性和运动。现有流匹配方法难以同时保证身份忠实度、图像质量和长视频实时生成。

Result: 作者提出 StreamingSwap，并构建含 1,000 小时电影训练数据及各 200 条短、长视频评测集的 HumanFaceVid。由于换脸只需改变身份但保留大部分画面内容，普通流匹配监督对身份变化不够敏感；Joint Rec-ID Loss 根据噪声水平平衡显式身份监督和潜空间重建。利用目标视频已有的非编辑属性及时间动态，Face-RoI Generation 只生成面部区域，Short-Context Rollout Training 缩短时间上下文。单张 NVIDIA RTX 4090 上达到 20 FPS，身份迁移和视觉质量具有竞争力，并能在最长 10 分钟视频上稳定生成。

Insight: 当编辑仅涉及局部身份属性时，保留目标视频其余空间和时间信息可减少生成开销；身份损失需随噪声阶段与重建目标平衡。摘要未提供相较其他方法的具体质量指标或长视频各长度的稳定性曲线。

Abstract: Video face swapping transfers the reference identity to a target video while preserving its other attributes and motion. Despite recent progress in reference-guided flow-based methods, jointly achieving faithful identity transfer, high fidelity, and real-time generation over long video streams remains challenging. To address this challenge, we rethink the task-specific structure of reference-guided video face swapping. First, face swapping changes only identity while preserving most visual content, yet flow-matching supervision is weakly sensitive to these changes. We therefore develop Joint Rec-ID Loss, which balances explicit identity supervision and latent reconstruction according to the noise level. Second, the target video already provides non-edited attributes and temporal dynamics, enabling generation with fewer spatial tokens and shorter temporal contexts through Face-RoI Generation and Short-Context Rollout Training. By integrating these designs, we propose StreamingSwap, a reference-guided flow-matching framework for real-time streaming video face swapping. We also introduce HumanFaceVid, comprising 1,000 hours of cinematic training data and evaluation sets of 200 short and 200 long videos. Experiments demonstrate that our method achieves competitive identity transfer and visual quality at 20 FPS on a single NVIDIA RTX 4090 GPU, with stable generation over videos up to 10 minutes long. Project webpage: https://aim-uofa.github.io/StreamingSwap/

[225] PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements cs.CVPDF

Kaiwei Chen, Moran Li, Zhenye Gan, Yabiao Wang, Chengjie Wang

TL;DR: PVBench 以物体级物理量测评视频生成的几何、运动、材料、因果与守恒合理性，覆盖 7 个领域、46 个维度和 600 个提示，其中 30% 含具身智能体，并通过人类评估验证自动指标一致性。

Details

Motivation: 视频模型已能生成视觉吸引人的结果，但其物理合理性缺少充分量化。现有基准往往依赖粗粒度 VLM 判断，或评估具身世界模型而非视频生成本身。

Result: PVBench 定义 7 个评测领域、46 个物理维度，涵盖指令与语义忠实度、行为和因果一致性、感知与时间质量、几何与空间一致性、材料与场状态、运动学与时空持久性，以及守恒定律等机械合理性。基准含 600 个精心设计提示，其中 30% 涉及机械臂、人形、四足和多智能体等具身对象。对可精确测量的难点维度，配套专门视觉模型提取物体级 6D 姿态、度量尺度、表面法线、运动学和数值守恒残差等可验证指标，同时支持文生视频与图生视频。人类评估用于验证自动指标一致性；作者也评估了主流开源和闭源模型。

Insight: 物理合理性应拆成可测量的对象级几何与动力学属性，而非只用整体视觉偏好分数；人类一致性验证可帮助判断自动指标是否有意义。摘要没有报告模型评测排名、自动—人类相关系数或各维度覆盖限制。

Abstract: Recent video generation models produce visually compelling outputs, yet their physical plausibility remains poorly measured. Physical plausibility asks whether generated objects obey geometry, kinematics, dynamics, and conservation laws. Existing benchmarks rely on coarse visual-language model judgments or target embodied world models rather than video generation itself. We introduce PVBench, a physical video benchmark that quantitatively measures the plausibility of generated videos across 46 dimensions organized into seven domains, spanning: i) instruction and semantic fidelity, ii) behavioral and causal consistency, iii) perceptual and temporal quality, iv) geometric and spatial consistency, v) material properties and field states, vi) kinematic and spatiotemporal persistence, vii) and mechanical plausibility with conservation laws. PVBench comprises 600 carefully crafted prompts, of which 30% involve physical embodied agents such as robot arms, humanoids, quadrupeds, and multi-agent systems. For challenging physically tractable dimensions, it pairs each with specialized vision models that derive verifiable, physically grounded measurements, such as object-level 6D pose, metric scale, surface normals, kinematics, and numerical conservation-law residuals. It thus avoids relying solely on subjective scoring, while supporting both text-to-video and image-to-video generation. Finally, we validate the agreement between automatic metrics and human judgments through human evaluation. Meanwhile, we evaluate mainstream open-source and closed-source models and present several insights. By releasing the benchmark, prompts, and calibration protocol, we aim to foster physically grounded video generation.

[226] A Flexible ROI Analysis Tool for Spatially Consistent Vegetation Index Extraction from Turfgrass Imagery cs.CVPDF

Kibon Ku, Thai Thanh Tuan, Jinhyun Ahn

TL;DR: 该开源工具用可调整方向和 N 路分割的 ROI 图形界面配合批处理脚本，从多光谱草坪影像中重复提取地块植被指数；保存的像素坐标可复用于同一时点的 NDVI、NDRE 和 SAVI 栅格，无需重新圈选。

Details

Motivation: 多时相多光谱成像可用于评估草坪状态，但相机位置随采集变化、自动地块跟踪不可靠时，逐地块提取植被指数仍费力，重复测量也容易因 ROI 不一致而降低可比性。

Result: 作者发布的软件包含两个部分：图形界面可交互放置方向可调的感兴趣区域，并按地块几何做 N 路分割；批处理脚本从单波段浮点栅格计算地块平均植被指数，并以 JSON 导出测量值和 ROI 坐标以供重用。坐标可直接应用于相同时间点的 NDVI、NDRE 和 SAVI 栅格。在 18 张轨道式多光谱草坪影像的演示中，作者一次会话提取地块均值后，直接复用坐标完成另外两种指数测量。

Insight: 保存并复用像素 ROI 能在相机位置变化、静态传感器部署不便时保持地块测量区域一致，降低重复人工勾画负担。摘要未提供工具耗时对照、测量误差或跨时间点配准能力。

Abstract: Time-series multispectral imaging is widely used to assess turfgrass condition; however, extracting plot-level vegetation index (VI) values remains labor-intensive when camera positions vary between acquisitions and automated plot tracking is unreliable. We developed an open-source software tool comprising (i) a graphical user interface (GUI) for the interactive placement of orientation-adjustable regions of interest (ROIs) with $N$-way subdivision to match plot geometry and (ii) a batch-processing script that computes plot-level mean VI values from single-band floating-point rasters and exports measurements and ROI coordinates in JSON format for reuse. Saved coordinates can be reapplied to measure NDVI, NDRE, and SAVI from identical pixel regions within a single time point without re-delineating ROIs. In a demonstration using rail-mounted multispectral turfgrass imagery (18 images), plot-level mean values were extracted in a single session, with coordinates reused directly for NDRE and SAVI rasters. The system reads non-georeferenced pixel-coordinate rasters natively, does not require GIS expertise, and supports reproducible VI quantification where static sensor deployment is impractical.

[227] CheckDistill: Adaptive Checklist Distillation for Efficient Distributional Reward Modeling in Image Editing cs.CVPDF

Yaqi Cai, Zuan Gao, Mengting Chen, Wei Li, Bo Zheng

TL;DR: CheckDistill 将图像编辑指令拆成可独立核验的自适应清单，并以四维分数分布及结构化证据蒸馏教师判断；在 MMRB2、MER-Bench 和 EditReward-Bench 上准确率分别为 72.30%、56.67% 和 82.90%，超过所比较的开源基线。

Details

Motivation: 指令图像编辑强化学习需要可靠、高效的奖励模型。单一标量压缩了多维偏好，生成式评审虽能给出推理，却因通用评分标准未覆盖每个实例的编辑和保留要求而不够具体，长推理轨迹又使在线优化昂贵。

Result: CheckDistill 是点式奖励建模框架，将每个编辑条件分解为原子、可独立验证的要求，并构造四维评分标签和结构化核验依据。学生先接受监督分数初始化，再由特权教师依据清单、条目级判断及维度级理由监督学生评分轨迹。可靠性加权的 on-policy 自蒸馏根据教师预期分数与参考标签的一致度调整分数 token 监督，并在截断 token 支持集内保留候选概率权重。部署学生无需清单或显式推理，直接预测四维分数分布并将期望聚合为标量奖励。MMRB2、MER-Bench、EditReward-Bench 上聚合准确率分别达 72.30%、56.67%、82.90%，均超过所比开源基线。

Insight: 把编辑要求拆为可核验子条件，能让奖励模型显式处理实例级修改与内容保留要求，同时减少部署时的长篇评审推理。摘要未给出各基线具体分数、训练开销或清单生成成本。

Abstract: Reinforcement learning for instruction-guided image editing requires reward models that provide reliable, accurate, and efficient feedback. Direct scalar prediction compresses multidimensional editing preferences, whereas generative evaluators can expose their judgment through explicit reasoning. However, broad scoring criteria do not specify all instance-dependent editing and preservation constraints, and generating long reasoning trajectories is costly for online optimization. We introduce CheckDistill, a pointwise reward modeling framework that internalizes instruction-adaptive checklists. We decompose each editing condition into atomic, independently verifiable requirements and construct four-dimensional scoring labels together with structured verification evidence. After supervised score initialization, a privileged teacher conditions on checklists, item-level judgments, and dimension-level rationales to supervise the student’s own scoring trajectories. A reliability-weighted on-policy self-distillation objective adjusts supervision at score tokens according to agreement between the teacher’s expected scores and reference labels, while retaining candidate probability weights in a truncated token support. The deployed student directly predicts four score distributions without checklist inputs or explicit reasoning and aggregates their expectations into a scalar reward. CheckDistill achieves aggregate accuracies of 72.30%, 56.67% and 82.90% on MMRB2, MER-Bench and EditReward-Bench, respectively, exceeding the compared open-source baselines on all three metrics.

[228] Beyond Omni: A Unified MLLM for Embodied Sensory Understanding cs.CVPDF

Shiyun Xiong, Dongming Wu, Peiwen Sun, Bokang Yang, Jiaxu Wang

TL;DR: Unisense 通过先将多种实体感知表示对齐到语言、再在共享 LLM 中调优模态专家，构建覆盖八种感觉模态的具身多模态模型；其 100 万样本训练集支持 12 个深度、点云、IMU、触觉、热成像、雷达、运动和嗅觉基准，并在多数任务超过通用及专门模型。

Details

Motivation: 具身系统越来越依赖多模态大语言模型的理解和推理，但其物理感知通常限于视觉和音频。扩展到更多环境感觉模态面临表示结构与统计差异大、且难与 LLM 共享的高层语义推理连接的问题。

Result: Unisense 以两阶段训练统一多种物理感知：先把感觉表示与语言对齐，建立语言可访问接口；再在共享 LLM 中调优模态专属专家。作者整理 Unisense-1M，含八种感知模态共 100 万训练样本，并在 12 个基准评估深度、点云、IMU、触觉、热成像、雷达、运动和嗅觉理解。跨异质任务，模型在多数基准上超过通用模型和专门模型。

Insight: 语言对齐接口与模态专家分工，可以在保留各感觉模态特性的同时共享高层理解和推理。摘要没有给出各基准的具体分数、资源成本或未覆盖传感器类型。

Abstract: Recent embodied systems increasingly build on multimodal large language models (MLLMs) for their pretrained understanding and reasoning capabilities, but their perception of the physical world remains largely limited to basic modalities such as vision and audio. However, broadening MLLMs to more diverse sensory modalities from the physical environment is challenging: their representations are often weakly connected to the MLLM and vary substantially in structure and statistics, while their interpretation still requires shared high-level understanding and reasoning. We therefore propose Unisense, a unified embodied MLLM framework that supports diverse physical sensory modalities within a shared multimodal architecture. Unisense follows a two-stage training process: (i) aligning sensory representations with language to establish a language-accessible interface; and (ii) tuning modality-specific experts within the shared LLM to provide specialized processing for different sensory modalities. To support broad sensory learning, we curate Unisense-1M, a large-scale embodied sensory training corpus comprising 1M samples across 8 sensory modalities. We further evaluate Unisense on 12 benchmarks spanning depth, point cloud, IMU, tactile, thermal, radar, motion, and smell understanding. Across these heterogeneous sensing tasks, Unisense achieves strong performance and outperforms both generalist and specialized models on most benchmarks, demonstrating the potential of pretrained MLLMs as unified models for embodied sensory understanding.

[229] V-JEPA Hand: Simple, Accurate and Fast 3D Hand and Camera Trajectory Recovery from Egocentric Videos cs.CVPDF

Ruicheng Li, Qixiu Li, Yu Deng, Sicheng Xu, Jiaolong Yang

TL;DR: V-JEPA Hand 用单次前向视频模型联合恢复第一视角双手动作和相机自运动，并将手轨迹变换到统一世界坐标；在 ARCTIC、HOT3D 和 EgoStandard 上取得可比基线中最低的覆盖率感知 MPJPE，单张 RTX A6000 吞吐为 93.4 FPS。

Details

Motivation: 从第一视角视频恢复公制 3D 手轨迹，既要估计手部运动，也要估计相机自运动，才能把手放进一致参考坐标。传统流水线往往把手检测、手部重建、相机跟踪和轨迹优化分开处理。

Result: V-JEPA Hand 是前馈视频模型，用自监督 V-JEPA 2.1 编码器从全帧视频提取时空特征，再通过查询式预测器输出逐帧双手姿态、手是否可见及相机运动，并用预测的相机运动把手轨迹变换到共同世界坐标。模型可使用已知焦距，或在焦距缺失时自行预测；整段视频一次前向完成重建。训练采用单阶段端到端混合数据，包含多样伪标签视频和高精度 3D 标注。在 ARCTIC、HOT3D、EgoStandard 上，覆盖率感知 MPJPE 低于可比基线；世界坐标轨迹误差也低于对比 SLAM 流程。消融支持视频预训练、时间上下文和混合监督的作用，单张 RTX A6000 前向吞吐为 93.4 FPS。

Insight: 手运动和相机自运动需要联合估计后才能获得稳定世界坐标轨迹；将二者拆成独立流水线可能引入坐标误差。摘要未提供误差绝对值、伪标签规模和跨设备部署结果。

Abstract: Egocentric video supports diverse applications, including robot learning, augmented reality, and human activity understanding. Recovering metric 3D hand trajectories from such video requires estimating both hand motion and camera egomotion to express the hands in a consistent reference frame. Many existing pipelines recover hand and camera motion in separate stages, combining hand detection, hand reconstruction, camera tracking, and trajectory refinement. We present V-JEPA Hand, a simple feed-forward video model that jointly reconstructs bilateral hand motion and camera motion from egocentric clips. It uses a self-supervised V-JEPA 2.1 encoder to extract spatio-temporal features from full-frame video and a query-based predictor to decode these features into per-frame hand poses, hand presence, and camera motion. We use the predicted camera motion to express the hand trajectories in a common world frame. The model accepts known focal lengths or predicts them when unavailable. It reconstructs each clip in a single forward pass. We train both components end to end in a single stage on a mixture of diverse pseudo-labelled videos and high-precision 3D annotations. A single model achieves the lowest coverage-aware mean per-joint position error (MPJPE) among comparable baselines on ARCTIC, HOT3D, and EgoStandard. It also yields lower world-frame trajectory errors than the compared SLAM-based pipelines. Ablations support the benefits of video pretraining, temporal context, and mixed supervision. V-JEPA Hand achieves a forward-pass throughput of 93.4 FPS on a single RTX A6000. Project page: https://qft-333.github.io/V-JEPA-Hand-Page/.

[230] Mise-en-Video: Fine-Grained 3D Control for Multi-Shot Video Generation cs.CV | cs.AIPDF

Yingshuang Zou, Xiaoyu Shi, Yawen Luo, Ziang Li, Qinghe Wang

TL;DR: Mise-en-Video 允许创作者先在 3D 空间安排多镜头舞台调度，再生成对应视频；层次注意力维持叙事一致性，3D 布局条件控制每镜头空间构图，并配套构建 16 万视频及方向 3D 框数据集。

Details

Motivation: 文本到视频模型已能生成高保真单镜头，但实际叙事创作还需跨镜头一致和细粒度控制。现有多镜头生成方法虽有较强镜头间一致性，创作者仍难精确控制场面调度。

Result: Mise-en-Video（MeV）让视频创作者在 3D 空间规划多镜头场面，再生成相应视频。框架结合多层叙事一致性层次注意力与 3D 感知布局引导，提供逐镜头空间控制。为训练和评测，作者建立标准化标注流程和 16 万视频数据集，含层次化多镜头文本及方向 3D 边界框。实验显示，MeV 在视频质量、镜头间一致性和主体级可控性上持续超过强基线。

Insight: 将多镜头生成分解为可显式规划的三维场景调度，可在保持叙事连续性的同时提升创作者对单镜头布局的控制。摘要未报告具体质量指标、训练数据分布或与逐镜头人工控制成本的比较。

Abstract: Recent advances in text-to-video generation have enabled high-fidelity single-shot synthesis, yet practical storytelling creation requires multi-shot videos with both strong inter-shot consistency and fine-grained controllability. Existing multi-shot video generation approaches achieve strong consistency but still fall short in controllability. In this paper, we propose Mise-en-Video (MeV), a highly controllable multi-shot video generation framework that allows video creators to conduct staging the multi-shot in the 3D space and then generate the corresponding multi-shot video. To achieve this, our framework combines a hierarchical attention mechanism for multi-level narrative consistency with 3D-aware layout guidance for precise, per-shot spatial control. To support training and evaluation, we develop standardized annotation pipelines and construct a 160K-video dataset with hierarchical multi-shot texts and oriented 3D bounding boxes. Extensive experiments demonstrate that MeV consistently outperforms strong baselines in video quality, inter-shot consistency, and subject-level controllability.

[231] What Language Adds, and at What Cost: Four Controlled Probes of Pathology Vision-Language Models cs.CV | q-bio.QMPDF

Wenhao Zhang, Zhongliang Zhou, Yi Ping, Zirui Wang, Sheng Li

TL;DR: 对病理视觉语言模型的四项受控探针显示，文本报告能在某些生存预测中提供窄范围增益，但提示词未带来可测分类精度提升，SlideChat 的问答在有无切片时准确率几乎相同，且改写问题可使 8–23% 回答翻转。

Details

Motivation: 病理 VLM 被认为可支持自然语言查询和报告草拟，但相较视觉骨干本身，语言究竟增加了多少价值尚不清楚；多数系统由开发者在自选测试集上评估，难以区分真正语言能力与数据、模型因素。

Result: 作者复用公开检查点和数据，通过一次只改变一个输入的四项探针检验语言贡献。跨 1,633 名 TCGA 患者，加入去除分期信息的冻结报告嵌入后，进展无病间期和总生存 pooled C-index 分别增加 0.019 和 0.016；恢复分期句没有额外增益。固定视觉特征和池化方式后，肺癌及肾癌亚型的零样本提示分类与监督线性探针统计上无差异。SlideChat 在带切片与不带切片时分别答对 74.82% 和 74.41% 的 SlideBench-VQA(TCGA) 项目，语言骨干单独为 43.52%；释义改写保持准确率，却使 8–23% 回答发生变化。论文认为现有证据还不足以把智能体导航和 few-shot 语言先验增益归因于语言。

Insight: 语言确有有限但狭窄的作用：报告文本可增加部分预后信号，但提示分类收益与切片问答 grounding 需独立验证。摘要建议未来评测采用控制输入的一次一变量对照；未给出四项探针之外的病理任务结论。

Abstract: Vision–language models (VLMs) promise to make computational pathology queryable and able to draft reports, but how much language adds beyond the vision backbone is unclear, because almost every system is evaluated by its developers on test sets of their own choosing. We grade published evidence by how it was produced and test four claims with controlled probes that reuse released checkpoints and public data, varying one input at a time. Report text adds prognostic signal: adding a frozen embedding of the stage-redacted report to a slide embedding raises the pooled C-index by 0.019 for progression-free interval and by 0.016 for overall survival across 1{,}633 TCGA patients, and restoring the stage statement adds nothing. Prompts add no measurable accuracy at inference: with features and pooling fixed, zero-shot classification is statistically indistinguishable from a supervised linear probe for lung and renal cancer subtyping. Slide-level question answering is barely grounded: SlideChat answers 74.82% of SlideBench-VQA(TCGA) items correctly with the slide and 74.41% without it, whereas its language backbone alone reaches 43.52%. Paraphrasing leaves accuracy unchanged but flips 8–23% of answers. Self-reported gains from agentic navigation and few-shot language priors cannot yet be attributed to language. Language thus adds real but narrow value, and we propose minimum controls that would expose these failure modes in future evaluations.

[232] State World Models: Execute the Rules, Generate the Pixels cs.CVPDF

Chan Yan Ki, Yanran Zhang, Wenzhao Zheng, Xiaofeng Wang, Zheng Zhu

TL;DR: Stateful Harness 将状态规则执行与视频像素渲染分离，以外部世界状态推进交互、由冻结视频模型呈现事件；在 50 对共享初始画面和动作但隐藏状态不同的反事实测试中，完整 harness 对两种情况都生成正确结局的比例达 92%。

Details

Motivation: 现有世界模型能从单张图像生成动作条件视频，常聚焦相机导航，但若动作结果取决于当前画面不可见的状态，单纯导航不保证遵守持续规则。

Result: Stateful Harness 在指定规则下维护演化的世界状态，以冻结视频模型渲染每个事件，并用生成片段更新下一轮状态。每轮先解析用户动作并将结果编译为指令；可编辑 3D 场景的预览片段提供空间和相机引导。作者构建 WorldState-Bench，包含 50 对反事实双生案例，每对共享初始画面和动作，但隐藏状态与正确结果不同。仅规则提示时，11 个视频后端中最佳者在 32% 案例对都产生所需结局；对 MiniMax-H3 明确指定结局后比例从 32% 升至 64%；完整 harness 在相同冻结模型下达 92%。

Insight: 隐藏状态规则应由显式状态机或 harness 执行，视频模型负责把执行结果呈现为像素；不能仅靠模型从当前画面推断不可见状态。测试集规模为 50 对案例，摘要未报告各规则类型分项或更大场景的扩展性。

Abstract: Existing world models can generate action-conditioned video from a single image. Many focus on camera movement, allowing users to explore the generated scene. However, navigation alone does not ensure that an action follows persistent rules when its outcome depends on state that is invisible in the current frame. We present Stateful Harness, a framework that separates world evolution from video rendering to support stateful and rule-governed interaction. It maintains an evolving world state under specified rules, renders each resulting event with a frozen video model, and uses the generated clip to update the world state for the next turn. At each turn, the harness resolves the action selected by the user and compiles the resulting event into an instruction, while a preview clip from an editable 3D scene provides spatial and camera guidance. To evaluate rule compliance, we build WorldState-Bench, a benchmark of 50 counterfactual twin pairs that share an initial frame and action but differ in hidden state and required outcome. Under rule-only prompting, the best of the 11 evaluated video backends produces the required endings for both cases in 32% of pairs. For MiniMax-H3, specifying the required endings raises this rate from 32% to 64%, while the full harness reaches 92% with the same frozen model. These results support separating rule execution from video rendering for world interactions with hidden state.

[233] PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models cs.CV | cs.ROPDF

Yude Zou, Yifei Yao, Guanbo Jiao, Jianming Ma, Guanjie Zheng

TL;DR: PICA-WAM 通过跨外观自对齐，让在线世界动作模型对增强后的视觉扰动保持动作表征不变：EMA 教师处理原始观察，在线模型将扰动观察的浅层动作表征对齐到教师深层目标；在四种 WAM 和仿真、真实机器人评测中提高鲁棒性。

Details

Motivation: 世界动作模型同时预测未来视频并生成动作，但仍可能对与任务无关的外观变化敏感。视频预测倾向保留外观信息，动作生成却未显式要求忽略这些变化；数据增强虽能展示扰动，却不强制原图与扰动图产生一致动作表征。

Result: PICA-WAM 将扰动不变性设计为辅助的跨外观自监督对齐任务。指数移动平均教师处理名义原始观察，输出深层目标表示；在线 WAM 处理受扰观察，并通过轻量投影头将较浅层动作表征对齐至该目标。四种强 WAM 在 RoboTwin 2.0、LIBERO-Plus、EBench 和真实机器人上评测，方法持续提高视觉扰动下鲁棒性、保持仿真名义表现并提高真实机器人成功率。逐层成对诊断还显示，视觉特征对无关外观更不敏感，其变化向动作传播的程度降低。

Insight: 只增加外观增强不够，需显式要求原始与扰动输入产生相容的动作表示；跨层对齐可在保留视频建模能力同时强化策略对任务无关外观变化的不变性。摘要未报告各基准提升值或扰动类别分项。

Abstract: World Action Models (WAMs) jointly predict future video and generate actions, yet remain sensitive to task-irrelevant visual perturbations, i.e., changes in visual appearance that should not affect robot behavior. We identify a gap in conventional WAM training: video prediction encourages representations to preserve appearance information, while action generation does not explicitly enforce invariance to task-irrelevant appearance. Data augmentation exposes WAMs to such perturbations but does not require perturbed and nominal observations to produce consistent action representations, and alignment with external pretrained features does not guarantee the desired invariance. To address this gap, we propose PICA-WAM, a self-supervised training framework that casts perturbation invariance as an auxiliary cross-appearance self-alignment task for WAMs. An exponential-moving-average (EMA) teacher processes the nominal observation to provide a deeper-layer target, while the online WAM processes a perturbed observation and aligns its shallower-layer action representation to this target through a lightweight projection head. Across four strong WAMs evaluated on RoboTwin 2.0, LIBERO-Plus, EBench, and a real robot, PICA-WAM consistently improves robustness under visual perturbations, preserves nominal performance in simulation, and raises the real robot success rate. PICA-WAM further reduces the sensitivity of video features to task-irrelevant appearance and its propagation into actions, as shown by paired layer-wise diagnostics. Code: https://github.com/YudeZou/PICA-WAM-release

[234] SpecKCast: Spectral Kalman Forecasting for Diffusion Acceleration cs.CVPDF

Javad Rajabi, Hakki Can Karaimer, David B. Lindell, Babak Taati, Michael S. Brown

TL;DR: SpecKCast 在频域预测扩散模型特征，用按频带噪声时间坐标运行的轻量 Kalman 滤波器跟踪每个 DCT 系数及变化率，并在预测不确定度超过阈值时才调用 Transformer；多个图像和视频模型上最高加速 6 倍，额外内存低于 0.05 GiB。

Details

Motivation: 扩散 Transformer 需要在多步去噪中反复前向计算。缓存可复用特征，预测方法则外推被跳过步骤的特征；但现有方法均匀外推所有特征，忽略不同空间频率在不同去噪阶段变化速度不同，以及去噪时间步并非均匀进展，导致跳步增多时预测误差扩大。

Result: SpecKCast 是免训练频域预测方法，对模型输出做二维离散余弦变换，并沿依据噪声调度构造的频带专属时间坐标，用轻量 Kalman 滤波器追踪每个系数的取值和变化率。只有当滤波器预测不确定性超过阈值时才执行 Transformer 计算。在 FLUX.1-dev、Qwen-Image、Wan2.1-14B 和 HunyuanVideo 上，SpecKCast 相较缓存和预测基线取得更好的加速—保真权衡，最高达到 6 倍加速，额外内存少于 0.05 GiB。

Insight: 去噪不同频带的特征变化速度与阶段不同，按噪声时间坐标逐频带预测比统一时间间隔外推更适合控制跳步误差。摘要未提供各模型对应的加速与保真数值。

Abstract: Diffusion Transformers (DiTs) achieve high-fidelity image and video generation, but they suffer from substantial inference costs due to repeated transformer evaluations across many denoising steps. Feature caching reduces this cost by reusing features across denoising steps. Recent forecasting methods extend this idea by extrapolating features to skipped steps instead of reusing them. However, existing forecasting methods extrapolate the features they predict uniformly, even though the spatial-frequency components of these features change most at different stages of denoising. They also overlook that denoising progresses nonuniformly, so equal step intervals can correspond to very different amounts of change. These mismatches amplify forecast errors as more steps are skipped, limiting the achievable acceleration. We propose SpecKCast, a training-free method that forecasts the model output in the frequency domain. SpecKCast applies a 2D discrete cosine transform (DCT) to the output and tracks each coefficient with a lightweight Kalman filter along a band-specific time coordinate derived from the noise schedule. The filter tracks each coefficient’s value and rate of change, and the transformer is evaluated only when the filter’s forecast uncertainty exceeds a threshold. Experiments on FLUX.1-dev, Qwen-Image, Wan2.1-14B, and HunyuanVideo show that SpecKCast achieves a better trade-off between acceleration and fidelity than existing caching and forecasting methods, reaching up to 6x speedup with less than 0.05 GiB of additional memory.

[235] GenOlympiad: Benchmarking Frontier Image Models at Their Capability Boundaries cs.CVPDF

Xi Hu, Martin Bell, Liang Zheng

TL;DR: GenOlympiad 用九类高难提示压力测试前沿文生图模型，并以可视化可核验的提示和自动分解评分框架检查能力边界；评测 20 多个模型后，最佳 Nano Banana Pro 严格得分仅 23.9/100。

Details

Motivation: 文生图模型已能生成高质量图像，但仍需明确前沿模型在哪些能力上失效。主观风格评价不易自动核实，论文选择数量、深度推理、物理一致性和复杂规则等可视觉验证的能力构造压力测试。

Result: GenOlympiad 包含九个能力维度的高难提示，设计使提示对齐可由图像直接验证，例如计数，而非依赖风格偏好。配套自动评估框架将提示拆为容易和困难的问题，并为多数难题提供证据依据以提高评分准确性。作者评测 20 多个强文生图模型；最佳模型 Nano Banana Pro 严格得分为 100 分中的 23.9。代码模型 GPT-6 Astra 的提示对齐分数更高，但图像观感不如最佳文生图模型。论文还分析模型演进、与 Elo 排名的一致性和 grounding 评估的效果。

Insight: 高保真并不意味着模型能遵守复杂、物理合理或数量精确的提示；用视觉可验证条件和困难项证据评分能暴露能力边界。摘要未提供九个维度分项或各模型完整排名。

Abstract: Text-to-image models have made tremendous progress and can produce high-quality images. However, a question which would be asked in any other discipline is where the frontier models fall short. We introduce GenOlympiad, a benchmark which provides the most challenging text prompts from nine capability axes for stress-testing frontier text-to-image models. The assessed capabilities include generating large quantities, performing deep reasoning, ensuring hard physical consistency, following complex rules and so on. The prompts are designed such that prompt alignment is visually verifiable (e.g., counting) instead of relying on subjectiveness (e.g., style). Based on this, we develop an automated evaluation framework which decomposes the prompt into easy and hard questions, and for most hard questions, draws help lines to ground evidence for accurate grading. We evaluate 20+ strong text-to-image models, including Nano Banana Pro, GPT-Image-2.5, Seedream-5.0 and so on. We report that even the best-performing T2I model (Nano Banana Pro) only achieves a strict score of 23.9 out of 100 and that coding models such as GPT-6 Astra have higher prompt alignment scores (but not as visually pleasing). We also derive insights in model evolution, alignment with Elo scores, and effectiveness of our grounding-based evaluation protocol. Code is available at https://github.com/GenOlympiad/GenOlympiad and the project page is at https://genolympiad.github.io/

[236] Manga2Animate: An Agentic Framework for Faithful Manga Animation with Script and Storyboard Planning cs.CVPDF

Sheng-Wei Zhang, Hsi-An Chen, Yan-Ting Liou, Chih-Hsuan Chen, Shih-Chen Tseng

TL;DR: Manga2Animate 以动画脚本和视觉分镜作为两个中间表示，并让生成代理与对照源漫画分格的审阅代理协作迭代；实验和用户研究显示，生成动画在源漫画忠实度与镜头间一致性上超过所比较先进方法。

Details

Motivation: 把现有漫画改编为连贯动画需要忠实保留源故事、人物和画风。只依据文字的现有方法会丢失图像内容，直接将漫画页送入视频模型则可能误解分格结构并生成错误剧情。

Result: Manga2Animate 采用多智能体框架，使生成每个阶段都以源漫画为依据，并把任务分成动画脚本和视觉分镜两个中间表示。每种表示由创作者代理生成，再由配对审阅代理对照漫画面板检查；方法结合表示专属选择和双层反馈，逐步细化两种表示后才合成最终动画。广泛实验和用户研究显示，该框架生成的动画较先进方法更忠实于原漫画，镜头间也更一致。

Insight: 在最终视频生成前显式规划脚本和分镜，并在每步回看源漫画，可减少模型误读漫画版式和剧情结构。摘要未给出具体用户研究规模、量化指标或生成成本。

Abstract: Adapting existing manga into coherent animation that stays faithful to the source is a valuable yet challenging task. Existing methods either access the manga mainly through text, losing its characters and art style, or feed manga pages directly into video generation models, which misinterpret the manga’s structure and produce incorrect storylines. In this paper, we propose Manga2Animate, a multi-agent framework that grounds every stage of generation in the source manga. We decompose the task into two intermediate representations, an animation script and a visual storyboard. Each representation is produced by a creator agent and reviewed against the source panels by a paired reviewer agent. We tailor representation-specific selection and dual-level feedback to progressively refine both before synthesizing the final animation. Extensive experiments and a user study demonstrate that Manga2Animate produces animations more faithful to the source manga and more consistent across shots than state-of-the-art approaches. Project page: https://annchenn.github.io/Manga2Animate/

[237] InfiniDream: Open-Ended Photorealistic 3D Worlds Through Procedural Generation cs.CVPDF

Yi-Chuan Huang, Hao-Jen Chien, Zheng-Hui Huang, Yu-Lun Liu, Kaipeng Zhang

TL;DR: InfiniDream 用 VLM 驱动开放式程序化 3D 世界生成：符号约束资产模块调用现有程序算法创造可编辑且支持随机种子的写实资产，再通过语义场景规划和确定性视锥网格布置生成大型场景；100 个物体及 20 个场景任务的四项感知指标均改善。

Details

Motivation: 程序化生成比静态网格更可控、易编辑且多样，但 VLM 代码生成常只得到简单 3D 资产；复杂写实程序系统需大量人工开发。既有 VLM 引导方法又受预定义工厂和 API 限制，难以创造新的资产类型。

Result: InfiniDream 是 VLM 驱动的开放式程序化世界生成框架。Symbolically Constrained Asset Generation 利用现有程序算法生成新颖写实资产，同时保留编辑性和基于随机种子的多样性；端到端文本到世界流程将语义场景规划与确定性视锥引导网格布置结合，实现大规模场景中精确物体放置、程序化填充和动画。在 100 个物体与 20 个场景生成任务上的评估显示，四项感知指标均有改善；多样性分析、消融和人类偏好结果进一步支持该方法。

Insight: 结合可复用程序算法与符号约束，可能突破固定资产工厂的开放性限制，同时保留生成场景的编辑能力。摘要未说明四项指标名称、改善幅度或与人工程序建模的成本差别。

Abstract: Procedural generation plays a key role in synthetic data creation and game development, offering greater controllability, editability, and diversity than static mesh representations. However, VLM-based code generation often produces relatively simple 3D assets, while sophisticated photorealistic procedural systems require extensive manual development. Existing VLM-guided procedural methods remain largely constrained by predefined factories and APIs, limiting their ability to create novel asset types. We introduce InfiniDream, a VLM-driven framework for open-ended procedural 3D world generation. First, we propose Symbolically Constrained Asset Generation, which leverages existing procedural algorithms to create novel, photorealistic assets while preserving editability and seed-based diversity. Second, we develop an end-to-end text-to-world pipeline that combines semantic scene planning with Deterministic Frustum-Guided Grid Placement to construct large-scale 3D worlds with precise object placement, procedural population, and animation. Extensive evaluations on 100 object and 20 scene generation tasks demonstrate improvements across four perceptual metrics. Diversity analysis, ablation studies, and human preference evaluations further validate the effectiveness of our framework. Project page: https://yichuanh.github.io/infinidream-page/

[238] Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience cs.CV | cs.ROPDF

Tianqi Liu, Hanfei Zhao, Qifeng Liang, Hui En Pang, Jiarui Xu

TL;DR: Ropedia W0 是跨人类与多种机器人形态的第一视角运动—操作世界基础模型，以 10,000 小时人类和 3,000 小时机器人数据训练，并用统一骨骼表示动作；在五个跨形态基准和留出人类视频上达到先进表现，蒸馏后可实时生成分钟级 rollout。

Details

Motivation: 世界模型可让具身智能体预测动作后果，但要在不同形态间联合建模行走和操作，面临动作空间异质、跨模态监督有限的问题。

Result: Ropedia W0 构建含 10,000 小时第一视角人类数据和 3,000 小时机器人数据的 Xp-Data，配有统一多模态标注与质量控制。统一运动学表示以相机位姿编码行走，以投影骨架编码操作；机器人骨架结合形态专属正向运动学和大量人工关键点标注校正。作者还建立基于高质量灵巧人形远程操作数据的长时基准 Xp-Teleop。评估涵盖 DreamDojo、Ψ0、ω-HOME、Xp-Teleop、AgiBot 五个跨形态基准和留出人类第一视角视频，涵盖零样本、少样本及全数据后训练，模型均达先进结果；扩大人类预训练数据可同时改善人类和机器人评估。蒸馏后支持实时、分钟级 rollout 和 VR 交互。

Insight: 人类第一视角经验可与机器人轨迹共享统一运动学表示，为跨形态预测和控制提供大规模监督；人类数据扩展也改善机器人任务。摘要未列五个基准的具体分数或各数据来源贡献。

Abstract: World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $Ψ_0$, $ω$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model’s utility for embodied decision-making.

[239] Reinforcement Learning with Visually Attributable On-Policy Self-Distillation cs.CV | cs.AIPDF

Shiao Xie, Siyu Chen, Jianwei Lv, Xiandong Li, Yujin Wang

TL;DR: VAG 用验证器优势作优化锚点，并比较特权教师在有视觉证据与证据退化视图下的修正方向，只保留与视觉证据一致的 on-policy 自蒸馏更新；六个视觉推理基准上持续超过 GRPO 和普通 OPSD，达到论文报告的平均最佳水平。

Details

Motivation: 带可验证奖励的多模态强化学习能提供可靠任务级监督，但难定位视觉推理错误；特权教师可给出密集 token 修正，然而修正可能混入语言先验、回答风格或教师视图差异，而不是真正由视觉证据支持。

Result: 论文提出视觉可归因 on-policy 自蒸馏范式，以任务级 RLVR 结合基于视觉证据的特权 token 修正，并将无法稳定归因到证据的修正定义为 privileged-correction contamination。实例化方法 Visual-Attribution Gating（VAG）以验证器导出的优势作为优化锚点，将教师在证据充分与证据退化视图中的修正进行对比，只接纳方向上受视觉差异支持的更新，抑制不一致残差。六个视觉推理基准实验显示，VAG 一贯超过 GRPO 与 OPSD 方法，并达到论文报告的新平均最佳性能。

Insight: 教师提供密集监督并不保证它由图像证据驱动；通过证据充分/退化对照验证修正方向，可减少视觉推理中的非视觉偏差。摘要未给出各基准准确率、增幅或门控阈值。

Abstract: Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections from a privileged teacher equipped with an evidence-centered visual view. However, the resulting corrections may conflate visual evidence with non-visual shifts induced by linguistic priors, response style, or the privileged view itself. We formalize this phenomenon as privileged-correction contamination: a token-level correction is contaminated when its update direction cannot be consistently attributed to the privileged visual evidence. Directly distilling contaminated corrections may therefore reinforce erroneous token-level updates. To address this issue, we propose reinforcement learning with visually attributable on-policy self-distillation, an evidence-consistent paradigm that augments task-level RLVR supervision with privileged token-level corrections grounded in visual evidence. We instantiate this paradigm with Visual-Attribution Gating (VAG), which uses the verifier-derived advantage as the optimization anchor and validates each privileged correction against the contrast between evidence-present and evidence-degraded teacher views. By admitting only corrections directionally supported by this visual contrast, VAG preserves visually grounded guidance while suppressing inconsistent residuals. Experiments across six visual reasoning benchmarks show that VAG consistently outperforms GRPO and OPSD methods, establishing new state-of-the-art average performance.

[240] VideoMaMa++: Temporally Consistent Video Matting via Preserve-and-Refine Token cs.CVPDF

Sangbeom Lim, Seoung Wug Oh, Heeji Yoon, Seungryong Kim, Joon-Young Lee

TL;DR: VideoMaMa++ 将 RGB 和粗掩码分成两个 token 流并用时空自注意力融合，再以 preserve/refine 可学习 token 跨短块传递已生成 matte；无需额外训练即可减少长视频分块边界闪烁，并在基准及真实长片段上提升细节和时间一致性。

Details

Motivation: 扩散式视频抠图依赖大规模预训练视频先验，合成数据上可零样本泛化，但计算量使推理只能处理 6–24 帧短块。常见方法独立处理每个块、只依赖粗掩码引导，导致块边界处边缘闪烁和透明度不连续。

Result: 该框架将 RGB 与掩码条件编码为独立 token 流，并通过完整时空自注意力交互，使模型能自适应融合细节纹理与粗略布局，而不使用固定通道融合。可学习的 preserve 和 refine token 提供逐帧条件接口，支持滑动窗口推理：已生成 matte 跨块传播以保持局部时间一致性，每块再共享全局参考锚点以避免长序列误差累积，无需额外训练。在标准视频抠图基准和真实长视频上，方法较既有扩散方法提升细节精度与时间一致性，并能维持长视频块边界的一致性。

Insight: 长视频分块处理需要显式在块间传递抠图状态，同时用全局参考防止局部误差累积；仅逐块独立推理难保持透明边界稳定。摘要未报告基准数值或滑窗推理额外成本。

Abstract: Diffusion-based video matting has recently emerged as a leading approach, leveraging video diffusion priors from internet-scale pretraining to achieve strong zero-shot generalization from limited synthetic data. However, the heavy compute of these diffusion backbones limits inference to a short chunk of just 6 to 24 frames, so even ordinary videos span multiple chunks. Because each chunk is matted independently from coarse mask guidance, existing methods fail to maintain consistency across chunk boundaries, producing flickering edges and discontinuous transparency. We address this with a mask-guided diffusion-matting framework built on two ideas. The RGB and mask conditioning are represented as separate token streams that interact through full spatio-temporal self-attention, letting the model adaptively combine fine RGB texture with coarse mask layout rather than committing to a fixed channel-wise fusion. On top of this, we propose a novel pair of learnable embeddings, the preserve token and the refine token, that act as a per-frame conditioning interface and enable a sliding-window inference scheme in which already-generated mattes are propagated across chunk boundaries for local temporal consistency, complemented by a globally shared reference anchor at every chunk to prevent error accumulation in long sequences without any additional training. Across standard video matting benchmarks and in-the-wild long footage, our framework outperforms prior diffusion-based methods in both fine-detail accuracy and temporal consistency, and uniquely maintains consistency across chunk boundaries on long videos.

[241] Can It Actually Work? Synthesizing Functional CAD Mechanisms cs.CV | cs.CL | cs.GRPDF

Wenqian Zhang, Yangyi Huang, Zigeng Xu, Weiyang Liu, Zhen Liu

TL;DR: FunCAD-Harness 将抽象物理功能要求转成针对具体候选 CAD 机构的模拟测试，并据反馈指导设计修订；FunCADBench 覆盖 46 个含相互依赖、多步操作的机构家族，前沿 LLM 智能体实验显示功能成功率超过标准仿真增强基线。

Details

Motivation: LLM 智能体能生成视觉上合理的 CAD 装配，却常无法实现预期物理功能。候选机构可能用完全不同部件、控制方式和操作顺序实现同一功能，因此验证抽象需求并不简单。

Result: FunCAD-Harness 将抽象功能要求翻译成候选专属的物理测试，再用模拟反馈指导设计迭代。作者构建 FunCADBench，涵盖 46 个机构家族及相互依赖、多步操作。前沿 LLM 智能体实验显示，FunCAD-Harness 能发现视觉检查难以察觉的机械故障，功能成功率显著高于标准仿真增强基线。

Insight: CAD 评估不能只检查装配外观或固定组件结构，应将用户功能要求转成每个候选机构可执行的物理测试。摘要未给出成功率数值、仿真精度或与人工机械工程设计的对比。

Abstract: While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions. We study functional CAD design, a setting where agents interpret underspecified requests to construct and iteratively revise mechanisms that satisfy functional requirements in physical simulation. Verifying these requirements is nontrivial: different candidate mechanisms may realize the same function through entirely different components, controls, and operating sequences. To bridge this gap, we propose FunCAD-Harness, a framework that translates abstract requirements into candidate-specific physical tests and uses simulation feedback to guide design revisions. To systematically evaluate this capability, we introduce FunCADBench, a benchmark spanning 46 mechanism families with interdependent and multi-step operations. Experiments across frontier LLM agents show that FunCAD-Harness exposes subtle mechanical failures invisible to visual inspection, substantially improving functional success rates over standard simulation-augmented baselines. Ultimately, our work marks a step toward CAD agents capable of translating human intent into working mechanisms.

cs.AI [Back]

[242] Beyond Explicit Intent: Discovering and Shaping Artistic Intent through Human-AI Co-Creation cs.AI | cs.CV | cs.HCPDF

Renye Yan, Jialun Zhong, Jikang Cheng, Yi Gao, Yanzeng Li

TL;DR: 论文提出通过扩大生成探索、让艺术家比较并挑选视觉结果来发现原本模糊或未表达的创作意图，再把对生成轨迹的策展偏好以不同强度用于强化学习，形成持续探索—识别—塑造循环；三项用户研究分别考察偏好发现、个性化对齐和模糊意图外化。

Details

Motivation: 生成式 AI 正成为艺术创作媒介，但现有交互通常假定创作者事先能明确表达意图并用提示词准确传达。真实艺术意图可能模糊、潜在，或只有看到具体视觉可能性后才被识别。

Result: 该人机共创范式先扩大模型探索空间，以多样视觉结果让艺术家通过观察、比较和选择发现未曾言明的偏好；随后把艺术家的策展判断融入强化学习，根据对不同生成轨迹的偏好分配不同学习强度，强化更被认可的方向，使模型逐渐适应个人审美。流程形成“探索—识别—塑造—再探索”的循环，让意图逐步被发现、外化并塑造。三项用户研究分别考察潜在偏好发现、个性化偏好对齐，以及难以言说的创作意图外化。

Insight: 创作提示未必能在生成前完整表达意图；让用户先比较具体结果，再把选择反馈给模型，能把生成过程变成意图探索工具。摘要未报告用户研究人数、偏好对齐效果数值或长期创作结果。

Abstract: Generative AI is becoming an important medium for artistic creation, yet existing interactions often assume that creators can specify intentions in advance and accurately communicate them through text prompts. Artistic intentions, however, may be vague, latent, or recognizable only after seeing concrete visual possibilities. We propose a human-AI co-creation paradigm beyond explicit intent expression. By expanding model exploration, we expose artists to diverse visual possibilities, allowing them to identify previously unarticulated preferences through observation, comparison, and selection. We then integrate artists’ curatorial judgments into reinforcement learning, assigning different learning strengths based on preferences over generation trajectories, so recognized directions are more strongly reinforced and the model gradually adapts to individual aesthetics. This creates a continuous “explore-identify-shape-re-explore” loop in which artistic intent is progressively discovered, externalized, and shaped. Through three user studies, we examine latent-preference discovery, personalized preference alignment, and the externalization of vague, difficult-to-express creative intentions.

[243] LogicVal:A Logic-Formalized Benchmark for Constraint Satisfaction in Text-to-Image Generation cs.AI | cs.CL | cs.LOPDF

Savitha Sam Abraham, Marjan Alirezaie, Hossein Rahnama

TL;DR: LogicVal 将文生图评测表述为逻辑约束满足问题，并结合神经视觉检测与逻辑推理判断生成图像是否符合提示；实验揭示多对象、多约束组合远难于单对象属性，LogicVal 在 CLEVR 上与人类判断的 ROC-AUC 为 0.88–0.98。

Details

Motivation: 文生图模型虽已能生成高保真图像，但属性绑定、计数、空间和常识推理等能力需要独立评估。已有基准通常从单项能力设计，论文提出把对象、属性、关系及其逻辑条件统一形式化，以衡量图像是否满足完整提示约束。

Result: LogicVal 将基准对象、属性、关系和约束写成逻辑规格，并用神经视觉检测加逻辑推理评估生成图像。作者评估四种先进文生图模型、九类从单对象属性到含全称、存在、计数及蕴含关系的多对象约束。表现最佳的模型在单对象类别准确率为 98%，复杂条件多对象关系仅 53%；组合泛化中，单约束提示准确率 84%，多约束同时满足时降至 20%；不可满足约束识别率仅 23%。LogicVal 在 CLEVR 上的 ROC-AUC 为 0.88–0.98，与人类判断较一致。

Insight: 单项属性生成成功不代表模型能组合多个对象关系或识别逻辑上不可满足的提示；评测应显式描述约束并单独考察复杂组合。CLEVR 上与人类判断一致并不自动证明其覆盖真实图像场景，摘要未报告其他视觉数据集结果。

Abstract: Text-to-Image (T2I) generative models have made significant progress in producing high-fidelity images from natural-language prompts. Alongside this progress, numerous benchmarks have been developed to evaluate capabilities such as attribute binding, numeracy, spatial reasoning, and commonsense reasoning. In this work, we propose a new perspective on T2I evaluation by casting it as a constraint-satisfaction problem. We introduce a logical framework that formally specifies the benchmark domain, including objects, attributes, and relations, alongside the constraints that generated images must satisfy. We evaluated four state-of-the-art T2I models across 9 classes of constraints ranging from a single-object-attribute to multi-object relationships constraints involving logical operators ($\forall, \exists, count, \rightarrow$). Our results reveal a substantial gap between single-object and complex conditional multi-object relationship constraint classes, with the best-performing model achieving 98% accuracy on the former and just 53% on the latter. We assess compositional generalisation, finding that the best-performing model`s accuracy drops from 84% on single-constraint prompts to 20% when satisfying multiple constraints simultaneously. Identifying unsatisfiable constraints is also challenging, with the best model correctly recognizing only 23% of such prompts. Building on this framework, we introduce \textbf{LogicVal}, an evaluation methodology that combines neural visual detection with logical reasoning to assess constraint satisfaction in generated images. On CLEVR, LogicVal achieves ROC-AUC scores ranging from 0.88 to 0.98, indicating strong alignment with human judgments. These findings underscore the value of logic-based evaluation for characterizing T2I model capabilities and limitations.

[244] Cache2LoRA: Understanding and Improving Context Internalization cs.AI | cs.CL | cs.IRPDF

Yi Fang, Gaurav Srivastava, Zhenyu Bi, Sirui Ding, Xuan Wang

TL;DR: Cache2LoRA 先从可复用的冻结文档缓存中按查询聚合证据，再生成 LoRA 更新；在 17 项任务上宏平均 F1 为 0.663，接近完整上下文 ICL 的 0.671，并在 32K 文档条件下取得其 5.9 倍解码吞吐。

Details

Motivation: Context-to-LoRA 方法将文档编码为低秩参数更新，使模型回答时不必把原文放入解码上下文，但事实保留、来源区分以及按不同指令组织回答的可靠性仍不清楚。论文通过控制诊断发现竞争来源会影响答案，事实恢复也不等于遵循回答指令。

Result: Cache2LoRA 针对每个查询，从可复用的冻结文档缓存中聚合证据后生成 LoRA 更新；可训练编码只处理查询和记忆 token 后缀，并以多风格监督同时训练证据保留和答案表达。作者在 17 项任务上评估，宏平均 F1 为 0.663，接近完整上下文 in-context learning 的 0.671。对 32K 文档、256 token 输出、batch size 16，解码吞吐为完整上下文 ICL 的 5.9 倍。在相同记忆大小下，遇到长且多样的中间文本时，生成的 LoRA 比 learned-KV 变体更可靠地保留答题准确率；与保留文档 KV 组合时，混合 F1 和 EM 更高。

Insight: 生成参数更新前按问题从共享缓存中筛证据，可以同时减少长上下文解码负担并改善事实与回答风格适配。吞吐结果对应所列文档长度、输出长度和 batch 设置；摘要未说明其他资源成本或在线构建 LoRA 的延迟。

Abstract: Context-to-LoRA methods encode documents as low-rank parameter updates, allowing language models to answer without retaining the source text in their decoding context. How reliably do these updates preserve facts and support their use under different instructions? Controlled diagnosis of existing context-to-LoRA methods separates fact discrimination, target retention, and response adaptation. Competing sources influence answers, while target-only construction and memory-state optimization improve recovery at fixed update size. Recovering facts also does not ensure that answers follow response instructions. We introduce Cache2LoRA, which uses each query to aggregate evidence from a reusable frozen document cache before generating LoRA updates. Trainable encoding operates only on the query and memory-token suffix, and multi-style supervision teaches both evidence preservation and answer expression. Across 17 tasks, Cache2LoRA improves on existing context-to-LoRA methods and achieves macro F1 of 0.663, approaching full-context in-context learning at 0.671. For 32K documents and 256-token outputs at batch size 16, its decoding throughput is 5.9$\times$ that of full-context ICL. At matched memory size, generated LoRA preserves answer accuracy more reliably than the learned-KV variant under long, diverse intervening text and combines more effectively with retained document KV, yielding higher hybrid F1 and EM.

[245] Joint, Decoupled, or Agent: When Each Multimodal Alignment Paradigm Wins cs.AI | cs.CL | cs.CV | cs.MMPDF

Yi Fang, Gaurav Srivastava, Sirui Ding, Xuan Wang

TL;DR: 论文比较联合预训练融合、独立表征融合和 LLM 智能体三种多模态范式，发现没有一种在 144 个数据集—监督预算组合中通吃；利用表示、任务和预算诊断进行两阶段选择，留一组评估下选中最佳范式的比例从元数据路由的 54.2% 提升到 71.4%。

Details

Motivation: 多模态信息可以通过联合预训练编码器、下游融合独立编码器，或在推理时让语言模型作为智能体整合。论文希望厘清不同任务和标注预算下何时适合哪种范式，而非假设单一架构普遍占优。

Result: 实验以公开检查点实例化三种方法：JOINT 冻结联合预训练编码器并训练融合头；DECOUPLED 对独立预训练编码器训练融合头；AGENT 向冻结 LLM 提供任务指令、标注示例、可用样本文本和单模态监督读数。跨五领域 16 组配对数据集和多个训练标注预算，共 144 个组合中没有范式全面占优；按平均效用计，DECOUPLED、JOINT、AGENT 分别赢得 54、44、46 个组合。留一组评估的两阶段路由器先选智能体或融合，再选 JOINT 或 DECOUPLED，选择胜者比例 71.4%，高于仅用元数据的 54.2%；平均相对遗憾为 0.028，对比 0.057。表示几何、监督表示诊断和文本—标签匹配共同使用时优于单一诊断源。

Insight: 多模态范式的相对优势受任务形式、标注预算和单模态表示质量共同影响；任务诊断可以在训练候选融合模型前帮助选择流程。结果来自所测试的 16 组数据集和路由方案，不能据此断言对所有新任务均能正确选型。

Abstract: Jointly pretrained multimodal encoders are a common way to combine modalities, while alternatives integrate their information during downstream training or inference. When does each approach win? We compare three paradigms, instantiated with public checkpoints, across 16 paired datasets in five domains and multiple training-label budgets. JOINT trains a fusion head on frozen, jointly pretrained encoders. DECOUPLED trains a fusion head on frozen, independently pretrained encoders. AGENT gives a frozen LLM task instructions, labelled examples, available sample text, and supervised unimodal readouts. No paradigm dominates the complete 144 dataset-budget combinations: DECOUPLED wins 54, JOINT 44, and AGENT 46 by mean utility. ScienceQA, MELD, and MetMeme shift from AGENT to fusion as supervision increases. Between the fusion paradigms, JOINT leads on four of five QA datasets across all evaluated budgets, while higher unimodal 5-NN accuracy is associated with a relative shift toward DECOUPLED across eight classification datasets (Spearman $ρ=-0.67$). We test whether representation and task diagnostics predict these preferences on held-out datasets before fitting their candidate systems. A two-step cascade combines these diagnostics with task and budget metadata to select AGENT or fusion, then JOINT or DECOUPLED. Under leave-one-group-out evaluation, it selects the winner in 71.4% of cells versus 54.2% for metadata-only routing, with mean relative regret 0.028 versus 0.057. Using geometry, supervised representation diagnostics, and text-label matching together yields higher accuracy and lower regret than using any one of these sources with metadata. These results connect paradigm preferences to supervision budget, task format, and unimodal 5-NN accuracy. Target-task diagnostics thus support selecting among the three paradigms before candidate fitting.

[246] Self-Improving Video-to-Action Models cs.AI | cs.CVPDF

Hyunwoo Ryu, Eric Ming Chen, Max Wilcoxson, Vincent Sitzmann

TL;DR: Mirror 通过在环境中尝试跟随视频规划并收集由自身猜测动作产生的视频—动作配对，迭代自监督学习视频到动作映射，无需预先动作数据或手工奖励；在 Atari、OGBench 和 DeepMind Control 上超过使用学习视频计划的观察学习方法。

Details

Motivation: 生成式视频模型可作为具身智能体的视觉规划器，为新任务生成完成过程，但执行计划通常还需依赖大规模人工视频—动作数据训练逆动力学模型，限制了扩展性。

Result: Mirror 在环境交互中自监督学习视频到动作映射：智能体尝试用当前最佳动作猜测执行视觉计划，把交互得到的视频—动作对加入训练，逐步提升后续跟随计划的能力。Atari、OGBench 和 DeepMind Control 实验中，Mirror 超过使用学习得到视频计划的领先观察学习方法；使用 oracle 视频计划时接近专家表现。Minecraft BASALT 上，它超过依赖专家动作或手工任务知识的方法；生存模式中，无奖励和专家动作标签首次通过自我改进取得铁镐。离线设置下有专家动作标签但无环境交互时，它在获取钻石任务上达到新最佳结果。

Insight: 环境交互可把模型自己的动作尝试转化为逆动力学训练信号，使视频规划能力提升不必伴随收集大量专家动作或设计奖励。离线钻石结果仍使用专家动作标签，和在线纯自改进设置不同。

Abstract: Video generative models are an emerging tool for embodied agents: they can serve as scalable visual planners, generating videos that depict how to complete previously unseen tasks. However, executing these plans typically requires a video-to-action or “inverse dynamics” model trained on large, human-collected video-action datasets, severely limiting scalability. We introduce Mirror, a self-supervised approach that learns video-to-action mappings through environment interactions, without prior action data or hand-crafted rewards. Mirror learns by attempting to follow a visual plan by executing its best guess of the underlying actions in the environment. Mirror then trains on the resulting video-action pairs, incrementally self-improving and following the visual planner more closely. Across Atari, OGBench, and DeepMind Control, Mirror outperforms leading learning-from-observation methods with learned video plans, and achieves near-expert performance with oracle plans. On Minecraft BASALT, it outperforms methods with access to expert actions or hand-engineered task knowledge. In Minecraft survival, it is the first method to acquire an iron pickaxe through self-improvement alone, without rewards or expert action labels. In the offline setting, with expert action labels but no environment interaction, it sets a new state of the art in obtaining diamonds. Our results suggest that advances in video models can translate into increasingly capable agents without a corresponding need for hand-crafted rewards or large collections of expert action data.

[247] Probing Ambiguity: Adversarial Intent Resolution for Robust Text-to-SQL cs.AI | cs.CLPDF

Youbing Yin, Xujun Peng, Muhammad S. Uddin, Erin Babinsky, Daben Liu

TL;DR: AIR 用多智能体流程主动解析模糊 Text-to-SQL 查询：生成多种意图解释、通过 SQL 执行确认结果确实不同，再针对数据差异提出澄清问题；在 AMBROSIA 的 schema-only 设置中，F1 为 0.679，高于 IntentRL 的 0.618，缺少完整数据库时性能下降约 7% 而对照下降约 20%。

Details

Motivation: Text-to-SQL 系统通常假定自然语言查询只有一个正确 SQL，但真实问题可能存在多种合理解释并产生不同结果。歧义若未识别，系统直接给单一 SQL 会把不确定意图伪装成确定答案；获取完整数据库供模型阅读在实际部署中也可能不可行。

Result: Adversarial Intent Resolution（AIR）无需任务专属训练，先构造竞争意图解释，通过执行 SQL 验证它们是否真的产生不同结果，再依据观察到的数据差异生成有针对性的澄清问题。在 AMBROSIA 的 schema-only 设置（不暴露完整数据库）下，AIR 在所有指标上超过强化学习方法 IntentRL，例如 F1 为 0.679 对 0.618；缺少完整数据库时，AIR 性能仅下降约 7%，IntentRL 下降约 20%。在歧义来自 schema 设计而非自然语言结构的 AmbiQT 上，AIR 在两种上下文设置下都获得更高 F1，且未针对基准适配。

Insight: 澄清应建立在可执行结果差异上，避免对并不会改变答案的表面歧义过度追问；schema-only 鲁棒性也说明不必依赖完整数据库暴露。摘要未提供所有指标的数值或澄清交互轮数。

Abstract: Text-to-SQL systems typically assume a one-to-one mapping between natural-language queries and correct SQL translations. In practice, however, many queries are inherently ambiguous, admitting multiple valid interpretations that yield distinct outputs. We introduce Adversarial Intent Resolution (AIR), a multi-agent framework that actively probes and resolves such ambiguity without any task-specific training. AIR first constructs competing interpretations, verifies through SQL execution that they produce genuinely different results, and generates targeted clarification questions grounded in the observed data differences. Evaluating on leading benchmarks for ambiguous questions, AIR demonstrates strong performance, efficiency, and robust cross-benchmark transferability. On the AMBROSIA benchmark under schema-only context, a realistic production setting where exposing the full database is infeasible, AIR outperforms IntentRL, the state-of-the-art reinforcement learning model, across all metrics (e.g., F1 0.679 vs. 0.618). Crucially, AIR exhibits only ~7% performance degradation without the full database, compared to ~20% for IntentRL. On the AmbiQT benchmark, where ambiguity arises from schema design rather than linguistic structure, AIR achieves higher F1 under both context settings without any benchmark-specific adaptation.

[248] MSCR: Jointly Balancing Utilization Across Modalities and Discovering Synergistic Information cs.AI | cs.CV | cs.LGPDF

Xinyu Chen, Liangjian Wen, Jiang Duan, Dongkai Wang, Yong Dai

TL;DR: MSCR 在多模态融合中同时抑制单模态过度主导并奖励融合后的协同预测，还用自适应门控将正则集中于需要协作的样本；五个基准中持续超过 MCR，并提高单模态误分类样本上的条件恢复率和 InterSHAP 所归因的跨模态交互占比。

Details

Motivation: 多模态训练中的模态不平衡会使某一模态主导优化，阻碍模型学习跨模态交互。现有动态平衡方法可调整模态贡献，却通常不考虑不平衡如何影响融合中协同信息的形成，而协同信息是单模态证据无法恢复的多模态增益来源。

Result: Multimodal Synergistic Competition Regularizer（MSCR）结合两个目标：反主导目标抑制过强单模态影响，为协同信息留下空间；协同目标鼓励融合预测优于非协同的单模态参考。自适应门控进一步将正则集中到更需要协作建模的样本。五个基准上，MSCR 持续超过 Multimodal Competition Regularizer（MCR），并强化协同能力：单模态预测错误样本上的融合正确率（conditional recovery）提升，InterSHAP 归因给跨模态交互的输出比例也提高。使用 Transformer 融合模块时条件恢复增益仍保持。

Insight: 平衡各模态使用只是手段，最终还需验证融合是否真正挖掘单一模态无法提供的协同信息。摘要未给出五个基准分数或正则权重敏感性。

Abstract: In multimodal learning, modality imbalance often lets one modality dominate optimization and keeps the model from fully exploiting cross-modal interactions. Balance-oriented methods such as the Multimodal Competition Regularizer (MCR) perform well by alleviating this imbalance during training. These dynamic balancing methods, however, mainly adjust modality contributions and do not consider how imbalance affects the formation of synergistic information during fusion. Synergy is a major source of multimodal advantage, because it is predictive gain that cannot be recovered from unimodal evidence alone. We propose the Multimodal Synergistic Competition Regularizer (MSCR), a fusion method that jointly promotes balanced modality utilization and the discovery of synergistic information. MSCR has two complementary objectives. An anti-dominance objective suppresses excessive unimodal influence and leaves room for synergistic information, and a synergy objective encourages the fused predictor to outperform a non-synergistic unimodal reference. An adaptive gate further concentrates the regularization on samples that need stronger collaborative modeling. Experiments on five benchmarks show that MSCR consistently improves performance over MCR and strengthens synergistic capability. It raises conditional recovery, the accuracy of the fused prediction on samples that the unimodal predictors misclassify, as well as the share of the model’s output that InterSHAP attributes to cross-modal interactions. Its gain in conditional recovery persists when the fusion module is Transformer-based. Code is available at https://github.com/chenxinyu991-collab/MSCR.

[249] Axiomatic World Modeling for Physics Reasoning cs.AI | cs.CLPDF

Xinye Yang, Zhenyang Liu, Yuxuan Wang, Yuanyuan Lei

TL;DR: RLAWM 把物理推理失败定位为前提层世界模型偏离题目所给物理公理，并分别奖励模型是否正确建构世界和是否在其中正确推导；无需增加推理期成本，在 PhysReason、PHYSICS 和奥赛级测试上均提升表现。

Details

Motivation: 近期高难物理推理基准显示，模型错误有时不是缺知识或不会运算，而是擅自改动已知量、编造定律或加入题目未给的假设，再在错误世界中推理。现有可验证奖励或过程奖励并未直接监督物理等公理化领域中的世界建模。

Result: 作者将模型内部公理世界模型与题目规定物理世界的差异定义为 modeling drift，并提出仅在训练期运行的 RLAWM。它在建模和求解两个层级探测推理并提供物理约束奖励：建模奖励检查计算前是否正确抽象世界公理，求解奖励检查具体推导是否与这些公理一致。相较基线，RLAWM 在 PhysReason 上最高提升 12.2 分、在 PHYSICS 上提升 5.2 分；对难度、来源和领域可零样本泛化。在奥赛级物理与数学上，相比最强基线取得 2.45 倍增益和 10 分提升，推理阶段无额外成本。

Insight: 物理推理评估和训练应分别检查模型相信自己处于什么世界、以及它如何在该世界中演算，避免只奖励最终数值而忽略错误前提。摘要未给出各数据集指标定义及绝对分数。

Abstract: Recent research-level physics-reasoning benchmarks for LLMs reveal a failure mode that lies not in knowledge or execution, but in modeling the world being reasoned about. Models often silently alter given quantities, fabricate laws, or inject unstated assumptions to force a problem into familiar patterns, and then reason inside this corrupted world. We define this premise-level divergence as modeling drift: the gap between a model’s internal axiomatic world model (AWM) and the true physical world specified in the problem. Although RLVR and PRM have yielded substantial gains in mathematics and code, their reward shapes do not directly supervise world modeling in axiomatic domains such as physics. We propose Reinforcement Learning with Axiomatic World Modeling (RLAWM), which optimizes the policy by penalizing divergence between its inferred axiomatic world model and the true physics world. RLAWM probes physics reasoning at two hierarchical levels (modeling and solving), each scored by a physically grounded reward. The modeling reward targets abstract knowledge formulation, evaluating whether the model captures the true axiomatic world before attempting calculation. The solving reward targets instantiated knowledge reasoning, checking whether the model’s concrete derivation remains consistent with the axiomatic world. This two-level structure separates what world the model believes it is reasoning in from how it reasons within that world. The framework operates exclusively during training, with no extra cost during inference. RLAWM improves performance by up to +12.2 points on PhysReason and +5.2 on PHYSICS over the baseline. It also generalizes zero-shot across difficulty, source, and domain, with a 2.45x gain and +10 points over the strongest baseline on olympiad-level physics and math.

[250] Verifying Human-Written Reasoning References cs.AI | cs.CL | cs.LOPDF

Ali Ansari, Farhana Ferdousi Liza

TL;DR: 该方法通过对齐检查、完整前提重检和失败归因审计人工推理参考链；P-FOLIO 的 386 条推导中仅 202 条（52.3%）通过全部检查，而被引用前提无法蕴含的 230 步里有 191 步可由完整前提推出，说明参考链本身也常有支持遗漏。

Details

Motivation: AI 推理系统常以人工推理链作为参考评估，但每一步是否真的由其引用的前提支持很少被系统核查。若参考推导或其形式化表示有误，与参考不一致就未必意味着 AI 系统推理错误。

Result: 作者提出以参考为中心的审计方法，结合对引用前提的蕴含检查、对全部原始前提的重检，以及错误来源归因。研究在 P-FOLIO 上用大语言模型将步骤结论转成一阶逻辑，再用 Z3 SMT 求解器检查蕴含。审计的 386 条推导中，202 条（52.3%）通过所有标准：每步均由所述支持蕴含，且可在现有一阶逻辑表示下推出目标。在 230 个未被所引用前提蕴含的步骤中，191 个用完整原始前提检查后变为可蕴含，说明多数初始失败来自支持遗漏或不匹配。其余 39 个经人工分析为 21 个人工推理错误、11 个逻辑翻译错误、7 个数据标注错误。另对 50 个随机抽取的通过步骤人工审计，46 个（92%）逻辑翻译忠实。

Insight: 推理基准中的参考链和形式化翻译并非无误差真值；评估分歧可能来自参考支持缺失、人工推理、翻译或标注。结果限于 P-FOLIO 和现有逻辑表示，摘要未讨论其他基准的参考质量。

Abstract: Evaluating the reasoning of AI systems often relies on human-written reasoning chains as references, yet the logical support for their individual steps is rarely verified. Although prior neuro-symbolic approaches have formalized and checked reasoning, their reported evaluations do not systematically assess whether step conclusions in human-written reference derivations are entailed by their cited premise sets or investigate the sources of failures in those references. To address these limitations, we present a novel reference-centred assessment method for human-written derivations that combines entailment checking against cited premise sets, full-premise rechecking, and failure attribution to examine the reliability of these derivations and their formal representations. We evaluate it on P-FOLIO, using a large language model (LLM) to translate step conclusions into first-order logic (FOL) and the Z3 SMT solver to check entailment. Of 386 derivations examined, 202 (52.3%) meet all our checking criteria, meaning that every translated step is entailed by its stated support and the required target is established under the available FOL representations. Among 230 steps not entailed by their stated support, 191 become entailed when checked against the complete original premises, indicating that most initial failures reflect incomplete or mismatched support rather than unsupported claims. Manual analysis of the remaining 39 cases identifies 21 human reasoning errors, 11 FOL translation errors, and 7 dataset annotation errors. A separate manual audit of 50 randomly sampled steps that passed the entailment checks found 46 faithful FOL translations (92%). These findings suggest that disagreement with a human-written reasoning reference may originate in the reference or its formal representation rather than in the AI system being evaluated.

[251] Neuro-Inspired Inhibitory Message Passing for Accurate and Token-Efficient Multi-Agent Reasoning cs.AI | cs.CL | cs.MAPDF

Nima H. Siboni, Magdalena Springer, Vahid Rostami

TL;DR: 受神经抑制启发的消息传递路由器按接收方和证据句去重，只转发接收方尚未掌握的事实并抑制重复内容；相较标准消息传递，多智能体准确率提升 39–77 个百分点、每次运行 token 减少 10–46%。

Details

Motivation: 多智能体协作能汇集单个智能体难以获得的信息，但消息积累会挤占上下文，使关键而稀有的证据被大量重复内容淹没。皮层回路中的抑制机制提示，可选择性压低共享输入，同时保留较弱而有用的信号。

Result: 作者提出免训练抑制式消息传递路由器，逐接收方、逐证据句抵消冗余信息。每轮仅转发其他智能体消息中复述的证据句：每条最多给同一接收方一次，且只发给自身尚无该证据的接收方；发送者的推理过程与答案本身不转发。对比标准消息传递，三个基准上的同伴群组准确率提升 39–77 个百分点，每次运行 token 减少 10–46%。收益覆盖 7 个 4B–32B 开源模型和两个商业模型、最多 64 智能体，以及点对点、编排器—工作者和多级层次架构。

Insight: 多智能体通信应传递稀缺的新证据，而非复制完整推理文本；按接收者过滤已知事实能同时减少上下文占用和证据遮蔽。摘要未提供各基准的绝对准确率或路由器通信延迟。

Abstract: Multi-agent systems promise to solve problems by bringing together information beyond the reach of any individual agent. Retrieval and tool use can distribute this information across agents, making communication essential to collaboration. A central challenge is keeping crucial evidence visible in each agent’s context as messages accumulate. Repeated information can crowd out rare, decisive evidence. Models of cortical circuits suggest a principle for addressing this imbalance: inhibition can selectively suppress shared inputs while preserving sensitivity to weaker signals. We adapt this principle into inhibitory message passing, a training-free communication router that cancels redundant evidence separately for each evidence sentence and receiver. Each round, it forwards evidence sentences restated in agents’ messages, delivering each at most once to receivers whose own evidence lacks it and excluding the senders’ reasoning and answers. Against standard message passing, accuracy in peer groups rose by 39-77 percentage points across three benchmarks, and tokens per run fell by 10-46%. Gains held across seven open models spanning four to thirty-two billion parameters and two commercial models, in groups up to sixty-four agents, and across multi-agent architectures including peer-to-peer, orchestrator-worker, and multilevel hierarchical organizations. These findings support a scalable design pattern for more accurate and token-efficient multi-agent reasoning.

[252] The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents cs.AI | cs.CL | cs.CR | cs.MAPDF

Jinwei Hu, Youcheng Sun, Yi Dong, Xiaowei Huang

TL;DR: One-Relation Attack（ORA）只用一句有来源依据、但限定关键隐含推理前提的话，就能使 LLM 智能体改变原本正确的信念；在 18 个模型上的 best-of-10 攻击平均翻转专有模型 54.9%、开源模型 80.0% 的初始正确回答。

Details

Motivation: 智能体的信念可能依赖证据与结论之间未明说的隐含前提。论文研究攻击者能否不伪造证据，而是通过针对这条推理关系的一句话限定其适用条件，从而误导智能体。

Result: ORA 先定位关键隐含推理前提，再注入一句有来源依据、但重新限定该前提的文本。在生物医学推断和漏洞检测任务、18 种模型的测试中，best-of-10 攻击平均翻转专有模型 54.9%、开源模型 80.0% 的初始正确答案；攻击可在没有受害模型反馈时迁移。配对控制支持攻击依赖目标前提定位和限定性改写。成功攻击句可通过来源支持检查并逃过识别；在平衡压力测试中，多智能体辩论反而使团队错误率从 50.0% 升到 59.5%。

Insight: 事实本身准确并不保证其推理使用安全；隐含的证据—结论关系也构成智能体认知攻击面。论文结果来自特定任务与 best-of-10 设置，摘要未报告防御方法或日常部署中的自然攻击率。

Abstract: Can one sentence manipulate an agent’s belief by targeting a single implicit warrant linking evidence to a conclusion? In this paper, we introduce the one-relation attack (ORA), a novel cognitive attack on this inferential relation. ORA localizes a critical implicit warrant and performs a one-relation intervention using a single source-grounded sentence that qualifies this warrant and steers the agent toward an incorrect belief. Across eighteen models on biomedical inference and vulnerability detection, best-of-$10 attacks overturn initially correct answers at mean rates of 54.9% for proprietary and 80.0% for open-weight models. Attacks also transfer without victim feedback, and paired controls support the roles of targeting and qualification. Moreover, successful sentences can pass source-support checks and evade identification, and multi-agent debate raises team error from 50.0% to 59.5% in our balanced stress test. These findings establish implicit warrants as a cognitive attack surface for manipulating agents’ beliefs.

[253] HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents cs.AI | cs.CLPDF

Soyoung Yoon, Peiran Xu, Han Wang, Ruofan Wu, Xiaodong Yu

TL;DR: HybridResearcher 是一个 4B 深度研究智能体，通过合成并筛选跨结构化数据库与非结构化网页的证据桥接轨迹，再以 SFT 和强化学习训练；HybridDeepResearch 上 pass@8 比 Qwen3.5-4B 基础模型高 30 分，并超过 Qwen3.5-397B-A17B 8.3 分。

Details

Motivation: 真实分析问题常需跨异构来源整合证据，例如同时计算内部数据库结果并查询网页。即使前沿智能体也难以跨越结构化与非结构化信息边界，现有基准暴露了问题，但缺少明确训练此能力的方法。

Result: HybridResearcher 从经验证的单源任务（如 Text-to-SQL 或网页搜索）出发，围绕种子输出中的桥接实体合成缺失来源，构造跨源推理轨迹。只有经过严格构造检查、确认数据库和网页两侧单独均不足以回答的问题才保留，确保数据需要真正的跨源依赖。模型通过监督微调后接强化学习训练。在 HybridDeepResearch 上，4B 模型比 Qwen3.5-4B 基础模型高 30.0 pass@8 分，并超过大得多的 Qwen3.5-397B-A17B 8.3 分；在 BrowseComp 和 BrowseComp-Plus 等留出的单源基准上也优于基础模型。

Insight: 跨源推理能力可以通过构造两侧均不可单独作答的桥接轨迹教给小模型，且作者报告未导致灾难性遗忘。摘要未提供具体单源基准分数、训练成本或合成轨迹规模。

Abstract: While LLM agents are increasingly capable of gathering and synthesizing evidence from a single homogeneous source, such as structured tables or unstructured data, real-world analytical questions frequently span heterogeneous sources. For example, answering one enterprise question may require both a result computed over an internal database and evidence found on the web. Recent evaluations on benchmarks like HybridDeepResearch show that even frontier agents struggle when evidence must cross this boundary. While these benchmarks have exposed the limitation, methodologies to explicitly teach cross-source reasoning remain absent. To address this, we introduce HybridResearcher, a 4B agent trained to bridge evidence across the structured and unstructured boundary. To build it, we synthesize cross-source bridge trajectories by starting with validated single-source tasks-such as existing text-to-SQL or web search queries-and fabricating the missing source around a bridge entity drawn from the seed’s own output. We retain a sample only when strict construction gates confirm that neither side alone suffices, ensuring the data requires true cross-source interdependence. Finally, we train on this curated dataset using supervised fine-tuning followed by reinforcement learning. Evaluated on the existing HybridDeepResearch benchmark, HybridResearcher improves over our base model, Qwen3.5-4B, by 30.0 points pass@8 and surpasses the far larger Qwen3.5-397B-A17B by 8.3 points, while also improving over its base model on established, held-out single-source benchmarks (BrowseComp and BrowseComp-Plus), demonstrating that this complex cross-source reasoning can be taught directly to small models without triggering catastrophic forgetting.

[254] A Cross-Domain Evaluation of Prompting Techniques for Large Language Models cs.AI | cs.CLPDF

Kevin Krazius, Tim Schlippe, Paul Libbrecht

TL;DR: 在同一 LLM 和六个覆盖算术、逻辑、常识推理及事实核验的基准上，Zero-shot、Few-shot、RAG 和自动提示优化没有显著的全域胜者；300 个检索示例时 RAG 平均准确率最高，为 78.33%，说明提示策略应依任务选择。

Details

Motivation: 提示方法常在不同模型、数据集和任务设置下分别评测，导致跨方法和领域的直接比较困难。论文用相同模型和多领域基准控制比较零样本、少样本、检索增强生成和自动提示优化。

Result: 研究用同一大语言模型在六个基准上评估 Zero-Shot、Few-Shot、RAG 和受 PE2 启发的自动提示优化。没有一种方法在所有领域始终领先；对各方法最高配置的成对统计比较也未发现显著差异。可提供 300 个检索示例时，RAG 总体平均准确率最高，为 78.33%；Few-shot 在若干结构化推理任务上较强，Zero-shot 对输入自包含任务仍有竞争力，自动优化方法结果稳定但未系统性占优。增加任务示例数并不保证所有提示方法都持续提升。

Insight: 提示策略收益依赖任务结构和领域，检索示例数量增加也不必然改善表现；应用时应依据任务特点选型，而非默认某种提示优化普遍最优。摘要未给出六个数据集的分项数值和显著性检验功效。

Abstract: Prompt engineering has become a central mechanism for guiding the behavior of Large Language Models (LLMs) across a wide range of reasoning and language understanding tasks [1], [2]. However, existing studies frequently evaluate prompting techniques using different LLMs, datasets, or task settings, making direct comparisons across methods and domains difficult [3]. Consequently, in this paper we present a controlled cross-domain evaluation of Zero-Shot prompting, Few-Shot prompting, Retrieval-Augmented Generation (RAG), and automatic prompt optimization based on a PE2-inspired optimization pipeline [4]-[6]. All experiments were conducted with the same LLM and evaluated on six benchmark datasets covering arithmetic reasoning, logical reasoning, commonsense reasoning, and factual verification. The results show that no prompting strategy consistently outperforms the others across all domains, and pairwise statistical comparisons of the highest-performing configurations revealed no significant differences. RAG achieves the highest overall average accuracy of 78.33% when 300 examples are available for retrieval, while Few-Shot Prompting achieves comparatively high accuracies across several structured reasoning tasks. Zero-Shot Prompting remains competitive on tasks with self-contained inputs, whereas the PE2-inspired optimization approach produces stable but not systematically superior results compared to the other evaluated prompting techniques. Furthermore, increasing the number of task-specific examples does not consistently improve performance across prompting techniques, indicating strong domain and task dependency. These findings suggest that prompting techniques should be selected adaptively based on task characteristics rather than treated as universally applicable optimization strategies.

[255] Revealing what benchmark scores hide: Jointly measuring the breadth and reliability of LLM capabilities using accessibility profiles cs.AI | cs.CL | cs.LGPDF

Kazem Faghih, Yize Cheng, Shoumik Saha, Mobina Pournemat, Armin Gerami

TL;DR: 论文提出用 accessibility profile 描述模型对同一问题不同等义表述的成功概率，并用 Access@5 与 Reliable@5 区分“多次尝试中能成功一次”和“多次都稳定成功”；跨六个基准、20 多个模型，平均表现相近的模型对有时会因此出现最多 10 分的相反排序。

Details

Motivation: 基准通常只用一种固定措辞评估每个问题，但用户可能以多种保持原意的表达提出同一请求。单一措辞的正确率以及平均、最好或最坏提示分数，会掩盖模型能力覆盖多少问题、又能否稳定被触发。

Result: 作者将一个问题的 accessibility 定义为模型在多种语义等价表述下成功的概率，并考察这些概率在问题间的分布。Access@5 衡量五种表述中至少成功一次，Reliable@5 衡量五种都成功；前者适合允许多次重试，后者反映反复可靠性。六个基准、20 多个模型的结果中，约半数平均 accessibility 接近的模型对会因采用“成功一次”还是“全部成功”而改变优劣排序；部分模型对的两项指标可向相反方向相差最多 10 分。规模扩展、推理训练和提示优化也可能在扩大可答问题范围的同时降低重复成功率。

Insight: 提示敏感性不仅是评测噪声，也反映能力广度与可靠性的不同维度；同平均正确率的模型，对需要重试和要求稳定成功的用户可能有相反价值。摘要未说明等义表述的具体生成和人工验证方法。

Abstract: Language model benchmarks assign each question a single outcome from one fixed wording, even though the same underlying request can be expressed in many equivalent ways. We argue that the quantity of interest should therefore not be correctness on one wording, but the accessibility of an inquiry, which we define as the probability that a model succeeds across meaning-preserving formulations. The distribution of accessibility across inquiries reveals a structure that accuracy and existing best-, worst-, or average-case prompt scores collapse: how broadly a model’s capabilities extend across inquiries and how dependably those capabilities can be elicited. Two models can have the same average performance while one succeeds across more inquiries unreliably and the other succeeds across fewer inquiries but more consistently, leading to opposite preferences depending on whether the task allows multiple retries until one success or when repeated consistent success is required. Across six benchmarks and more than twenty models, this tradeoff is common: for about half of the model pairs with nearly equal mean accessibility, which model looks better depends on whether one success across five formulations is enough or all five must succeed, quantities we call Access@5 and Reliable@5, respectively. For some pairs the two measures diverge significantly, by up to 10 points in opposite directions. Accessibility also reveals that scaling, reasoning, and prompt optimization can redistribute capability across inquiries in ways aggregate accuracy obscures, sometimes broadening the set of inquiries a model can succeed on while making success on them less repeatable. Accessibility turns formulation sensitivity from a nuisance into a measurable property of capability, distinguishing what a model can sometimes succeed on from what users can actually depend on.

[256] CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents cs.AI | cs.CLPDF

Zhongwei Xie, Yiming Du, Jiaxin Bai, Baixuan Xu, Chunyang Li

TL;DR: CoworkEnv 根据工作流需求合成可执行的 AI 同事训练环境，并用程序验证与反馈修复检查任务规格和评分行为；Qwen3-VL-8B-Instruct 经 SFT 后 Claw-Eval 平均分由 46.6 升至 64.1，多轮任务提高 31.9 分。

Details

Motivation: AI 同事需要协调工具、跨轮沟通、操作变化中的工作区状态并遵循流程约束。训练这类能力需要彼此一致的请求、服务记录、工具接口和评分检查，手工构建成本高且合成环境可能自身不可靠。

Result: CoworkEnv 以能力为导向合成可执行训练环境，由工作流要求指导任务构造，再用程序化验证和反馈修复检查生成规格与评分逻辑。教师执行提供监督微调示范，学生交互则产生强化学习 rollout 和奖励。基于这些资源训练 PocketWorker 紧凑模型；Qwen3-VL-8B-Instruct 经 SFT 后，Claw-Eval 平均分从 46.6 提升到 64.1，多轮任务增加 31.9 分。SFT-only 扩展实验显示方案也适用于 ClawMark。论文还比较不同奖励处理和信用分配下的 RL 动态，当前改善训练奖励的结果被作者定位为环境支持策略优化的初步证据。

Insight: 训练环境不仅要生成任务和工具，还应程序化核验其执行与评分逻辑；教师示范和学生 rollout 可由同一可执行环境支持。强化学习证据仍属初步，摘要未报告最终泛化性能。

Abstract: AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints. Constructing training experience for these capabilities requires coherent requests, service records, tool interfaces, and grading checks. We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers. Workflow requirements guide task construction, and programmatic validation with feedback repair checks generated specifications and grading behavior. Teacher executions yield curated demonstrations for supervised fine-tuning (SFT), while student interactions provide rollouts and rewards for reinforcement learning (RL). Using these resources, we train compact PocketWorker models. For Qwen3-VL-8B-Instruct, SFT raises the mean score from 46.6 to 64.1 on Claw-Eval, including a 31.9-point gain on Multi-turn tasks. An SFT-only extension to ClawMark demonstrates applicability across coworker benchmarks. We further compare RL training dynamics across configurations that vary reward processing and credit assignment. Improving training rewards provide preliminary evidence that the constructed environments support policy optimization.

[257] Aligning Language Models with Indigenous Community Cosmovision through Constitutional AI cs.AI | cs.CL | cs.CYPDF

Silvia Fernández-Sabido, Jorge Reyes-Magaña, Jesús Adrián Chay Tun, Jean Carlos Buenfil, Victor González Nahuat

TL;DR: 论文以拉美原住民社群原则构建 12 条文化宪章，并用 SFT+DPO 对 Qwen2.5-7B 和 1.5B 模型进行价值对齐；仅用 24 组偏好数据也观察到代际传承、语言谨慎及咨询同意方面的有限改善。

Details

Motivation: 大语言模型训练语料主要反映西方英语世界观，当模型回答领土、社群权威或人与自然关系问题时，可能默认使用与使用者社群不相符的前提。论文探索如何以拉美原住民社群伦理原则构建参与式价值对齐流程。

Result: 作者访谈五位原住民领袖和倡导者，形成基于原住民宇宙观的 12 项操作性文化宪章，并对 Qwen2.5-7B-Instruct 与 Qwen2.5-1.5B-Instruct 实施两阶段 Constitutional AI：监督微调后进行直接偏好优化。使用 24 组偏好数据后，在代际传承、语言谨慎、咨询与同意方面观察到幅度不大但可检测的改善。定性分析认为主要对齐转变来自 SFT，DPO 进一步调整特定类别。系统还部署为带检索增强生成的对话界面，使社群能够补充知识库。

Insight: 价值对齐可以从具体社群参与制定的原则出发，并通过可更新知识库形成参与式循环；小样本结果是初步证据，作者也指出扩大语料和计算资源仍是后续条件。摘要未给出各类别量化指标或更大样本验证。

Abstract: Large language models (LLMs) are predominantly trained on corpora that reflect Western, anglophone worldviews. This asymmetry is not neutral: when a model answers questions about territory, community authority, or the relationship with nature, it does so from assumptions that may be alien to the communities using the system. This paper presents a system for aligning language models with the ethical principles of Indigenous communities in Latin America. Drawing on interviews with five Indigenous leaders and defenders, we built an operational cultural constitution grounded in an Indigenous cosmovision, with 12 principles. On this basis, we implemented a Constitutional AI (CAI) pipeline in two phases: supervised fine-tuning (SFT) and direct preference optimization (DPO). The base models were Qwen2.5-7B-Instruct and Qwen2.5-1.5B-Instruct. With only 24 preference pairs, improvements were modest but detectable in intergenerational transmission, linguistic caution, and consultation and consent. Qualitative analysis reveals that the SFT phase was responsible for the primary shift in alignment, while DPO refined specific categories. The system was deployed in a conversational interface with retrieval-augmented generation (RAG) that allows communities to feed the knowledge base, turning alignment into a participatory cycle. We argue the architecture is functional and will scale qualitatively once the corpus is expanded and computing constraints are resolved. The work constitutes a first iteration of an epistemic justice mechanism for value alignment in the Global South.

[258] Outcome-Anchored Credit Reallocation for Agentic Knowledge Graph Question Answering cs.AI | cs.CL | cs.LGPDF

Mingyuan Li, Yucheng Li, Xiaxia Wang, Gong Cheng

TL;DR: OACR 用只在训练期运行的知识图谱状态结果模型估计相邻搜索状态间的答案奖励变化，再把轨迹级优势重新分配到图谱操作和生成 token；三个基准上整体优于已有 KGQA 方法，并减少推理期 LLM 调用与 token。

Details

Motivation: 智能体知识图谱问答需多轮查询图谱。结果监督强化学习通常从最终答案得到可验证奖励，却把同一轨迹优势均匀分给所有策略生成 token，忽略不同图谱操作对搜索过程和最终答案的作用不同，形成轨迹级奖励与操作级贡献不匹配。

Result: Outcome-Anchored Credit Reallocation（OACR）在训练时使用知识图谱状态结果模型，估计相邻图谱搜索状态的结果奖励变化，并将其转成操作级信用，再为 token 生成权重。这些乘法权重调整 token 级更新幅度，同时保留每个生成 token 上原始结果优势的正负号。三个基准实验中，OACR 的整体性能超过现有知识图谱问答方法，并减少推理阶段的大语言模型调用和 token 数。

Insight: 逐操作估计搜索状态变化可将最终答案的结果奖励细化到对搜索真正有帮助的动作，且不反转原始成功/失败信号。摘要没有提供各基准准确率、调用量减少幅度或状态模型误差。

Abstract: Agentic knowledge graph question answering (KGQA) requires large language models (LLMs) to solve complex questions through multi-turn interactions with a knowledge graph (KG). Outcome-supervised reinforcement learning derives verifiable outcome rewards from final answers but uniformly assigns each rollout’s outcome-level advantage to all policy-generated tokens, creating a trajectory-to-operation mismatch since KG operations within a rollout can have varying effects on the search process and final answer. To address this, we propose Outcome-Anchored Credit Reallocation (OACR), which employs a training-only KG-state outcome model to estimate outcome-reward changes between adjacent KG search states and converts them into operation credit, from which token-wise weights are derived. These multiplicative weights reallocate token-level update magnitudes while preserving the sign of the outcome-level advantage at every generated token. Across three benchmarks, OACR achieves stronger overall performance than existing KGQA methods while using fewer inference-time LLM calls and tokens.

[259] A failure of Imagination: VLMs fail to use visual representations for text, but self-distillation can help cs.AI | cs.CL | cs.CVPDF

Shashwat Singh, He He, Tal Linzen

TL;DR: 研究发现常规 VQA 对齐虽让模型能回答输入图像问题，却未让文本表示自动获得视觉结构；作者将上下文中的视觉嵌入蒸馏到文本嵌入后，模型在仅以文字描述几何形状的任务上改善，且无需该任务的显式监督。

Details

Motivation: 视觉语言模型通常通过图像问答微调语言模型关注视觉特征，但这是否改变其纯文本表征、使文本中隐含的可视化结构也能被利用，尚不清楚。论文区分仅能回答图像问题的弱多模态与文本表示本身视觉对齐的强多模态能力。

Result: 作者构造形状计数任务，问题可用等价的图像规格或线段坐标文本描述：给图像时很简单，纯文本时需在脑中构造几何图形。比较指令微调的 Qwen-2.5、Mistral 与视觉对齐版本发现，多模态模型虽擅长图像版本，在文本版本上却无提升；探针实验也显示视觉对齐解码器嵌入并不比纯文本嵌入包含更多视觉结构。随后作者提出把上下文中的视觉嵌入蒸馏到文本嵌入的优化目标，提升纯文本任务表现，且无需该任务的显式监督，对复杂度高于对齐样本的测试也有帮助。

Insight: 图像输入上的强表现不意味着视觉知识会迁移到纯文本推理；跨模态迁移需要直接对齐文本表示和视觉表征。实验显示常规 VQA 对齐留下了可利用的迁移空间，但摘要未给出准确率增幅或更多几何任务验证。

Abstract: Vision-language (VL) models are typically built by fine-tuning a language model to attend to visual features in service of a Visual Question-Answering objective. We ask whether this alignment changes how language models represent text—especially on tasks where visualization would help, like geometry questions posed in words. We distinguish weak multimodality (the model can answer questions about an input image) from strong multimodality (the model’s text representations themselves become vision-aligned, improving performance on text-only tasks with latent visual structure). To test for strong multimodality, we construct a task of shape-counting problems with paired equivalent text and image specifications. This task is trivial when the image is provided, but requires visualization when specified in text as line coordinates. We compare instruction-tuned Qwen-2.5 and Mistral models against their vision-aligned counterparts on the text variant of the task. We find that the multimodal models do no better in this setting, despite strong performance on the image version of the task. Probing experiments corroborate this: vision-aligned decoder embeddings carry no more visual structure than their text-only counterparts. Finally, we provide an optimization objective that distills in-context visual embeddings into text embeddings. This alignment induces gains on the text variant of the task without explicit task supervision, even on samples more complex than seen during alignment. This suggests that VQA-style alignment leaves available cross-modal transfer untapped.

[260] Auditing Tool Traces in Agentic LLMs: Source Localization and Counterfactual Dependence Testing cs.AI | cs.CLPDF

Junhao Liu, Tianxiao Li, Yongxian Wu, Ziheng Duan, Ren Li

TL;DR: 该审计先用逐工具 TF-IDF 分解定位带标签预测力的工具结果，再通过 token 类别遮蔽反事实测试冻结 LLM 是否依赖捷径；在 TCR-pMHC 结合预测中，去除单个工具结果中的数据集标识符后，工具增强性能从表面优势降至随机水平。

Details

Motivation: 智能体通过序列化工具结果获取科学证据，但其中可能隐藏数据集编号、模板痕迹等标签捷径，使模型在没有真正生物学推理时也表现成功。仅看到工具增强性能提升，无法确认模型用的是科学证据还是工具输出中的泄漏信号。

Result: 论文提出两阶段审计：Per-Tool TF-IDF Decomposition（PT2D）将标签预测模式定位到单个工具结果，避免把所有工具轨迹拼接后稀释局部信号；Token-Class Redaction 再反事实遮蔽所定位的 token 类别，检验冻结模型是否依赖它。在 TCR-pMHC 结合预测中，工具增强相比无证据基线显著提升，表面上像是模型利用了科学证据；审计却发现最强信号来自某工具结果中的标识符，遮蔽后性能降至随机水平。

Insight: 工具调用轨迹应按来源逐项审计，并对疑似捷径做反事实删除；整体基准增益可能由单一来源中的数据泄漏驱动。结果集中在 TCR-pMHC 任务，摘要未给出原始增幅或其他生物学任务的普遍性。

Abstract: Agentic LLMs increasingly support scientific prediction by grounding decisions in external evidence. In biology, they promise accurate, interpretable predictions for important problems by integrating complex experimental and computational evidence. However, serialized tool results can contain hidden label shortcuts, such as dataset identifiers or template artifacts, that create apparent success without biological reasoning. To address this risk, we introduce a two-stage audit. Stage 1, Per-Tool TF-IDF Decomposition (PT2D), localizes label-predictive patterns to individual tool results while avoiding signal dilution in the concatenated tool trace. Stage 2, Token-Class Redaction, counterfactually tests whether the frozen LLM depends on the localized shortcut. We apply this audit to TCR-pMHC binding, where tool augmentation substantially outperforms a no-evidence baseline. At face value, this suggests that the LLM has learned to use scientific evidence. The audit overturns this interpretation: PT2D localizes the strongest signal to identifiers in one tool result, and redacting them reduces performance to chance. These findings show how hidden label shortcuts can masquerade as scientific reasoning or discovery, motivating source-level audits before benchmark gains are trusted.

[261] Standing on the Shoulders of Giants: Enhancing LLM Agents for Automated Fact-Checking with Human Expert Skills cs.AI | cs.CLPDF

Xinwen Chen, Haorui He, Yupeng Li

TL;DR: SkillCheck 将 AFP Fact Check 公开指南蒸馏为可复用的 Professional Fact-Checking Skill，并在推理时动态指导证据检索和主张核验；AVeriTeC 上准确率提高 2.2–4.8 个百分点、Macro-F1 提高 1.5–6.3 点，GPT-4o-Mini 还以约低 61% 成本超过无技能 GPT-4o 基线。

Details

Motivation: LLM 智能体自动事实核查仍受证据检索和主张验证限制，已有改进常依赖昂贵后训练或难以扩展的人类反馈。论文研究能否将专业核查经验转化为推理时可调用技能，避免额外训练和持续人工监督。

Result: 作者将 AFP Fact Check 公开指南整理为模块化的 Professional Fact-Checking Skill（PFC-Skill），并构建 SkillCheck，让智能体在推理时动态使用该技能指导证据检索与主张核验。多个骨干模型上的 AVeriTeC 实验中，较不使用技能的 DEFAME 基线，准确率提高 2.2–4.8 个百分点，Macro-F1 提高 1.5–6.3 点。GPT-4o-Mini 配合 SkillCheck 的准确率比 DEFAME 的 GPT-4o 高 3.4 个百分点，推理成本约降低 61%；增益也延伸至开放网页检索。消融显示检索阶段引入技能贡献最大，两阶段都使用效果最好，案例分析显示检索更有针对性、证据审查更仔细。

Insight: 专业流程知识可作为推理时模块化技能复用，不必微调模型或持续依赖人工反馈；证据检索阶段的技能使用是主要收益来源。结果针对 AVeriTeC 等评测，摘要未给出不同主张类别的分项成绩。

Abstract: Automated fact-checking (AFC) systems powered by large language model (LLM) agents still face challenges in evidence retrieval and claim verification. Existing approaches improve their capabilities through computationally costly post-training or human-in-the-loop feedback that is difficult to scale. To transfer professional fact-checking expertise to LLM agents efficiently, we distill publicly available guidelines published by Agence France-Presse (AFP) Fact Check into a reusable, modular Professional Fact-Checking Skill (PFC-Skill). We then present SkillCheck, the first AFC system that gives LLM agents dynamic access to PFC-Skill to guide evidence retrieval and claim verification at inference time. Extensive experiments demonstrate that SkillCheck improves upon skill-free DEFAME baselines across multiple LLM backbones, with gains of 2.2–4.8 pp. in Accuracy and 1.5–6.3 pp. in Macro-F1 on the AVeriTeC benchmark. Notably, SkillCheck with GPT-4o-Mini exceeds DEFAME with GPT-4o by 3.4 pp. in Accuracy while cutting inference cost by roughly 61%. The performance gains also extend to open-web retrieval. Ablations show that incorporating the skill during evidence retrieval provides most of the gains, while incorporation into both stages yields the best overall performance. Case studies show that PFC-Skill enables more targeted search and closer scrutiny of evidence. To conclude, our work shows that LLM agents can “stand on the shoulders of giants” by drawing on the best practices of professional fact-checkers, substantially enhancing AFC performance without extra training or continuous human supervision.

[262] GLOBE: Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories cs.AI | cs.CLPDF

Xinyue Peng, Zhenlin Liang, Meng Cai, Hua Wang

TL;DR: GLOBE 在推理时把已完成思路段嵌入单位球面，以短答案试探和语义轨迹反复反向且净进展很小作为提前停止信号；两个模型、五类任务上输出 token 减少 49.7%、端到端活跃时间减少 51.0%，五任务平均准确率 80.58%，与完整 CoT 相当。

Details

Motivation: 思维链能改善问题求解，但模型有时在正确答案已可得到后仍继续生成，增加服务成本，甚至用错误后续推理替换正确中间结果。论文提出无需训练、无需额外模型的在线停止算法。

Result: Geometry-Led Online Behavioral Early Exit（GLOBE）将每段已完成推理嵌入单位球面，并读取语义轨迹形状。试探阶段在独立分支生成短答案，若置信度超过阈值则停止主流生成；若继续推理，则在轨迹反复改变方向且净进展有限时提前退出。在五个涵盖数学推理、工具调用和智能体记忆的基准、两个模型上，GLOBE 平均减少 49.7% 输出 token 和 51.0% 活跃端到端耗时，五任务平均准确率为 80.58%，与完整 CoT 相当。

Insight: 答案可达性与推理轨迹是否继续取得进展可作为互补的提前退出信号，避免无效长链扩大成本和引入新错误。摘要未给出各任务准确率和停止阈值敏感性。

Abstract: Chain-of-thought (CoT) reasoning can improve problem solving, but models often keep generating after the correct answer is already reachable; the extra steps raise serving costs and may replace a correct intermediate result with an incorrect one. We propose GLOBE (Geometry-Led Online Behavioral Early Exit), an online, training-free algorithm that terminates the model’s CoT process early. GLOBE embeds each completed reasoning segment as a point on a unit sphere and reads the shape of the resulting semantic trajectory. GLOBE first applies a trial phase that generates short trial answers in a separate branch, stopping the main stream on an accepted trial after checking confidence against a threshold. If generation continues, a trajectory phase stops the stream once this path repeatedly reverses direction with little net progress. On five benchmarks including mathematical reasoning, tool calling, and agent memory, GLOBE cuts output tokens by 49.7% and active end-to-end wall time (E2E) by 51.0% across two models, and achieves a five-task average of 80.58%, comparable to Full CoT.

[263] Understanding When Reflective Prompt and Skill Optimization Helps on Locally Served Reasoning Models: A Two-Model Study of GEPA and SkillOpt cs.AI | cs.CLPDF

Andrea Giummarra, Carmelo Baglieri, Antonio Di Maria, Andrea Farruggio, Michele Piatto

TL;DR: 在本地部署的 DeepSeek V4 Flash 和 Qwen3.8-27B 上，GEPA 与 SkillOpt 的收益取决于任务；两者均未改善 HotpotQA，GEPA 却在 MATH-500 上把准确匹配率从 0.828 提至 0.904、另一模型从 0.852 提至 0.908，显示验证集改进可能无法泛化。

Details

Motivation: 反思式文本优化器 GEPA 和 SkillOpt 已报告在前沿及中型模型上取得明显增益，论文考察它们能否在本地服务推理模型上复现，并控制模型、任务和评测协议来区分提示或技能效果。

Result: 研究比较 DeepSeek V4 Flash（通过 DwarfStar 服务）和 Qwen3.8-27B（通过 LM Studio）上的最小基线、GEPA 提示、SkillOpt 技能和人工提示，在 HotpotQA 与 MATH-500 使用同一保留任务协议。主要 Qwen v2 重跑中，GEPA 在 HotpotQA 选择了原始种子，行为与基线完全相同；SkillOpt 的验证提升在留出集上反转为退化。MATH-500 上 GEPA 将 exact match 从 0.828 提至 0.904，同时降低延迟但增加 token 消耗；DeepSeek 上则从 0.852 提至 0.908。两模型上 GEPA 最强收益都出现在 MATH-500，HotpotQA 上两种优化器均未改善基线。较早一次 Qwen 运行作为服务敏感性证据保留。

Insight: 提示或技能优化不是普适增益，必须按模型—任务组合和留出集评估；延迟下降也不意味着 token 成本下降。论文提出把内容推理、答案实现与评测器对齐、服务完成可靠性分别分析；样本规模仅为所列两模型和两任务。

Abstract: Reflective text-space optimizers such as GEPA and SkillOpt have reported large gains on frontier and mid-size models. We test replication across two reasoning-native target-serving configurations: DeepSeek V4 Flash through DwarfStar and Qwen3.8-27B through LM Studio. For each target, we evaluate the same held-out tasks and four conditions - minimal baseline, GEPA prompt, SkillOpt skill, and manually engineered prompt - on HotpotQA (n=500 per condition) and MATH-500 (n=250). In the primary Qwen v2 rerun under the same frozen protocol, GEPA selects the HotpotQA seed and exactly reproduces baseline behavior, whereas SkillOpt’s validation gain reverses into held-out degradation. On MATH-500, GEPA raises exact match from 0.828 to 0.904 while reducing latency but increasing token consumption, so lower response time does not imply lower token cost. DeepSeek likewise shows its strongest GEPA effect on MATH-500 (0.852 to 0.908). Across both targets, GEPA’s strongest gains occur on MATH-500, while neither optimizer improves the HotpotQA baseline. Optimization is therefore task- and method-dependent, and validation gains may reverse on held-out evaluation. An earlier Qwen run is retained as serving-sensitivity evidence. These results suggest that textual headroom for a model-task pair and the observed effect of a text-space optimizer are more informative when decomposed into content/reasoning, answer-realization/evaluator alignment, and serving-completion/reliability. The cross-benchmark pattern further suggests that optimization may benefit from low-ambiguity, actionable evaluation signals.

[264] Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law cs.AI | cs.CLPDF

Daan Boven, Gregor Behnke, David Graus

TL;DR: 该神经符号系统用 Catala 形式化荷兰破产法第 42 条撤销权，并让神经模块从事实和判例中处理开放文本法律概念；100 份人工标注判决上的分段与标注精确率、召回率较高，10 个案例测试中多数标准处理合理，但知识要件仍困难。

Details

Motivation: 法律条文适用常需同时进行结构化规则推理，并理解判例形成的开放文本概念。荷兰《破产法》第 42 条的 actio pauliana（债权人损害交易撤销权）需要两者结合。

Result: 作者提出第 42 条的神经符号模型，符号核心以文学式 Catala 编写，使法条与形式化规则保持对应；开放文本要件作为布尔输入连接神经模块和符号求解器。神经流程先把自然语言案件事实整理为布尔值和日期，再用荷兰民事判决的案例推理解决开放概念：检索时按待判断要件筛选判决，并沿案件程序链扩展。对 100 份人工标注判决的分段和标注达到较高精确率与召回率；十个手工标注案例中，模型能合理处理多数法律标准，但在“知情”要件上困难。改变判例检索深度没有明确量化效果，不过检索判决为判断提供了明确判例依据。

Insight: 可形式化的条文规则适合交给符号求解器，开放概念则可由经过筛选的判例检索支持；两者之间的接口质量和知识要件判断仍是瓶颈。摘要未提供具体精确率/召回率或十案逐项准确性。

Abstract: Applying a legal provision frequently requires both structured rule-based reasoning and sensitivity to open-textured terms developed in case law. The actio pauliana of Article 42 of the Dutch Bankruptcy Act (Faillissementswet, Fw), which allows an insolvency practitioner to avoid transactions that prejudice creditors, requires both. We present a neuro-symbolic model of Article 42 Fw. Its symbolic core is written in Catala in a literate style that keeps statute and formalisation aligned and was developed through pair programming, with open-textured terms represented as boolean inputs where neural components couple to the symbolic solver. A neural routine structures a case’s natural-language facts into boolean and date inputs. Neural subroutines resolve the open-textured terms through case-based reasoning over Dutch civil judgments, retrieved through a pipeline that filters on the requirement being resolved and expands along the procedural chain of a case. Segmenting and labelling those judgments attains high precision and recall against a manually annotated gold standard of 100 judgments. On ten hand-annotated cases the neuro-symbolic model handles most criteria reasonably but struggles with the knowledge requirement. Retrieval depth shows no clear quantitative effect, although retrieved judgments provide explicit grounding in prior case law.

[265] The Scheduler Is a Hidden Retriever: How Asynchrony Shapes Evidence Acquisition in Search Agents cs.AI | cs.CL | cs.IRPDF

Ruiyang Ren, Yuhao Wang, Yinan Zhang, Chunyan Miao

TL;DR: 研究发现异步结果到达顺序会持续改变搜索智能体后续证据获取，即使初始证据相同；CEA 在每个子问题证据齐备后才释放局部推理，并独立分配共享搜索，使获取分歧接近重复运行噪声，且比等待全部证据更快。

Details

Motivation: 搜索智能体常并行推理和异步收集信息，但不同工具结果到达时间可能改变智能体在全部结果到齐前的搜索决策。论文检验临时看到的信息差异是否会在之后持续影响智能体选择搜索什么证据。

Result: 两种模型家族、三个基准的受控重放显示，即使初始证据相同，结果到达时机也会改变搜索决策；获取分歧最多比同顺序重复运行差异高 24 个百分点。差异在全部初始证据到齐及第二次搜索后仍持续，形成短暂可见信息对后续搜索的长期影响。作者给出获取遗憾界及匹配构造，并提出 Confluent Evidence Acquisition（CEA）：只有子问题证据完整后才释放局部推理，并让共享搜索分配不依赖完成顺序。在固定计划与稳定条件响应规律下，证明决策、信息获取和答案联合不变。三个基准的问题专属计划实验中，CEA 将获取分歧降至仅比重复运行波动高 1.0 个百分点，准确率接近完整输入对照，并比早期承诺最高高 12.6 个百分点。争用时将固定分配换成 first-ready 分配会使获取分歧增加 21.6 点；CEA 比等待所有初始证据更早作决定、响应更快。

Insight: 异步调度实际上参与了检索决策；只在有局部证据后释放推理、并使共享资源分配与完成顺序解耦，可避免偶然到达时机改变证据链。摘要未给出各基准绝对准确率或不同计划复杂度下的扩展结果。

Abstract: Language-model agents increasingly solve complex tasks through parallel reasoning and asynchronous information gathering. Controlled replay across two model families and three benchmarks shows that delivery timing redirects search decisions made before all results arrive, despite identical initial evidence. Acquisition disagreement reaches up to 24 percentage points above same-order repeat variation. This dependence persists after all initial evidence arrives and through a second search, turning temporary differences in what agents have seen into lasting differences in what they acquire. An acquisition-regret bound with a matching construction shows that invariance at a fixed early arrival count can limit the use of evidence already received. We introduce Confluent Evidence Acquisition (CEA), which instead releases local reasoning when each subquestion’s evidence is complete and allocates shared search independently of completion order. For a fixed plan and stable conditional response law, we prove joint invariance of decisions, acquisitions, and answers. With question-only plans across three benchmarks, CEA brings acquisition disagreement to within 1.0 percentage point of repeat variation and achieves accuracy comparable to complete-input controls, improving over early commitment by up to 12.6 percentage points. Replacing CEA’s fixed allocation with first-ready allocation under contention adds 21.6 percentage points of acquisition disagreement, while local query-target disagreement remains near repeat variation. Timed comparisons show that CEA reaches decisions earlier and responds faster than the same program waiting for all initial evidence.

[266] VibeTraining-Bench: Data-Trainer-Harness Co-Design for Iterative System Improvement cs.AI | cs.CLPDF

Ruotian Peng, Yuanyi Ren, Shuaiqi Liu, Xin Guo, Fang Zhu

TL;DR: VibeTraining-Bench 评估研究智能体在固定资源下协同改进数据、训练器和推理 harness 的能力；GPT-6-Astra 在 SWE-bench Verified 上将目标系统从 32.9% 提升到 53.7%，接近人类开发指令模型的 55.4%。

Details

Motivation: 递归自我改进要求智能体诊断系统瓶颈，并在训练数据、训练器和推理流程间协调优化。论文将这种跨组件共同设计能力作为评测对象，同时避免研究智能体通过测试泄漏或环境作弊获得虚假收益。

Result: VibeTraining-Bench 覆盖终端操作、软件工程、多轮工具使用和数学推理。各任务中研究智能体从相同初始系统出发，在固定资源预算和验证反馈下迭代修改数据、训练器与推理 harness，再提交系统在测试集评估。防护包括隔离测试数据、分开研究与评分容器以及限制研究智能体联网。评测显示前沿研究智能体能显著提升目标系统：GPT-6-Astra 在 SWE-bench Verified 上将成绩从 32.9% 提至 53.7%，接近人类开发指令模型的 55.4%。

Insight: 智能体研究能力应评估其是否能跨数据、训练和推理系统迭代协同优化，而非只生成一次方案；固定资源和隔离测试有助于检验可复现提升。摘要未给出其他智能体、任务及累计改进的详细数值。

Abstract: Recursive self-improvement (RSI) requires agents to diagnose system bottlenecks and coordinate improvements across training and inference. We refer to the agents’ capability to jointly optimize training data, the trainer, and the inference harness as cross-component co-design. We introduce VibeTraining-Bench, a benchmark for evaluating this capability across tasks spanning terminal operation, software engineering, multi-turn tool use, and mathematical reasoning. Within each task, research agents start from an identical initial agent system under a fixed resource budget. Guided by validation feedback, they iteratively refine data, trainers, and inference harnesses, then select and submit a final agent system for evaluation on the test set. To prevent reward hacking, we enforce three safeguards: test data isolation, separate research and grading containers, and network restrictions for research agents. Our evaluation results show that frontier research agents substantially improve target agent systems. For example, on SWE-bench Verified, GPT-6-Astra raises the target agent system’s performance from 32.9% to 53.7%, comparable to the human-developed instruct model’s 55.4%. However, there is still room to improve their ability to co-design and iteratively refine data, trainer, and harness to achieve cumulative performance gains. We hope that VibeTraining-Bench will serve as a unified platform for advancing RSI, deepening our understanding of cross-component optimization for agent systems. Code and data are available at https://github.com/D2I-ai/VibeTrainingBench.

[267] When the harness decides success: state that cannot be re-observed cs.AI | cs.CL | cs.LGPDF

Raimondas Lencevicius

TL;DR: 受控任务显示，harness 丢弃上下文时若事实还能重新查到，低成本推理模型成功率不受影响；若事实不可重观察则降至随机，二者差 92.5 个百分点。ARC-AGI-3 单模型四游戏分解中，改为摘要旧轮次且保留记笔记提示可恢复标准与 Provider Adapter harness 63 分差距的 86%，但每种配置仅一次运行。

Details

Motivation: 智能体 harness 决定多轮间保留、摘要或重放哪些历史内容，论文研究这些选择何时改变成功率，而非只比较成本；尤其关注被丢弃的状态能否在后续重新观察，以及模型自身早期推理是否值得重放。

Result: 预注册合成测试发现，当所需事实可再次查询时，截断上下文对低成本推理模型成功率没有影响，模型会重新查找；无法重观察时成功率降至随机水平，两种情况相差 92.5 个百分点（90% CI [89.5,95.5]）。第二提供商一个模型的预注册复现得到相同差异，事实丢失后模型选择不作答而非猜测。ARC-AGI-3 中，作者对一个模型、四款游戏逐项分解标准 harness 与 Provider Adapter 的差距：不再丢弃而是摘要旧轮次，同时保留标准 harness 的记笔记提示，可在不重放模型旧推理的情况下追回 63 分差距的 86%；没有记笔记提示时追回 43%，加入提供商原生推理重放后再追回其余 57%。这些是每种配置和游戏各一次计数的点估计，推理重放的占比受单次运行影响。已到达关卡上 Provider Adapter 优势在深层更大，但作者指出这基本复述了运行停止位置，不能区分优势是否随深度增加。

Insight: 上下文管理的影响取决于丢失事实能否重新观察；摘要历史与记笔记可能恢复长程任务表现，不一定需要重放模型先前推理。ARC-AGI-3 的 harness 结果样本很少，不能视为稳定估计。

Abstract: Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model’s own earlier reasoning. We ask when these choices change whether an agent succeeds, not only what it costs, and report three findings. First, in a pre-registered synthetic probe, truncating the context window costs a low-cost reasoning model nothing in success when a needed fact can be looked up again (it looks the fact up instead), but drops its success to chance when the fact cannot be re-observed: the two penalties differ by 92.5 percentage points (90% CI [89.5, 95.5]). A pre-registered replication on one model from a second provider showed the same contrast. Once the fact was lost, that model chose no answer rather than guessing. Second, on ARC-AGI-3, where the benchmark’s Standard harness trails its Provider Adapter harness by a wide margin, we decompose that gap one component at a time, for one model on four games. Summarizing old turns instead of dropping them, while keeping the Standard harness’s invitation to take notes, recovers 86% of a 63-point gap without replaying the model’s reasoning. The same harness without the invitation recovers 43%, and adding the provider’s native replay of the model’s reasoning recovers the remaining 57%. These are point estimates from one counted run per configuration and game, and the replay share depends on a single run. Third, in the benchmark’s published runs and in ours, the Adapter’s advantage in the probability of clearing a level it has reached is larger at deep levels than at early ones. This largely restates where runs stopped and does not separate a growing advantage from a constant one.

cs.CR [Back]

[268] Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents cs.CR | cs.CL | cs.CVPDF

Wonjun Lee, Vaidehi Patil, Zaid Khan, Yue Zhang, Hyunji Lee

TL;DR: LookAhead Defense 利用模型发布时必须表现安全这一点，构造不依赖外部安全回答的 Safety Buffer，并预演良性微调、只惩罚预测会削弱安全性的更新；在语言模型、语言智能体和多模态智能体中均降低后门行为，同时维持较好效用并控制过度拒答。

Details

Motivation: 休眠投毒可在模型发布时隐藏恶意行为，待模型经过看似良性的微调后再触发；智能体还可能把潜伏行为扩展到未见动作，多个恶意行为也可能共同被触发。现有防御常需外部安全回答或可信干净参考模型，而智能体动作场景中这些资源难以获得，过度惩罚也会造成过度拒答或效用下降。

Result: 作者利用投毒者必须让发布模型表面安全的特性，自动构造 Safety Buffer：将有害输入与发布模型原有回答配对，并加入由同一模型生成、去除有害意图的良性版本。LookAhead Defense 用缓冲区预演良性微调更新，只惩罚预测会削弱有害输入安全性，或增加对良性版本拒答概率的更新。在语言模型、语言智能体和多模态智能体上，相较普通微调，有害智能体行为比例最高从 93.9% 降至至多 14.2%，语言模型有害回答最高从 92.5% 降至至多 2.0%，同时保持有竞争力的效用和较低过度拒答；对其他投毒方法也有效。

Insight: 发布阶段的安全行为本身可作为后续微调的安全参照，不必依赖外部干净模型或人工安全回答；预测性筛选更新还可避免把所有变化都一概惩罚。摘要没有给出各任务细分结果或安全性与效用的具体数值。

Abstract: Recent work shows that attackers can implant dormant malicious behaviors that remain hidden at release but emerge after benign fine-tuning. As foundation models are used as agents, whether dormant poisoning extends to harmful agent actions and multimodal settings remains unclear. We show two agentic risks: (1) harmful behaviors activated by benign fine-tuning can extend to unseen actions; (2) multiple harmful behaviors can coexist and be jointly activated by benign downstream fine-tuning. Existing defenses often rely on external safe responses, which are hard to obtain for agent actions, or a trusted clean reference model. We exploit a key property of dormant poisoning: the attacker must preserve safe behavior at release to conceal poisoning, and this behavior can serve as a safety reference. Leveraging this property, we automatically construct a Safety Buffer pairing harmful inputs with the released model’s responses and benign versions generated by the same model to remove harmful intent. The Safety Buffer can replace external safe responses in existing defenses, but these defenses do not distinguish updates that weaken safety from those that preserve it, often increasing over-refusal or reducing utility. To address this limitation, we introduce LookAhead Defense, which previews benign fine-tuning updates using the Safety Buffer and penalizes only updates predicted to weaken safe behavior on harmful inputs or increase the likelihood of refusals on benign versions in the Safety Buffer. Across language models, language agents, and multimodal agents, LookAhead Defense reduces harmful behavior in agents from up to 93.9% to at most 14.2% and harmful responses in language models from up to 92.5% to at most 2.0% versus vanilla fine-tuning, while maintaining competitive utility and low over-refusal. It remains effective against other poisoning methods.

cs.DC [Back]

[269] StreamAhead: Real-Time Streaming Video Question Answering via Adaptive Precomputation cs.DC | cs.CVPDF

Haoyang Fan, Dhruv Parikh, Rajgopal Kannan, Viktor Prasanna

TL;DR: StreamAhead 将实时视频片段保持在视觉特征到 KV 缓存的不同预计算层级，只对检索到的片段补全昂贵缓存，并用 CPU/GPU 流水并发隐藏查询计算；可持续 30 FPS 且 P95 首 token 延迟低于 0.5 秒，最高可达 57 FPS、P95 低于 2 秒。

Details

Motivation: 实时流式视频问答要求任意时刻到达的问题都能低延迟回答，同时持续处理应用要求的视频帧率。现有系统通常随视频到达就为全部内容构建 KV 缓存，但查询往往只用到此前少量帧，因此前处理与查询延迟存在权衡。

Result: StreamAhead 按视频片段独立编码，允许各片段停留在不同表示层级，从低成本视觉特征到高成本 KV 缓存；只有查询检索到的片段才在问题到达时补全缓存。系统离线测量片段推进到各层级的成本，运行时根据这些配置，在查询处理期间并发安排片段计算，并在异构 CPU/GPU 上流水化，同时维持视频处理速率。准确率与先进方法相当，首次 token 延迟最高比先进方法低 10.4 倍、比先进流系统低 5.9 倍；可维持 30 FPS 且 P95 延迟低于 0.5 秒，或达 57 FPS 且 P95 延迟低于 2 秒。

Insight: 按需预计算比对整段视频提前生成完整 KV 缓存更节省资源；需结合视频处理和查询执行动态调度以同时满足帧率与响应时延。摘要未给出不同查询密度下的成本或模型具体准确率。

Abstract: Real-time streaming video question answering with vision-language models must answer queries that arrive at any time with low latency while sustaining the video-processing rate required by the application. Existing systems construct key-value (KV) caches for the entire video as it arrives, although a query typically uses only a few of the preceding frames as evidence. This highlights a fundamental trade-off: performing more computation as video arrives reduces query-time computation but can increase processing lag, whereas deferring computation can increase the video-processing rate but requires additional query-time computation for retrieved content. We present StreamAhead, an algorithm-system co-design that exploits this trade-off through adaptive precomputation: dynamically controlling how much computation to complete for each video segment before a query arrives. StreamAhead encodes video segments independently and lets each remain at a different representation level, from inexpensive visual features to costly KV caches, completing KV caches only for the retrieved segments when a query arrives. StreamAhead profiles offline the cost of advancing a segment to each representation level. Its runtime uses these profiles to schedule segment computation concurrently with query execution, pipelining it across heterogeneous CPU and GPU resources while maintaining the required video-processing rate. StreamAhead achieves accuracy competitive with state-of-the-art methods with up to 10.4x lower time-to-first-token (TTFT), and up to 5.9x lower TTFT than state-of-the-art streaming systems. It sustains 30 FPS with P95 TTFT under 0.5 s, and up to 57 FPS with P95 TTFT under 2 s. To our knowledge, StreamAhead is the first real-time streaming video QA system with competitive accuracy, setting a new Pareto frontier between video-processing rate and TTFT.

cs.HC [Back]

[270] GazeAgent: VLM-Guided Implicit Gaze Calibration from Egocentric Multimodal Behavior for AI Glasses cs.HC | cs.CVPDF

Yuhong Zhang, Khanh An Nguyen, Xinyu Tian, KungYi Chang, Abraham Padua

TL;DR: GazeAgent 从 IMU、视线和佩戴者手部信号中识别自然发生的精确视觉注意事件，并将功能目标作为伪校准目标持续更新 AI 眼镜视线校准；10 人研究覆盖 820 个预设事件，设备受扰时中位视线误差比显式校准降低 44.1%。

Details

Motivation: AI 眼镜通常需要用户专属的显式眼动仪校准，设备在使用中移动后准确度还会漂移。若校准要求用户中断当前任务，交互负担较高。

Result: GazeAgent 是 VLM 引导的持续校准框架，从 IMU、眼动和佩戴者手部等多模态流识别精确视觉注意事件，验证其任务相关性，并把功能目标定位为伪校准目标，无需用户停下活动。在 10 名参与者的脚本研究中，系统检测到全部 820 个预设注意机会；设备受扰时，较显式校准中位视线误差降低 44.1%。另一个五人案例研究中，自动校准的工作负荷和中断评分最低，所有参与者都偏好日常使用自动校准。

Insight: 自然交互中已有的手部动作和注意线索可提供持续校准信号，减少专门校准步骤和设备位移造成的漂移。研究样本为 10 人和 5 人，摘要未报告设备型号、误差绝对值或长期佩戴效果。

Abstract: AI glasses typically require explicit, user-specific eye-tracker calibration, and accuracy can drift when the device shifts during use. We present GazeAgent, a VLM-guided framework that continuously and adaptively improves gaze-calibration accuracy from naturally occurring behavior without asking users to interrupt their activities. GazeAgent identifies precise visual attention (PVA) events from multimodal streams of inertial measurement unit (IMU), gaze, and wearer-hand signals, validates their task relevance, and localizes functional targets as pseudo-calibration targets. In a 10-participant scripted study, the system detected all 820 predefined PVA opportunities and, under device disturbance, reduced median gaze error by 44.1% compared with explicit calibration. In a five-participant case study, autonomous calibration produced the lowest workload and interruption ratings, and all participants preferred it for routine use. By deriving task-critical targets from natural multimodal behavior, GazeAgent demonstrates the potential for continuous, adaptive, and low-burden gaze interaction on egocentric AI glasses.

[271] Small Average Error, Wide Individual Error: General-Purpose LLMs Estimating Energy Intake from Meal Photos in a Feeding Trial cs.HC | cs.CV | q-bio.QMPDF

Xiaomeng Li

TL;DR: 在 802 餐受控膳食试验中，六种通用多模态模型估计“吃了多少能量”时，餐后照片对餐级误差仅再改善 2.2–3.4 分；没有模型满足预设的个体一致性界限，且所有模型都低估蛋白质。

Details

Motivation: 饮食能量照片估计可能减轻膳食评估负担，但多数评测只拍餐前照片并与供应能量比较，临床饮食评估实际需要估计摄入量。论文在有称重摄入真值的受控试验中比较餐前、餐后照片使用方式。

Result: 作者在 ACETADA 受控膳食试验的 802 餐上比较两家厂商六个模型，输入为餐前照片、分别调用餐前和餐后照片后相减，或一次同时输入两张照片。相减会使六个模型中的五个低估摄入量，因为即使托盘几乎空了，模型仍报告数百千卡。将餐前估计按平均吃掉比例 0.87 缩放，已获得餐后图像大部分收益；在此基础上，第二张图只再降低 2.2–3.4 分餐级误差，主要帮助有剩餐的情况。148 个膳食日中，三个模型的平均误差 90% 置信区间落在预设 10% 边界内，但没有模型满足预设 30% 一致性界限。所有模型均低估蛋白质，蛋白质密集餐最明显；重复调用误差变化最多 1.4 分；高推理预算只让最小模型提升 4.2 分，其余不到 1 分。

Insight: 较低平均偏差不等于个体摄入估计准确；作者明确认为这些模型应辅助而不能替代营养师评估。餐后照片主要改善有剩餐场景，蛋白质估计仍普遍偏低。

Abstract: Photo-based estimation of food energy could ease dietary assessment, and general-purpose multimodal large language models (LLMs) make such estimates easy to obtain. Most evaluations use one photo taken before eating and compare with energy served, while dietary assessment needs energy eaten. We evaluated six models from two vendors (Claude Haiku, Sonnet and Opus 5.5; GPT-6 luna, 6.1 sol and 6 astra) on 802 meals from ACETADA, a controlled feeding trial with weighed intake, with the before photo only, before and after photos in separate calls (subtraction), or both photos in one request. Three results ran against intuition. Subtraction under-estimated intake for five of six models, because the models still report hundreds of kilocalories for a nearly empty tray. Scaling the before-only estimate by the average share eaten (0.87) matched much of the gain from the after photo; beyond that constant, the second photo lowered meal-level error by only 2.2 to 3.4 points, mainly in meals with food left. With one photo, two models had a near-zero intake bias only because they under-estimated the served energy and ignored uneaten food. In ablations, adding the after photo improved accuracy under both prompt formats for five models, while prompt effects varied across models. Over 148 feeding days, three models kept the 90% confidence interval of their mean error within a pre-set 10% margin, but no model reached the pre-set 30% limits of agreement. Protein was under-estimated by every model, most in protein-dense meals; repeated calls changed error by at most 1.4 points; a high reasoning effort helped only the smallest model (4.2 points) and changed the others by less than 1 point. A low average bias is not evidence of accurate individual intake estimation; such estimates should support, not replace, assessment by a dietitian.

cs.IR [Back]

[272] Cross-Pollination Retrieval: Finding Transferable Ideas Across Fields in Scientific Literature cs.IR | cs.AI | cs.CLPDF

Seokwon Jung, Yerang Kim, Yejin Kim, Alexander Rubinstein, Arnas Uselis

TL;DR: CPR-Dataset 将跨领域可迁移思想检索形式化，并提供 26 个领域的 1,300 个测试查询与 10,573 个训练对；对比学习微调将 Recall@10 提至最高 0.70，并迁移改善类比和推理密集检索，在 BRIGHT 上 nDCG@10 提升最高 70%。

Details

Motivation: 许多科学进展来自把一个领域的思想迁移到另一领域，但跨学科论文检索尚未充分研究，现有基准只覆盖少数领域。论文提出跨领域思想迁移检索任务并构建大规模数据资源。

Result: CPR-Dataset 从跨主题引用中挖掘训练与评测样本，测试集 CPR-eval 含 26 个领域中的 1,300 个查询，训练集 CPR-train 含 10,573 对。CPR-eval 上，11 种现有检索器中最佳 Recall@10 仅为 0.26，LLM 重排器为 0.81；在 CPR-train 上标准对比微调可将 Recall@10 提高至最高 0.70。增益可迁移至训练未见领域和类比、推理密集检索：AnaloBench 上相较未训练检索器 nDCG@10 最高提高 33%，BRIGHT 上提高最高 70%，超过使用三种既有检索数据集训练的两倍收益。

Insight: 跨领域检索需要识别可迁移的思想，而非只匹配主题相近文本；专门构造的跨主题引用监督还可能提升类比和推理型检索。摘要未报告 LLM 重排器成本或不同领域的细分结果。

Abstract: Many scientific advances come from cross-pollination, carrying an idea from one field over to a problem in another. Yet retrieving papers across fields remains underexplored, with existing benchmarks confined to a narrow range of fields. We study cross-pollination retrieval, the problem of identifying transferable ideas across research fields, and introduce CPR-Dataset, a large-scale dataset constructed by mining cross-topic citations. Its evaluation split CPR-eval has 1,300 queries in 26 fields, and its training split CPR-train has 10,573 pairs. On CPR-eval, the best of eleven existing retrievers reaches a Recall@10 of only 0.26, whereas an LLM reranker reaches 0.81. Standard contrastive fine-tuning on CPR-train raises Recall@10 to as high as 0.70. The gain carries over to domains held out from training and to analogy and reasoning-intensive retrieval. On AnaloBench, it raises nDCG@10 by up to 33% over the untrained retriever, whereas training on three existing retrieval datasets lowers it. On BRIGHT, it raises nDCG@10 by up to 70%, more than twice the gain from those datasets. CPR-Dataset is available at https://huggingface.co/CPR-Dataset, and the code at https://github.com/SeokwonJung-Jay/Cross-Pollination-Retrieval.

[273] MAPS: A Multi-Agent Personalized Script-Driven Video Advertising System Enhanced by Reinforcement Learning cs.IR | cs.CV | cs.MAPDF

Zhongyu Wang, Chen Zuo, Qingxuan Sun, Jiayang Qiao, Qiuhao Qian

TL;DR: MAPS 用四个协作智能体组成用户建模、脚本生成、视频合成和反馈迭代闭环，并以多维强化学习奖励训练脚本智能体；快手特定场景的在线 A/B 实验较生成式基线 CTR 提高 2.87%、CVR 提高 2.74%、ARPU 提高 9.85%。

Details

Motivation: 常见个性化视频广告通过特征匹配，从预制创意中推荐相似候选，个性化粒度有限，也容易造成不同用户群效果差异。论文希望直接生成更贴合个人偏好和商品信息的广告创意。

Result: MAPS 将用户画像建模、广告脚本生成、视频合成和反馈驱动的迭代优化分配给四个协作智能体，形成闭环创意生成流程。脚本智能体通过强化学习训练，使用多维奖励同时引导脚本匹配用户内容偏好与目标商品信息。代表性数据集的离线实验显示个性化创意生成和整合表现积极；快手广告系统一个特定场景的在线 A/B 测试中，较生成式基线 CTR 提升 2.87%、CVR 提升 2.74%、ARPU 提升 9.85%。

Insight: 将创意生成拆成画像、脚本、视频和反馈优化阶段，使个性化能作用于广告内容本身，而不只是在预制候选中排序。在线收益来自特定广告场景，摘要未说明其他场景或用户群的效果。

Abstract: With the rapid growth of online video consumption, video advertising has become a dominant component of digital advertising. Most personalized video advertising systems follow a retrieval-based paradigm, which matches ad content with user preferences in a feature space and recommends the most similar candidates. However, since the candidate creatives are pre-produced, this paradigm limits the granularity of personalization and leads to performance disparities across diverse user groups. To address this limitation, we propose a multi-agent personalized script-driven video advertising system (MAPS). We design four collaborative agents that form a closed-loop pipeline of user profile modeling, script generation, video synthesis, and feedback-driven iterative refinement. To train the Scripter Agent, we develop a reinforcement learning framework equipped with a multi-dimensional reward function, which steers the agent toward producing scripts that simultaneously align with user content preferences and target product information. Extensive offline experiments on representative datasets demonstrate that our approach achieves encouraging performance in generating and integrating personalized creatives. Online A/B experiments in a specific scenario in the Kuaishou advertising system show that MAPS improves CTR by 2.87%, CVR by 2.74%, and ARPU by 9.85% over the generative baseline.

cs.IT [Back]

[274] Dependency, Compression, and Synergy: A Unified Information-Theoretic View of Multimodal Learning cs.IT | cs.AI | cs.CV | cs.LGPDF

Liangjian Wen, Linjie Li, Jiang Duan, Yong Dai, Jianzhuang Liu

TL;DR: 该综述将互信息、信息瓶颈和部分信息分解统一为多模态信息处理视角：分别刻画模态依赖、任务导向压缩下的信息保留，以及冗余、独有与协同信息；它整理 170 篇近年研究和 12 篇基础工作并提出统一信息拉格朗日坐标系。

Details

Motivation: 多模态基础模型的发展使理解模态如何共享、保留和互补信息更重要，但已有研究常把互信息（MI）、信息瓶颈（IB）和部分信息分解（PID）视为互不相干的工具。

Result: 综述把 MI、IB、PID 视为逐步细化的信息处理视角：MI 描述模态间依赖，IB 解释压缩时如何保留任务相关信息，PID 将保留信息分成冗余、独有和协同部分。作者回顾 170 篇 2018–2026 年研究和 12 篇基础工作，按跨模态对齐、信息高效融合、交互类型刻画及多模态基础模型扩展四项挑战组织文献，并把医疗、机器人、推荐、情感计算和无线通信等应用视作信息原理的实证场景。综述进一步提出 Generalized Multimodal Information Lagrangian 统一坐标系，将现有多模态范式视为精确或近似参数特例，并标出尚未探索的参数区域及候选方法族。

Insight: 统一的信息论坐标系可帮助比较多模态学习方法在依赖、压缩和协同信息上的取舍，而不是按应用领域孤立分类。摘要指出的开放问题包括高维信息估计扩展性、标准化评测、PID 组合复杂度，以及从事后分析转向信息感知训练。

Abstract: Recent advances in multimodal foundation models have intensified the need to understand how different modalities share, preserve, and complement information. Mutual Information (MI), the Information Bottleneck (IB), and Partial Information Decomposition (PID) provide complementary perspectives, yet existing studies often treat them as isolated tools. This survey presents an information-theoretic perspective connecting these principles as progressively refined views of multimodal information processing: MI characterizes inter-modal dependency, IB explains task-oriented information preservation under compression, and PID decomposes preserved information into redundancy, uniqueness, and synergy. We review 170 recent studies (2018–2026) and 12 foundational works, organizing multimodal learning around four challenges: cross-modal alignment, information-efficient fusion, interaction-type characterization, and scaling to multimodal foundation models. Rather than using application domains as primary taxonomy axes, we interpret healthcare, robotics, recommendation systems, affective computing, and wireless communications as empirical validations of these information principles. Beyond taxonomy, we organize existing multimodal paradigms within a single information-theoretic coordinate system – the Generalized Multimodal Information Lagrangian – in which they occupy exact or approximate parameter corners, and whose unoccupied regions name candidate method families the literature has not yet built. We further discuss how emerging multimodal foundation models instantiate these principles at scale and identify open challenges including scalable information estimation in high-dimensional settings, standardized evaluation across information-theoretic methods, combinatorial complexity of multimodal PID, and the transition from post-hoc information analysis toward information-aware multimodal learning.

cs.LG [Back]

[275] SAGE-CP: When Marginal Coverage Hides Route-Conditional Failures in Adaptive VLM Ensembles cs.LG | cs.CVPDF

Mushir Akhtar, M. Tanveer

TL;DR: SAGE-CP 将自适应 VLM 集成的共形校准按早停与升级路径分别进行，避免边际覆盖掩盖某一路径严重失准；10 个基准、4 个冻结 VLM 上，最差路径覆盖率从 0.623 提至 0.887，路径间覆盖差距从 0.324 降至 0.025。

Details

Motivation: 自适应 VLM 集成可对简单图像提前停止、对困难样本升级模型以省计算。单一共形阈值虽能保证整体边际覆盖，却可能让个别计算路径的条件覆盖率很低，造成实际部署风险。

Result: SAGE-CP 先用开发集训练路由器并冻结所有路由决策，再对提前退出和升级输入分别校准返回分数；校准采用针对路由所形成操作组的 Mondrian split conformal prediction，在可交换假设下为每条路径提供有限样本覆盖率保证。跨 10 个基准、4 个冻结 VLM、5 个对齐划分，整体目标覆盖率 0.90 时，普通全局校准和 SAGE-CP 均实现 0.900 边际覆盖，但路径可靠性不同：SAGE-CP 将平均最差路径覆盖率从 0.623 提至 0.887，将平均路径覆盖差从 0.324 降至 0.025，源模型调用数相同，为每张图 1.177 次。预测集平均大小从 1.893 增至 2.212，仍小于完整集成的 2.486，同时源调用少 70.6%。相较完整推理，五个数据集上 batch-one 延迟降低 58.8–63.4%，GPU 能耗降低 55.7–57.5%；更细的退出深度组在稀有路径上需要更多校准数据。

Insight: 整体覆盖保证无法自动保护路由产生的每一种操作路径；需在冻结路由后分组校准并报告最差路径可靠性。更强路径保证会扩大预测集合，而罕见路径仍受校准样本量限制。

Abstract: Adaptive VLM ensembles save computation by stopping early on easy images and escalating harder ones. Although a single conformal threshold can make the final predictor marginally valid, it can still fail severely along individual computational routes. We characterize this failure precisely: if a route has prevalence pi, marginal 1-alpha coverage allows its conditional coverage to fall to max{0, 1-alpha/pi}. To address this mismatch, we learn the router on development data, freeze all routing choices, and calibrate the returned score separately for early exits and escalated inputs. We call this pipeline SAGE-CP. Its calibration step applies Mondrian split conformal prediction to the operational groups created by the router. Under exchangeability, the resulting prediction sets achieve finite-sample coverage within each route. Across 10 benchmarks, 4 frozen VLMs, and 5 aligned splits, both global calibration and SAGE-CP achieve 0.900 marginal coverage at a 0.90 target, but their route behavior differs substantially. SAGE-CP raises mean worst-route coverage from 0.623 to 0.887 and reduces the mean route-coverage gap from 0.324 to 0.025, using the same routed scores and 1.177 source calls per image. The stronger guarantee increases the mean prediction-set size from 1.893 to 2.212; nevertheless, the sets remain smaller than those of the full ensemble (2.486) while requiring 70.6 percent fewer source calls. The route-level reliability persists across 6 fixed query fractions, while finer exit-depth groups demonstrate the need for additional calibration data on rare routes. Across 5 datasets, SAGE-CP also reduces measured batch-one latency by 58.8 to 63.4 percent and gross GPU energy by 55.7 to 57.5 percent relative to full inference. Thus, routing provides computational savings, while route-stratified calibration attaches coverage guarantees to the operational paths created by the router.

[276] How Data Scales in Agentic Reinforcement Learning: Laws and Synthesis Strategies cs.LG | cs.CLPDF

Bowei He, Yankai Chen, Xiaokun Zhang, Changjiang Han, Ye Yuan

TL;DR: 论文把智能体强化学习数据拆成任务、环境和轨迹三个可扩展维度，并拟合多模型、多领域的缩放关系；实测环境规模的缩放效应强于任务、任务又强于轨迹，按推导的预算配方训练可比主流“只扩轨迹”方案提高最高 1.7 倍计算效率。

Details

Motivation: 预训练缩放通常把数据简化为 token 数，但智能体强化学习的数据管线可以分别合成任务、环境和交互轨迹，且各自成本不同。给定合成预算，应该扩大哪类数据成为新的设计问题。

Result: 研究将智能体数据分为任务数、环境数、轨迹数三个轴，并将奖励密度作为控制变量；在数学与代码、软件工程、网页及 UI/工具应用领域，对不同语言模型拟合各轴缩放关系。所测模型规模、模型家族、领域及 PPO/GRPO 算法中，指数排序均为环境大于任务、大于轨迹。拆分环境多样性后，优势主要来自世界本身的多样性，环境结构多样性则约与任务多样性相当。作者进一步为数据合成方法估算各轴效率折扣和合成/真实比例坍缩阈值，并据此构造受约束的计算最优预算配方，在留出规模验证，最高获得 1.7 倍计算效率。

Insight: 智能体数据的“量”是多维的，环境多样性可能比堆积更多交互轨迹更有效；且合成数据风险阈值随数据轴相差一个数量级，不能用统一比例控制。报告的最高效率提升是在所测配方和实践对照范围内。

Abstract: Pretraining scaling laws treat training data as a one-dimensional quantity: a token count. Agentic reinforcement learning (RL) inherits this scalar abstraction, but modern synthesis pipelines now generate \textit{tasks}, \textit{environments}, and \textit{trajectories} independently and at very different costs, turning data into a structured design space. This raises a question pretraining never had to answer: given a synthesis budget, \textit{which axis} should be scaled? We address this question end-to-end with three tightly coupled contributions. First, we decompose agentic data into three cardinality axes (task, environment, trajectory), plus a reward-density control variable, and fit per-axis scaling relations on different large language models (LLMs) across math+code, SWE, web, and UI/tool application domains. Across all model sizes, model families, domains, and RL algorithms (PPO and GRPO) we measured, the scaling exponents order as environment $>$ task $>$ trajectory. When environment diversity is decomposed into world and environment-structure diversity, its advantage is mainly driven by the former, while the latter is just comparable to the task diversity. Second, we use the fitted laws as a common ruler to price synthesis methods, capturing each method’s efficiency as a per-axis discount factor and identifying critical synthetic-to-real ratios at which collapse begins; these ratios differ by an order of magnitude across axes. Third, we cast budget allocation as a constrained optimization under our laws and discounts, deriving a Chinchilla-style compute-optimal recipe and validating it at held-out scale. The empirical results demonstrate that following the recipe yields up to $1.7\times$ compute efficiency over the dominant ``scale-the-trajectories’’ practice.

[277] DISCERN: Identifiable Disease Evidence for Generalizable Medical Multimodal Learning cs.LG | cs.CV | q-bio.QMPDF

Xiaoguang Zhu, Pengyi Jiang, Juncen Guo, Jingyi Wu, Lianlong Sun

TL;DR: DISCERN 将跨队列医疗多模态偏差建模为潜在结构因果偏移，并从跨队列状态对齐和自然参数变化识别最小选择统计量；合成反向入组实验的状态 AUROC 从 0.043/0.265 提升至 0.883/0.825，eICU 与 MIMIC-IV 死亡率迁移较 ERM 高 0.036/0.035。

Details

Motivation: 临床多模态模型可能在多个模态中共同学习入组捷径，在新队列上失效。论文将数据表示为潜在结构因果模型：疾病状态和背景因素经共享机制产生生理观测，而不同队列在样本选择和测量方式上存在差异。

Result: 在采集方式固定的假设下，DISCERN 不需观察疾病状态标签或恢复完整因果模型，也可识别条件疾病证据。基于共同高斯坐标，跨队列状态对齐和自然参数变化可识别最小选择统计量；对其条件化可去除依赖入组的证据，并在指定偏移族内得到可识别参考后验。状态相关交互可能携带单模态后验看不到的证据，因此框架结合信息格编码器、潜状态模型和结局头联合训练。已知潜变量坐标的反向合成入组实验中，条件化使两组状态 AUROC 分别从 0.043、0.265 升到 0.883、0.825，随机条件化无此效果。eICU 与 MIMIC-IV 跨数据库死亡率探索实验中，AUROC 较匹配 ERM 高 0.036 和 0.035。

Insight: 跨队列泛化需要区分疾病状态信息与队列入组/测量机制；联合多模态交互也可能保留单模态后验无法体现的疾病证据。真实数据库结果是探索性比较，摘要未证明所有坐标或偏移设定下都可识别。

Abstract: Clinical multimodal models can learn enrollment shortcuts shared across modalities and fail on new cohorts. We model these data through a latent structural causal model: disease state and background factors generate physiology through shared mechanisms, while cohorts differ in selection and measurement. Holding acquisition fixed, we identify conditional disease evidence without observing disease-state labels or recovering the full causal model. In common Gaussian coordinates, cross-cohort state alignment and natural-parameter changes identify a minimal selection statistic; conditioning on it removes enrollment-dependent evidence and yields an identifiable reference posterior within the specified shift family. State-dependent interactions can carry evidence invisible to individual-modality posteriors, motivating joint, information-preserving representations. DISCERN combines an information-lattice encoder, a latent-state model, and an outcome head through joint training. Under reversed synthetic enrollment on known latent coordinates, conditioning raises state AUROC from 0.043 to 0.883 and from 0.265 to 0.825, while random conditioning does not. Exploratory cross-database mortality experiments between eICU and MIMIC-IV improve AUROC over matched empirical risk minimization by 0.036 and 0.035. Learned-coordinate diagnostics distinguish preservation of state information from identification of the coordinates required by the estimator.

[278] Inside the Loop: A Mechanistic Study of Weight-Tied Transformers on Depth-Bound Algorithmic Tasks cs.LG | cs.CLPDF

Magnus Sesodia, Juil Sock, Philip Torr, Christian Schroeder de Witt

TL;DR: 在有逐跳真值的置换遍历任务上，循环 Transformer 表现得像迭代程序学习器：四层模块循环八次时所有随机种子均解出任务，而同有效深度、相同训练方式的非共享权重模型一次也未成功。

Details

Motivation: 逐步算法任务通常需要增加 Transformer 深度或延长思维链，分别提高参数规模或推理计算。循环 Transformer 反复使用同一层模块，可以不增加参数和 token 而提高有效深度，但每轮到底计算什么、它是否只是模拟更深的普通堆叠仍不清楚。

Result: 作者研究有明确中间真值的有向置换遍历：模型从起点沿 H 跳前进并报告终点。结果显示任务需要最低有效深度，超过后权重共享程度成为关键因素。四层模块循环八次的模型在每个随机种子均成功；相同有效深度、相同训练方式的非共享模型从未成功，只有降低学习率或监督其中间层后才可解。对残差流的探测和干预显示，循环模型逐轮推进中间节点，形成随迭代预算调整的阶梯；降低学习率的非共享模型则只在最后几层显露中间节点。

Insight: 循环结构可能通过在每次复用中执行算法的下一步来学习迭代程序，而非仅以较少参数模拟普通深层网络。实验限定于置换遍历任务，摘要未报告自然语言推理或其他算法任务的验证。

Abstract: Solving algorithmic tasks with many sequential steps typically requires scaling up transformer depth or chain-of-thought length, incurring unfavorable parameter and inference-compute costs, respectively. Looped transformers, which reuse one block of layers across many iterations, are poised to address this, since they scale effective depth without adding parameters or tokens. Yet, it remains unclear what each iteration is actually computing, or whether the loop merely simulates a deeper untied stack. Here, we study directed permutation traversal, in which a model follows $H$ hops from a starting node and reports the node it reaches, so every intermediate hop has a known ground truth. We find that a minimum effective depth is necessary to solve the task, beyond which the degree of weight tying becomes the determining factor. For instance, a looped transformer with a four-layer block iterated eight times solves the task in every seed, while an identically-trained untied transformer of the same effective depth never does, succeeding only with a lower learning rate or supervision of its intermediate layers. Probing and intervening on the residual stream, we find that the looped model advances through the intermediate nodes gradually, forming a staircase that adapts to its iteration budget, whereas the untied model, at that lower rate, exposes them only in its final layers. Taken together, our results recast looped transformers as iterative program learners, which allocate their iterations to successive steps of the algorithm and solve tasks that identically-trained untied models of the same effective depth do not. Code is available at https://github.com/magnus-sesodia/insidetheloop.

[279] Continual Test-Time Learning with Evolution Strategies cs.LG | cs.CLPDF

Guangya Hao, Yiting Duan, Longyan Tan, Yunbo Long, Hao Wu

TL;DR: ESRL 用锚点策略附近的一组扰动策略共同产生无标签组相对学习信号，避免持续测试时适应只依赖正在变化的单一策略而导致监督漂移；在数学、常识和医学任务上平均准确率较基础模型最高提高 31.7%，较最强测试时适应基线提高 16.6%，并减轻灾难性遗忘。

Details

Motivation: 开放环境中的语言模型需持续适应变化任务，但标签、示范或奖励等外部监督可能昂贵或不可用。测试时适应可用模型自生成监督，但若监督也由正在更新的单一策略产生，策略漂移会使学习信号跟着变差，最终导致崩溃、推理退化和遗忘。

Result: Evolution Strategy Reinforcement Learning（ESRL）在参数空间扰动锚点策略形成一组邻近策略，每个策略采样一条轨迹，再汇总答案形成无标签组相对信号来训练中心学生。锚点在每个训练块内固定，并在块边界依据学生刷新，使策略群围绕稳定且不过时的中心提供多样监督。单任务测试时适应覆盖数学推理、一般知识和医学领域，平均准确率相对基础模型最多提升 31.7%，相对最强测试时适应基线最多提升 16.6%；持续学习实验中显著减少灾难性遗忘，并取得最高最终表现。

Insight: 多个邻近策略的群体信号可降低单个自适应策略同时充当学习者和监督者所带来的反馈漂移。报告的最高增幅跨任务平均，摘要未提供各领域分项结果、扰动群体计算开销或持续学习任务序列细节。

Abstract: Large language models increasingly operate in open-ended environments, where they are expected to continually adapt to evolving tasks and distributions, often without access to external supervision. Existing continual learning approaches rely on external supervision, such as labeled data, demonstrations, or reward signals, which can be costly or unavailable in practice. Test-time adaptation removes this dependence by learning from self-generated supervision, but existing methods derive that supervision from the single policy being updated, so once the policy drifts, its own learning signal drifts with it, leading to policy collapse, degraded reasoning, and catastrophic forgetting. In this paper, we propose \textbf{Evolution Strategy Reinforcement Learning (ESRL)}, a continual test-time learning framework that draws its self-generated supervision from a population of policies instead of the single policy being trained. ESRL perturbs an anchor policy in parameter space to form a population of nearby policies, samples one trajectory from each, and aggregates their answers into label-free group-relative learning signals for the central student. The anchor is held fixed within each block and refreshed from the student at block boundaries, so the population stays centered on a stable yet current policy and continues to supply diverse and reliable supervision as the student moves. In single-task test-time adaptation across mathematical reasoning, general knowledge, and medical domains, ESRL improves average accuracy by up to 31.7% relative to the base model and 16.6% relative to the strongest test-time adaptation baseline. Notably, ESRL substantially reduces catastrophic forgetting and achieves the strongest final performance in continual learning.

[280] ProbePO: Sparse Probes, Redistributed Credit for Critic-Free LLM Reinforcement Learning cs.LG | cs.CLPDF

Diyi Hu, Ruizhong Qiu, Lin Qiu, Hanghang Tong, Hanqing Zeng

TL;DR: ProbePO 用有限的续写预算挑选不确定的部分推理状态进行稀疏验证，再将探针价值重新分配到原轨迹 token，同时保持轨迹平均优势不变；在 18 个模型—基准组合中有 16 个取得四种方法中的最高准确率，并比 VinePPO 少 76% 端到端训练时间。

Details

Motivation: 可验证奖励强化学习通常只给整条推理轨迹一个终端奖励，无法指出哪些步骤促成结果。GRPO 将组归一化优势广播到所有响应 token；更细粒度方法则可能需要密集额外 rollout 或分支，并改变用于策略更新的数据。

Result: ProbePO 将中间环节视为预算受限测量：根据不依赖奖励的不确定性分数选取部分解题状态，在受限续写预算内估计其期望验证器奖励。随后，它重新分配而非替换 GRPO 的轨迹级优势，并保持每条轨迹 token 级优势的均值不变；探针续写只用于测量。在 1.5B–7B 的 DeepSeekMath 和 Qwen 系列三个模型上，ProbePO 一贯优于 GRPO，并在 18 个模型—基准比较中的 16 个获得四种被评方法里最高准确率。对 DeepSeekMath-7B，ProbePO 在六个基准上全部高于密集探针 VinePPO，端到端训练时间减少 76%。

Insight: 稀疏测量中间步骤的价值可以改善信用分配，而不必把额外探针 rollout 混入策略训练数据或改变轨迹整体优势。摘要未解释其余两个比较未夺冠的具体结果，也未给出推理/验证额外开销。

Abstract: Reinforcement learning from verifiable rewards assigns one terminal reward to an entire reasoning trajectory, obscuring how individual reasoning steps contributed to the outcome. GRPO broadcasts a group-normalized advantage to every response token, while finer-grained approaches often rely on dense auxiliary rollouts or branching schemes that can also alter the data used for policy updates. We cast the missing middle as budgeted measurement: with a limited rollout budget, which partial solutions should be evaluated, and how should their value estimates refine credit? ProbePO selects partial-solution states with a reward-free uncertainty score and estimates their expected verifier rewards under a capped continuation budget. ProbePO redistributes, rather than replaces, GRPO’s trajectory-level advantage across tokens while preserving each trajectory’s mean token-level advantage; probe continuations serve only as measurements. Across three DeepSeekMath and Qwen-family models spanning 1.5B-7B parameters, ProbePO consistently outperforms GRPO and achieves the highest accuracy among the four evaluated methods in 16 of 18 model-benchmark comparisons. Against dense-probing VinePPO on DeepSeekMath-7B, ProbePO achieves higher accuracy on all six benchmarks with 76% less end-to-end training time.

[281] Structured Latent Dynamics for Long-Horizon Embodied World Models cs.LG | cs.CVPDF

Kevin Chen, Yu-Wei Zhan, Haoran Liang, Sen Cui, Changshui Zhang

TL;DR: HWM 将世界模型潜变量拆为广义坐标与动量，并以耗散 Hamilton 动力学传播未来状态，再映射回原潜空间；受控模拟中速度、加速度、急动度估计更准确且长程能量漂移降低，RoboCasa 成功率相对提升 2.7%。

Details

Motivation: 具身世界模型可生成视觉上合理的动作条件未来，但无约束潜空间转移在递归 rollout 中可能逐渐偏离连贯物理演化，影响长时预测稳定性与规划价值。

Result: Structured Latent Dynamics（HWM）把骨干模型潜表示转换为学习的相空间，包括广义坐标和动量；未来状态通过耗散 Hamilton 动力学传播，再由可学习逆接口映射回原始潜表示。该结构可接入现有世界模型，不必重做感知或解码模块，多步转移监督使 rollout 对齐真实观察潜轨迹。受控模拟中，HWM 更准确估计速度、加速度和急动度，并显著减少长程能量漂移。进一步在 Ctrl-World 单臂多视角、EVAC 双臂单视角和 Cosmos-Predict2.5-2B 任务导向单臂等三类世界模型评测：RoboCasa 上任务成功率相对提高 2.7%；真实 AgiBot World 轨迹上，Ctrl-World 的 nDTW 提高 9.7%，EVAC 的 Logic 指标提高 8.4%。

Insight: 为潜在状态显式加入坐标—动量结构和耗散动力学，可为多步世界预测增加物理稳定性约束，同时保留原模型接口。摘要未给各模型全部任务结果或仿真物理指标数值。

Abstract: Reliable long-horizon prediction remains a central challenge for embodied world models. Although recent generative models can synthesize visually plausible action-conditioned futures, their latent dynamics are typically learned through unconstrained transition modules, which may gradually deviate from coherent physical evolution as predictions are recursively rolled forward. We investigate whether explicitly structuring latent state evolution can improve the stability and utility of such imagined trajectories. To this end, we introduce HWM, a structured dynamics framework that represents backbone latents in a learned phase space consisting of generalized coordinates and momenta. Future states are propagated through dissipative Hamiltonian dynamics and mapped back to the original latent representation through a learnable inverse interface. The resulting formulation can be integrated into existing world-model architectures without redesigning their perception or decoding components, while multi-step transition supervision encourages the structured rollout to remain aligned with the latent trajectories induced by real observations. Controlled simulation experiments show that HWM produces more accurate estimates of velocity, acceleration, and jerk and substantially reduces long-horizon energy drift relative to unstructured latent dynamics. We further evaluate the approach across three distinct embodied world-model settings, covering single-arm multi-view prediction with Ctrl-World, dual-arm single-view prediction with EVAC, and task-oriented single-arm prediction with Cosmos-Predict2.5-2B. On RoboCasa, incorporating HWM into Cosmos-Predict2.5-2B yields a 2.7% relative improvement in task success. On real-world AgiBot World trajectories, it improves nDTW by 9.7% with Ctrl-World and Logic by 8.4% with EVAC.

[282] Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design cs.LG | cs.AI | cs.CLPDF

Boyi Liu, Yite Wang, Xinyu Lian, Fan Shu, Sida Li

TL;DR: Arctic-SWE 通过模型—系统协同构建异步智能体强化学习基线，使用小任务池过拟合测试揭露早期奖励掩盖的问题；仅用原始 R2E-Gym 训练 Qwen3.5-4B 即在 SWE-bench Verified 达到 62.0% pass@1，约为 FrogNano rollout 数的六分之一。

Details

Motivation: 软件工程等长时智能体强化学习快速发展，但强基线的训练配方和基础设施难以复现，使新算法或数据的收益可能只是弥补了基线调优不足。论文希望建立可复现的异步 RL 基础，并分辨任务训练成绩与留出泛化。

Result: Arctic-SWE 是全异步 RL 框架，以小任务池过拟合作为测试平台，整合门控奖励、重要性权重控制、训练与推理引擎对齐及策略更新间状态一致性等组件。使用原始 R2E-Gym、没有新算法或数据处理时，它将 Qwen3.5-4B 训练到 SWE-bench Verified 62.0% pass@1，与依赖在线任务合成的 FrogNano 61.5% 相当，rollout 约少六倍。进一步筛选和整理 R2E-Gym、SWE-rebench V2、Scale-SWE 后，4B 在 Verified/Pro-v1/Pro-v2 上由 47.4%/25.0%/37.9% 升至 65.0%/38.3%/59.7%；9B 则由 56.2%/31.9%/48.8% 升至 65.7%/39.6%/70.2%。受控比较显示数据整理和加强测试均改善留出表现，但更强测试会降低表面上的训练任务掌握度。

Insight: 长时智能体 RL 的成绩高度依赖训练与推理基础设施以及任务池测试强度，不能把早期奖励提升直接归因于算法进步。论文分别报告原始数据基线和整理数据结果，摘要未列训练总时长或资源配置。

Abstract: Long-horizon agent training, including software engineering, is advancing rapidly, but the recipe and infrastructure needed for a strong reinforcement-learning baseline remain difficult to reproduce. This makes gains from new algorithms or data hard to distinguish from improvements to an under-tuned baseline. We present Arctic-SWE, a fully asynchronous RL framework that establishes such a baseline through model-system co-design, using overfitting on a small task pool as the testbed. This testbed exposes failures that early reward gains conceal and guides the integration of established components: gated reward, importance-weight control, aligned training and inference engines, and consistent state across policy updates. No single component is new, yet together they form the missing foundation. On raw R2E-Gym data, without new algorithms or data work, Arctic-SWE trains Qwen3.5-4B to 62.0% pass@1 on SWE-bench Verified. This is comparable to concurrent FrogNano (61.5%), which relies on online task synthesis, after about one-sixth as many training rollouts. Building on this foundation, we further filter and refine R2E-Gym, SWE-rebench V2, and Scale-SWE. With this curated subset, the 4B model improves from 47.4%/25.0%/37.9% to 65.0%/38.3%/59.7% on SWE-bench Verified/Pro-v1/Pro-v2, and the recipe also improves Qwen3.5-9B from 56.2%/31.9%/48.8% to 65.7%/39.6%/70.2% on Verified/Pro-v1/Pro-v2. Controlled comparisons show that both curation and test strengthening improve held-out performance, although stronger tests reduce apparent training-task mastery. Arctic-SWE provides a documented starting point for studying better algorithms, data and environments for asynchronous agent RL.

[283] CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning cs.LG | cs.AI | cs.CLPDF

Hao Wang, Zihan Wang, Kai Wu, Jian Yao, Yiwen Ye

TL;DR: CasePlay 从临床病例报告构造医疗推理自博弈训练信号，以病例知识点引导提问并用量规同时评估答案与问题质量；七个医学基准上优于自博弈基线，较 SFT 热启动模型平均提升 3.1 分，LiveClin 提升 6.3 分。

Details

Motivation: 临床推理强化学习通常依赖专家标注问题或成本较高的合成数据，而病例报告包含从初始表现到治疗决策的高质量临床过程。既有自博弈容易反复围绕少数显著线索提问，也可能生成难度合适但浅显或设计不佳的问题。

Result: CasePlay 包含知识条件提问生成器和量规评审推理器：前者将问题锚定在病例特定知识点上，扩大对来源报告证据的覆盖；后者除答案准确性外，还评估线索整合、推理深度和选项质量，以促使提问更适合训练推理器。七个医学相关基准实验中，CasePlay 优于已有自博弈基线；相较 SFT 热启动模型，平均提高 3.1 分，其中 LiveClin 提高 6.3 分。

Insight: 未标注病例文献可以转化为持续强化临床推理的训练资源，但自博弈的质量取决于问题是否覆盖多样证据，以及难度是否与推理深度匹配。摘要未报告各基准分项结果、成本或病例来源覆盖范围。

Abstract: Reinforcement learning (RL) is a promising way to improve the clinical reasoning abilities of large language models (LLMs), but current pipelines often rely on expert-labeled questions or costly synthetic-data generation. Meanwhile, the medical literature offers a wealth of high-quality unlabeled documents, including clinical case reports that trace the full clinical course from initial presentation to final treatment decisions. We study whether document-grounded self-play can convert these reports into effective RL training signals. We identify two challenges that limit existing self-play methods in this setting: knowledge-point fixation}, where the proposer repeatedly asks about a narrow set of salient clues in a source report, and difficulty-quality mismatch, where appropriately difficult questions may still be poorly formed or shallow, limiting their value for reasoner training. To address these challenges, we propose CasePlay, a self-play framework with two key components: (a) a Knowledge-Conditioned Proposer, which anchors question generation to case-specific knowledge points so that questions draw on a broader range of evidence from the source report; and (b) a Rubric-Judge Reasoner, which augments answer-accuracy feedback with rubric-based quality feedback on clue integration, reasoning depth, and option quality, guiding the proposer toward reasoning-intensive questions that better support reasoner training. Extensive experiments across seven medically related benchmarks demonstrate that CasePlay outperforms existing self-play baselines and improves the SFT warm-start model by 3.1 points on average, highlighted by a 6.3-point gain on LiveClin. This approach offers an effective path toward continually improving medical reasoning ability through self-play RL.

[284] Beyond the Teacher: Understanding and Preserving Distillation Overshoot in Agentic On-Policy Distillation cs.LG | cs.CLPDF

Jike Zhong, Ming Li, Yuxiang Lai, Xi Xiao, Sheng Zhang

TL;DR: SPOD 针对 on-policy 蒸馏中学生早期超过教师、随后因教师抑制学生正确选择而退化的 overshoot 现象，选择性屏蔽这类冲突监督并设置软概率下限；三个智能体基准中最终成功率最高提升 15.2 个百分点，且比教师高至 20.0 个百分点。

Details

Motivation: On-policy distillation 让智能体沿自身轨迹接受更强教师监督。作者发现多轮任务中，起初较弱的学生可能在训练中超过教师，继续蒸馏却又回退；原因可能是教师修复学生错误的同时，也磨损教师不认可但实际成功的学生决策。

Result: 作者将早期学生超越教师、后续性能退化称为 distillation overshoot，并分析修复弱决策与侵蚀成功决策这两种竞争效应。Success-Preserving On-Policy Distillation（SPOD）对学生成功但与教师指导冲突的选择，选择性屏蔽教师监督并加入软概率下限，以减缓侵蚀，同时保留其他位置的教师引导。该方法只需二元 episode 成功标记及蒸馏过程已有模型概率。在两个模型家族、三个智能体基准的六种设置中，SPOD 均优于普通 OPD 和所测变体，最终成功率最高提高 15.2 个百分点、超过教师最高 20.0 个百分点；六种设置的最终检查点都超过普通 OPD 最佳中间检查点。

Insight: 教师并非对所有成功学生决策都更好，持续蒸馏可能把教师不偏好的有效行为消除；仅对成功冲突决策保护可保留并继续扩大超教师收益。摘要未给出模型规模及各任务的具体收益。

Abstract: On-policy distillation (OPD) trains agents on their own trajectories using supervision from a stronger teacher. We uncover a counterintuitive phenomenon in multi-turn agentic tasks: a smaller, initially weaker student can substantially outperform its teacher during training, only to regress as distillation continues. We term this phenomenon distillation overshoot. Through decision- and task-level analyses, we trace overshoot to two competing effects of OPD: repairing weak student decisions and eroding successful student choices that the teacher disfavors. Early in training, repair outpaces erosion, allowing the student to combine the teacher’s corrections with its own strengths and surpass the teacher. As training continues, erosion overtakes repair, causing these gains to vanish. Motivated by this finding, we propose Success-Preserving On-Policy Distillation (SPOD), which selectively masks conflicting teacher supervision on successful student choices and introduces a soft probability floor to counteract their erosion, while retaining teacher guidance elsewhere. SPOD requires only binary episode success and model probabilities already available during OPD. Across two model families and three agentic benchmarks, SPOD consistently outperforms ordinary OPD and evaluated variants, improving final success by up to 15.2 percentage points and surpassing the teacher by up to 20.0 points. Moreover, SPOD’s final checkpoints outperform even ordinary OPD’s best intermediate checkpoints in all six settings, demonstrating that beyond-teacher gains can be not only preserved but further extended. Code: https://github.com/Jike338/SPOD

[285] Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of “Unsafe” You Trained On cs.LG | cs.AI | cs.CL | cs.CRPDF

Vahid Jebraeeli, Jonathon Romero, Yijin Wang, Adrian Chan, Jianbang Zhang

TL;DR: 研究指出安全基准分数可能主要反映训练标签采用的“不安全”定义，而非模型检测能力：固定架构下 HarmBench-copyright 在十种护栏训练配置间分数跨度达 0–100，OR-Bench-hard 过度拒答率跨度为 3.4%–56.4%，仅约 0.03% 参数即可改变分数范围。

Details

Motivation: 护栏分类器常被当作判断文本是否不安全的能力模型，但标签实际承载具体操作政策，例如“请求有害内容”与“审核员会拦截”并不相同。论文利用同一批约 5.3 万提示在两种定义下被标注并训练的十个模型配置，审计基准结果是否受政策标签、数据污染和蒸馏软标签影响。

Result: 作者先发现两个过度拒答基准存在训练污染：JailbreakBench-benign 有 88%、OR-Bench-hard 有 82% 出现在训练池，其余报告均去污染。固定架构与优化器时，十种训练臂的基准分数跨度依基准不同：HarmBench-copyright 为 0.99、WildGuard-benign 为 0.043，相差约 23 倍；约 0.03% 参数即可让分数跨越整段范围。HarmBench-copyright 覆盖 0–100，OR-Bench-hard 过度拒答率覆盖 3.4%–56.4%，并随最后监督信号携带的政策定义分处两端。蒸馏学生即使硬标签采用另一政策，教师软标签仍可带入其政策。唯一评估所有模型的验证集准确率与标注者政策一致性相关系数为 0.74，与检测能力仅 0.34。

Insight: 安全基准分数可能测量模型学到哪套标签定义，而不是抽象、统一的安全检测能力；数据污染、种子下限、标注定义和教师软标签来源都应作为评测报告变量。摘要中的极端分数跨度来自特定训练设置，不能外推为所有护栏系统的普遍幅度。

Abstract: A guardrail classifier is reported as if unsafe were a property of text. It is not: it is a property of a policy, and the policy enters the model through its labels. We exploit an accident of an industrial deployment in which the same ~53k prompts were labelled twice, under two operational definitions (“is this a request for harmful content?” and “would a moderator block this?”), and used to train ten guardrail arms sharing architecture, codebase and optimiser family, at two model scales. First, the guardrail’s own training corpus already contained the benchmarks: JailbreakBench-benign is 88% and OR-Bench-hard 82% present in the training pools, two over-refusal benchmarks, the kind a guardrail reports to argue it does not over-block. Everything below is decontaminated. Second, safety benchmarks stratify sharply by how much of their score is decided by the label definition rather than by the model: with architecture and optimiser fixed, the range across the ten arms is 0.99 on HarmBench-copyright and 0.043 on WildGuard-benign, a 23x spread. Third, a benchmark score can be moved across its entire range by 0.03% of the parameters. Across these arms HarmBench-copyright spans 0-100 and OR-Bench-hard over-refusal 3.4-56.4%, and the four arms whose last supervision signal carries a given policy occupy one extreme of both (exact rank p=0.005, or 0.04-0.14 at worst), at unchanged detection capability. For a distilled student the teacher’s soft labels do this even when every hard label carries the other policy, so teacher provenance is an unrecorded confound in any released checkpoint. Fourth, validation accuracy on the one split that grades all twenty arms tracks agreement with the annotator’s policy (rho=+0.74), not detection capability (rho=+0.34). We argue that guardrail evaluations should report contamination, a seed floor and the definition they were graded against as first-class variables.

[286] Trackable Skill Library Lifecycle Management for Agentic Reinforcement Learning cs.LG | cs.CLPDF

Xi Zhu, Yuanlin Duan, Changyi Yang, Kai Mei, Riyang Bao

TL;DR: TrackRL 在每个 GRPO 组内对技能独立随机注入，通过同任务组内对比和跨组可靠性加权估计技能的因果效用及模型掌握度，再据证据管理技能从生成、准入到保留、退休或移除；实验中同时改善技能辅助和无技能评测。

Details

Motivation: 智能体强化学习中的技能可能随策略学习出现不同生命周期：有的被模型内化后变冗余，有的仍作为外部脚手架有用，有的则变成有害。现有技能库因多个共同检索技能相互搭便车、任务难度混淆且缺少技能级反事实证据，难以可靠评估各技能贡献。

Result: TrackRL 将每个 GRPO 组内的采样转化为在线随机对照实验：独立随机注入技能，并在同一任务的组内对比，以隔离单项技能贡献；再跨组按可靠性加权累计证据。系统持续追踪技能效用（其作为外部脚手架的因果贡献）和模型掌握度（对应能力已内化多少），并把生命周期管理做成证据门控状态转换，涵盖生成、准入、受保护孵化、保留、退休和移除，同时用软容量上限控制技能库增长。实验显示，TrackRL 在技能辅助和不使用技能的评测中均优于近期技能增强 RL 基线。

Insight: 技能是否仍有价值应分清外部脚手架贡献与模型已内化能力，并用单技能随机对照证据决定生命周期，而非只看共现时的奖励。摘要未报告具体性能数字、技能库规模或控制成本。

Abstract: Skill-augmented agentic RL has emerged as an effective approach to improving LLM agents on long-horizon tasks. Despite this progress, individual skills may follow distinct trajectories as the policy evolves, with some becoming redundant after the model internalizes their capabilities, others remaining valuable as external scaffolding, and still others turning harmful. However, existing skill library systems fail to identify each skill’s true contribution due to free-riding among co-retrieved skills, difficulty confounding across tasks, and the absence of skill-specific counterfactual evidence, which prevents reliable skill lifecycle management. We propose TrackRL, an agentic RL framework where skill tracking drives skill library full lifecycle management. Specifically, within each GRPO group, we turn grouped sampling into online randomized controlled experiments. Independent random injection and within-group contrasts on the same task isolate each skill’s true contribution, and reliability-weighted aggregation accumulates evidence across groups. As evidence accumulates, we continuously track skill utility, which measures the skill’s causal contribution as external scaffolding, and model mastery, which reflects how far the corresponding capability has been internalized. Along this line, we formulate skill lifecycle management as an evidence-gated state-transition process that governs each skill from generation and admission through protected incubation to retention, retirement, or removal, while a soft capacity bound controls library growth. Experiments show that TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, with further analyses revealing how skill value shifts during policy learning and how evidence-driven state transitions shape skill library evolution. Our code is available at https://github.com/xizhu1022/TrackRL.

[287] ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference cs.LG | cs.AI | cs.CL | cs.DC | cs.PFPDF

Mobina Kashaniyan, Ali Jannesari

TL;DR: ATEST 在测试时扩展推理预算时同时考虑额外采样的预期收益和实时 GPU 能耗；在线版无需离线校准即可监控负载并决定停止，在四个 A100 基准上所有问题均未超能耗预算且准确率最高或统计持平，高负载 GSM8K 准确率为 60.0% 且无延期。

Details

Motivation: 测试时扩展通过采样多个答案提升推理，但固定采样数忽略了问题难度和 GPU 争用下能耗变化。作者测得相同四样本查询在负载增加时能耗最多上升 4.1 倍，因此固定预算可能不适应资源动态。

Result: ATEST 联合评估增加样本的预期价值和当前 GPU 成本。默认在线模式 ATEST-Online 无需离线校准，逐步生成样本、估计多数答案稳定性、监测实时 GPU 条件并在线学习单步能耗，以决定何时停止；ATEST-Calibrated 则利用离线准确率和成本模型按问题选择预算，适合负载稳定场景。在 A100 上以 Phi-3-mini 测试四个推理问答数据集并控制变化负载，另测试迁移到 L40S 和 Qwen2.5-3B。四个 A100 数据集上，在线版每题都未超能耗预算，准确率最高或与最高统计持平，并会随负载升高减少采样。高负载 GSM8K 上在线版准确率 60.0%、无延期；校准版为 44.0%、延期率 42.4%。计入离线分析后，在线版相较校准版少用 42 倍 GPU 时间和 48 倍能耗。在线控制在短任务上以更高尾延迟和较低吞吐为代价，换取负载变化下鲁棒性。

Insight: 采样预算应随查询难度和当下硬件负载变化，而不应把静态准确率成本估计套到所有推理时刻。在线成本学习是适应负载的关键消融结果；负载稳定时离线校准仍有竞争力。

Abstract: Test-time scaling (TTS) improves large language model (LLM) reasoning by sampling multiple answers, but fixed sampling budgets ignore both query difficulty and the changing energy cost of inference under GPU contention. We measure up to a 4.1-fold increase in energy for the same four-sample query as GPU load rises. We present ATEST, a resource-aware TTS framework that jointly considers the expected value of additional samples and their current GPU cost. ATEST-Online, its default mode, requires no offline calibration: it incrementally generates samples, estimates the stability of the plurality answer, monitors live GPU conditions, and learns step energy cost online to decide when to stop. ATEST-Calibrated instead selects a per-query budget using offline accuracy and cost models, favoring deployments with stable loads. We evaluate both modes on four reasoning and question-answering datasets using Phi-3-mini on an NVIDIA A100 under controlled varying load, and test transfer to an NVIDIA L40S and Qwen2.5-3B. Across the four A100 datasets, ATEST-Online answers every query without exceeding the energy budget, while achieving the highest or statistically tied accuracy and reducing sampling as load increases. On GSM8K at high load, it achieves 60.0% accuracy with no deferrals, versus 44.0% accuracy and 42.4% deferrals for ATEST-Calibrated. Ablations show that online cost learning drives adaptation to load. Including offline profiling, ATEST-Online uses 42 times less GPU time and 48 times less energy than ATEST-Calibrated for the evaluated workload. Online control trades higher tail latency and lower throughput on short tasks for robustness under varying load; offline calibration remains competitive under stable conditions.

Nozomu Fujisawa

TL;DR: Finite Interaction Lens（FIL）在有限干预评估中提高单干预组件数值精度，并在联合候选中复用这些数值贡献，同时保留混合残差和非线性路径；被测 Qwen3-VL-4B 上，合成族及两组四图 COCO 研究的交互误差分别降低 94.4%、99.85% 和 99.35%，但作者说明其实际价值取决于下游任务是否需要这类精度。

Details

Motivation: 要判断 VLM 中两种视觉干预是否相互作用，需比较它们联合和单独的四种输出；交互差可能很小，容易被舍入误差淹没。

Result: FIL 是评估有限干预族的混合精度方法。在选定线性投影处，它将各候选输入分为原生状态、两项单独干预分量和全维混合残差；单独分量用更高精度投影，并在联合候选中复用相同数值贡献，同时保留全部混合残差和非线性轨迹。作者推导局部抵消恒等式及平滑非线性下误差传播的条件界。在 Qwen3-VL-4B 的一个 MLP 上，相较同名义 BF16 乘积预算、但独立计算每个联合候选的控制，FIL 在 16 个合成族上交互 MSE 降低 94.4%，两项独立四图 COCO 研究分别降低 99.85% 和 99.35%；每项研究的绝对增益主要来自一张图。对所有候选使用更高精度多乘积 BF16 在相近总查询时间下总体误差更低；单乘积 BF16 已能恢复测试的两个候选选择任务中的参考选择。

Insight: 固定名义精度预算时，复用单干预计算可明显提高交互量测的数值保真度；但更高精度的全候选计算可能有更低总体误差，且已有选择任务不一定需要额外精度。论文将结果限定在开发输入，未证明一般任务收益。

Abstract: Identifying which visual interventions interact in a vision-language model requires comparing their joint and separate effects. The resulting four-output difference can be small enough for rounding error to obscure it. We introduce the Finite Interaction Lens (FIL), a mixed-precision method for evaluating finite intervention families. At each selected linear projection, FIL decomposes each candidate’s input into a native state, two single-intervention components, and a full-dimensional mixed residual. It projects the single-intervention components at higher precision and reuses the same numerical contributions in the joint candidates, while retaining every mixed residual and nonlinear trajectory. We derive a local cancellation identity and a conditional bound on error propagation through a smooth nonlinearity. In one MLP of Qwen3-VL-4B-Instruct, FIL reduces final interaction MSE by 94.4% on 16 synthetic families and by 99.85% and 99.35% in two separate four-image COCO studies, relative to a control that has the same nominal BF16-product budget and higher-precision single interventions but projects each joint candidate independently. One image accounts for most of the absolute gain in each COCO study. Higher-accuracy multi-product BF16 arithmetic on all candidates gives lower aggregate error at similar total query times. Single-product BF16 already recovers the reference choices in both candidate-selection tasks we tested. On these development inputs, FIL improves interaction fidelity at a fixed nominal precision budget. Its practical value depends on a task that needs the extra fidelity.

[289] Curriculum Learning Teaches LLMs Long-Horizon Implicit Reasoning, Up To A Limit cs.LG | cs.CLPDF

Jimson Huang, Maxwell Nickell, Ruowang Zhang, Abulhair Saparov

TL;DR: 课程学习按词汇规模和前瞻步数逐渐增加隐式推理难度；在匹配计算量下，Qwen3-0.6B 对最多需向前搜索 256 步的问题准确率达到 98%，无课程训练为 45%，但持续微调呈饱和趋势，表明隐式推理深度存在学习上限。

Details

Motivation: LLM 的显式思维链会把中间步骤写成 token，但长时任务上即使强推理模型也会突然失效。隐式推理可让模型在单次前向中提前搜索下一步，减少错误路径和回溯；既有研究显示其深度浅且单纯扩大规模无法解决，论文研究渐进式课程能否提升该深度。

Result: 作者通过逐步增加词汇规模和 lookahead（判断正确下一步所需向前搜索的步数）微调模型。在相同计算量下，课程训练能达到无课程训练无法达到的前瞻深度。Qwen3-0.6B 在 lookahead 最多 256 的问题上，课程训练准确率为 98%，无课程为 45%。继续微调的进步速度按饱和指数减缓，提示可学习隐式推理深度有内在上限；更大的模型拟合上限更高，预训练模型微调远快于从头训练，过大的课程步长会使进展停滞。合成任务能力还迁移至 11 个自然语言推理基准，在匹配计算量下超过无课程训练。

Insight: 逐步扩展模型需隐式预演的搜索范围，可以提高其单步规划深度，但并未消除能力上限；模型规模、预训练和课程步幅都会影响最终可达范围。摘要未给各自然语言基准具体成绩或拟合上限值。

Abstract: LLMs “reason” by externalizing intermediate steps as tokens through chain-of-thought (CoT), yet even frontier reasoning models hit abrupt performance cliffs on long-horizon tasks. Implicit reasoning (reasoning within a single forward pass) is useful during CoT generation, as the model can “search ahead” to decide the correct next move without explicit search. Models with weak implicit reasoning are more likely to commit to wrong paths and must attempt to recover via backtracking, while models with strong implicit reasoning identify the correct proof strategy directly, shortening the CoT. However, prior work shows LLMs struggle to reason implicitly beyond shallow depths, with no relief from scale. We investigate whether curriculum learning can extend this depth by fine-tuning on examples with progressively increasing vocabulary size and lookahead (the number of steps a model must search ahead to identify the correct next move). At matched compute, curriculum-trained models achieve lookaheads where no-curriculum training fails. Qwen3-0.6B reaches 98% accuracy on problems with lookahead up to 256 with curriculum, compared to 45% without. However, progress slows as fine-tuning continues, following a saturating exponential that suggests a maximum intrinsic lookahead to which models can learn to reason implicitly. Larger models reach higher fitted ceilings, fine-tuning pretrained models far outpaces training from scratch, and curriculum step sizes that are too large stall progression. Finally, the capability acquired on synthetic data transfers to natural language reasoning, with curriculum-trained models outperforming models trained without curriculum at matched compute on 11 natural language reasoning benchmarks.

[290] Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents cs.LG | cs.AI | cs.CLPDF

Pedram Mouseli, Massieh Moayedi

TL;DR: 在需要跟踪文件锁状态的受控任务中，Qwen3-8B 和 Qwen3.5-9B 的规则遵循收益主要来自每次行动前新生成的推理；把旧推理保留在上下文帮助有限，甚至会让模型服从来自不同状态历史的错误推理。

Details

Motivation: 智能体常先生成推理再行动，监控推理文本是否能预测动作取决于推理与状态跟踪的关系。论文设计一个文件系统任务：背景事件只告知哪个锁被切换、不告知新值，智能体必须根据历史重建当前锁状态才能安全移动文件。

Result: 作者用 Qwen3-8B 和 Qwen3.5-9B 分离考察行动前生成推理与在提示中保留历史推理的作用。大部分锁违规命令减少来自每次行动时当下生成推理；保留更早推理对遵循规则帮助很小，但能帮助其中一个模型完成任务。收益依赖推理内容：填充文本、针对其他历史生成的推理以及准确状态表都不能复现效果；若给模型一段仅有一个锁值不同的错误历史推理，它会采纳该推理，即使与自身历史冲突。没有显式推理时，线性探针能从激活中恢复状态，准确率与从文本记录中训练的分类器相近，但状态解码准确度无法预测干预是否改变动作。

Insight: 在此设置中，规则遵循更依赖行动前逐步生成的推理，而不是把旧推理当作可靠状态记忆；可解码的内部状态也不等于能预测或控制模型行为。结论限于单一受控文件锁任务，摘要未证明其对现实智能体的普遍性。

Abstract: Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions. We ask how an agent’s actions depend on its reasoning when the correct action depends on a changing state. In a small file-system task, background events announce which file’s lock has been toggled but not its new value. The agent therefore has to reconstruct the current locks from its history before it can move files safely. With Qwen3-8B and Qwen3.5-9B, we separate two roles of reasoning: generating it before each action, and retaining earlier reasoning in the prompt. Most of the reduction in lock-violating commands comes from reasoning written at the moment of action. Retaining earlier reasoning adds little to rule-following, although it helps one model complete the task. The benefit depends on what the reasoning says. Filler text, reasoning written for another history and an accurate table of the state do not reproduce it. When we give the agent reasoning written for a history that differs in one lock value, the agent acts on that reasoning even against its own history. Without reasoning, linear probes recover the state from the model’s activations about as well as a text classifier recovers it from the transcript. However, decoding accuracy does not predict which interventions change the agent’s actions. In this setting, the benefit of reasoning for rule-following comes mainly from generating it at each step. The state that can be decoded without reasoning is a limited guide to how the agent acts.

[291] Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding cs.LG | cs.AI | cs.CL | stat.AP | stat.MLPDF

Minoo Ahmadi, Seyedarmin Azizi, Mehdi Kamal, Massoud Pedram

TL;DR: LP-SMC 用带输入权重下限的 Chopthin 重采样，在扩散语言模型的序贯蒙特卡洛解码中保留低权重但仍可能成功的推理谱系；两种基座模型的数学推理和代码任务上，oracle coverage 全部提高 1.8–25.6 分，统一奖励模型选择器下 12 种设置中 11 种端到端准确率提高。

Details

Motivation: 推理时 SMC 可按奖励重采样扩散语言模型的部分生成，以改善准确率，但重采样既需聚焦有希望候选，也要保留不同推理路径。多项式重采样可能复制高权重粒子、丢弃早期低权重但后来能答对的谱系；有效样本量 ESS 可能在谱系已经消失时仍显示为最大值。

Result: Lineage-Preserving SMC（LP-SMC）将 Chopthin 重采样与输入权重下限结合。Chopthin 不再把权重完全均等化，而限制返回权重最大值与最小值的比例并将不等权重继续传递；下限则限制权重在重采样前可低于最大值多少，从而让早期低权重谱系保有恢复机会。两种基座模型的数学推理和代码基准中，oracle coverage（至少有一个正确粒子的题目比例）在每种设置提高 1.8–25.6 个百分点；用共同奖励模型选择器计算端到端准确率时，12 种设置有 11 种提高，LLaDA-1.5 上最高提高 7.9 点，Dream-7B 上最高提高 18.3 点。外部奖励模型下 coverage 增益仍存在。

Insight: 推理时搜索不能只追踪当前奖励最高的路径；限制权重过度集中可保留可能后来恢复的推理谱系，而且多样性保留与奖励引导分别贡献收益。摘要未给出各基准逐项数值或额外采样成本。

Abstract: Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward. Its effectiveness therefore depends on the resampling step, which must concentrate computation on promising candidates while retaining enough diversity to explore alternative reasoning paths. The reward can be internal, the decoder’s own confidence on the committed prefix, or external, a process reward model scoring that prefix. Under either reward, multinomial resampling can disrupt this balance by duplicating high-weight particles and discarding lineages that may still reach a correct solution. The effective sample size (ESS) does not register this loss: multinomial resampling resets it to its maximum even as distinct lineages disappear. To retain such lineages, we introduce lineage-preserving SMC (LP-SMC), which combines Chopthin resampling with an input weight floor. Instead of equalizing weights, Chopthin bounds the ratio between the largest and smallest weights it returns and carries the unequal weights forward. The floor limits how far below the maximum a weight may fall before resampling so that lineages with low early weight keep a chance to recover. Across mathematical-reasoning and code-generation benchmarks on two base models, LP-SMC raises oracle coverage, the fraction of problems with at least one correct particle, in every setting, by 1.8 to 25.6 points, and raises end-to-end accuracy under a common reward-model selector in eleven of twelve settings, by up to 7.9 points on LLaDA-1.5 and 18.3 on Dream-7B. The coverage gain persists under the external reward model, which suggests that diversity-preserving resampling and the guidance signal contribute separately to inference-time search.

[292] Correctness-Dependent Asymmetric Target Distributions for Self-Distillation Policy Optimization cs.LG | cs.AI | cs.CLPDF

Suwan Yoon, Seongryong Jung, Changhee Lee

TL;DR: APSD 按 rollout 正误构造非对称自蒸馏目标：对正确轨迹在学生与自教师分布间插值以保留成功行为，对错误轨迹外推以加强纠正；在多个 Qwen3 规模上，平均用 2.6 倍更少训练步数即可匹配 SDPO。

Details

Motivation: 自蒸馏策略优化能通过反馈条件化自教师提供密集 token 监督，但直接匹配自教师可能破坏正确 rollout 中已有的成功行为，同时对错误 rollout 的纠正又不够强。论文希望按结果正确性分别设置监督强度。

Result: Asymmetric Policy Optimization via Self-Distillation（APSD）根据 rollout 正误，将学生与自教师分布混合形成不同目标：正确 rollout 使用插值保留学生的有效行为；错误 rollout 则外推到自教师分布之外以加强纠正。论文从散度分解、反向 KL 重心几何和梯度分析解释方法，并指出梯度需通过依赖学生分布的目标传播。在可验证奖励强化学习和丰富反馈设置下，APSD 整体表现优于 RL 与自蒸馏基线；跨 Qwen3 多个模型规模，平均训练步数约为 SDPO 的 1/2.6，即减少约 2.6 倍后达到相当表现。

Insight: 成功轨迹与失败轨迹不应被同一种自教师匹配目标处理；保留前者、强化纠正后者有助于减少学习过程中的能力破坏。摘要未报告各规模详细指标及训练步数外的计算成本。

Abstract: Self-distillation policy optimization provides dense token-level supervision for LLM post-training through feedback-conditioned self-teachers. However, directly matching the self-teacher can disrupt successful behaviors on correct rollouts while providing insufficient correction on incorrect ones. We propose Asymmetric Policy Optimization via Self-Distillation (APSD), which constructs correctness-dependent targets by mixing student and self-teacher distributions: interpolation preserves successful behaviors on correct rollouts, while extrapolation beyond the self-teacher strengthens correction on incorrect ones. We analyze APSD through divergence decomposition, reverse-KL barycentric geometry, and gradient analysis, highlighting the importance of gradient flow through student-dependent targets. Experiments under reinforcement learning with verifiable rewards and rich feedback show that APSD achieves strong overall performance against reinforcement-learning and self-distillation baselines and matches SDPO performance with 2.6x fewer training steps on average across Qwen3 model scales.

[293] Intra-Rubric Reward Divergence: Mitigating Rubric Hacking through Latent Dimension Recovery and Balancing cs.LG | cs.AI | cs.CLPDF

Jianqiang Xia, Han Xu, Dianxi Shi

TL;DR: FRDA 让生成结果在各 rubric 子维度内分别竞争以缓解维度权重失衡，RSDC 则从整体偏好信号中恢复未明确写出的评价维度以补全覆盖；聊天、科学和医疗实验中，FRDA 训练更稳定，RSDC 通常优于单奖励和朴素多奖励聚合。

Details

Motivation: 基于 rubric 的强化学习会出现 rubric hacking。作者把评分拆成子维度后发现，总分提高时某些子维度得分反而系统性下降，说明模型可能牺牲特定质量方面来优化聚合目标；原因包括被重点覆盖的维度权重失衡，以及 rubric 缺少真实质量中的某些维度。

Result: 论文观察到 Intra-Rubric Reward Divergence，即 rubric 总分上涨时部分子维度分数下降。为缓解维度失衡，Fair-Race Dimension Aggregation（FRDA）让生成结果在每个维度分别竞争；为缓解维度缺失，Residual Semantic Dimension Completion（RSDC）从整体偏好信号中挖掘隐含评价维度以补全 rubric。聊天、科学和医疗领域实验显示，FRDA 减少 rubric hacking 并带来更稳定的提升；RSDC 优于单奖励训练和朴素多奖励聚合，通常产生质量更高的回答。

Insight: rubric 总分可能掩盖某些质量维度被牺牲，既要避免少数高权重维度支配训练，也要检查 rubric 是否遗漏整体偏好中的隐含标准。摘要未给出各领域具体数值或子维度诊断成本。

Abstract: Rubric hacking limits both performance and training robustness in rubric-based reinforcement learning. By decomposing rubrics into sub-dimensions, we reveal the Intra-Rubric Reward Divergence: as the total rubric score increases, certain sub-dimension scores systematically decrease, suggesting the model sacrifices specific quality aspects to optimize the aggregated objective. From the sub-dimension perspective, we observe two phenomena: (1) dimension imbalance, causing the policy to favor heavily covered dimensions; (2) dimension absence, leading to a systematic gap between rewards and true quality. To mitigate these issues, we propose Fair-Race Dimension Aggregation (FRDA), which has generations compete within each dimension separately to mitigate rubric imbalance, and Residual Semantic Dimension Completion (RSDC), which mines evaluation dimensions implicitly encoded in a holistic preference signal to complete rubric coverage. Experiments in chat, science, and medical domains indicate that FRDA mitigates rubric hacking with more stable improvements, while RSDC outperforms both single-reward training and naive multi-reward aggregation, yielding generally higher-quality responses.

[294] CAVA: Complementary-Label-Aware Data Valuation cs.LG | cs.CVPDF

Tan-Ha Mai, Hsuan-Tien Lin, Vu Nguyen

TL;DR: CAVA 将互补标签学习中的数据估值按“该样本不是类别 c”这一标签语义重新建模，通过互补标签成本和标签间隔稳定检测错标；在合成及人工标注基准上，能近乎完美分离对角噪声，并用部分干净验证数据改善所有测试的 CLL 目标。

Details

Motivation: 互补标签学习用“该样本不属于类别 c”的标签训练分类器，众包标注者或 VLM 可较低成本提供此类标签；但一旦错误地把真类标为互补类，模型就会被明确训练去压制正确答案。直接迁移最优传输数据估值还会混淆训练集互补标签与验证集正类标签的语义。

Result: 作者证明，直接迁移的 OT 标签组会变成依噪声率变化的混合：只有噪声足够高时分数才有信息，在 M 类问题的交叉率 1/M 处失去信息，低于此率时排序反转，而实际 CLL 标注往往位于低噪声区。CAVA 引入互补标签成本和标签间隔，以隔离每个互补标签对传输势的贡献，使间隔在不同噪声率下方向一致；它用无需调参的闭式精度阈值检测对角噪声，并将检测样本回收为普通类别监督。合成及人工标注基准上，CAVA 几乎完美区分对角噪声，而直接迁移 OT 估值还不如随机；仅用部分干净验证数据，也能提升所有被评 CLL 目标。

Insight: 数据估值必须尊重互补标签和普通标签的相反语义，否则低噪声时可能反转样本价值排序。显式建模该语义后，错误互补标注还可转为常规正类监督。摘要未给出各数据集具体噪声率或分数。

Abstract: Complementary-label learning (CLL) trains a classifier from labels that name a class an instance does not belong to. Such labels are cheap to collect from crowdworkers or vision-language models (VLMs), making them attractive for large-scale annotation. However, both sources can mistakenly assign the true class as the complementary label. Such annotation errors are particularly harmful in CLL because the model is explicitly trained to suppress the correct answer. Optimal transport (OT) data valuation identifies harmful samples by comparing training data with a clean validation set, but transferring it directly to CLL creates a semantic mismatch: label $c$ means not class $c$ in the training but class $c$ in the validation. We show that this makes the OT label groups noisy mixtures whose behavior depends on the noise rate. The transferred score is informative only at sufficiently high noise, becomes uninformative at the crossover rate $1/M$ for $M$ classes, and reverses below it, where practical CLL annotation typically operates. We propose CAVA, which introduces a complementary label cost and a label margin that isolates the contribution of each complementary label to the transport potential. The margin keeps a consistent direction across noise rates, detects diagonal noise with a parameter-free threshold of closed-form precision, and allows detected samples to be recycled as ordinary-label supervision. On synthetic and human-annotated benchmarks, CAVA separates diagonal noise almost perfectly where directly transferred OT valuation performs worse than random. With only a fraction of the clean validation data, it also improves every CLL objective we evaluate, showing that OT-based valuation becomes reliable once the semantics of complementary supervision are modeled explicitly.

[295] recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models cs.LG | cs.AI | cs.CLPDF

Saurav Jha, Qianli Liao, Robert Vernon, Sarath Chandar, Quentin Fournier

TL;DR: recAMPLIFY 将权重共享 Transformer 递归预训练用于掩码蛋白语言模型，以 155M 独立参数和与 AMPLIFY-350M 相同的 2T token 训练；参数少 56%，CASP 接触恢复更好，四项监督任务中三项提升，折叠准确率提高 22.5%。

Details

Motivation: 蛋白质语言模型可用于蛋白注释和突变效应预测，但进步主要依赖扩大模型和数据，成本高且易记忆。自回归语言模型的测试时扩展方法难直接迁移到掩码编码器，因为其没有生成循环，氨基酸也不以自然语言推理表达。

Result: recAMPLIFY 将递归预训练应用于掩码蛋白模型，以权重共享 Transformer 核心在可调整循环次数下迭代细化残基对齐隐藏状态。模型有 155M 独立参数，以同 AMPLIFY-350M 相同的 2T token 训练；尽管参数少 56%，CASP 接触恢复更好，四项监督任务中的三项提升，折叠准确率增加 22.5%。由于最佳递归深度因任务和实验而异，作者提出 Gated Mixture Recurrence（GMR），按概率混合递归各轮的隐藏状态，并以学习残差门控制每轮更新幅度。GMR 在 recAMPLIFY 上将期望层深最多降低 37%，但不同任务准确率权衡不一；机制分析显示门控可在训练深度之外控制循环稳定性，停止深度也与生物属性相关。

Insight: 对掩码蛋白模型，重复细化残基表示提供测试时计算扩展途径；递归次数应适应任务，门控有助于控制更深循环的稳定性。摘要指出 GMR 减少深度时存在任务相关的准确率权衡，未给出各任务完整数值。

Abstract: Protein language models (pLMs) are powerful tools for protein annotation and mutation-effect prediction, but progress has largely depended on scaling model size and data, making them expensive and prone to memorization. Test-time scaling offers a complementary path, but methods developed for autoregressive LLMs do not transfer directly: masked encoders lack a generation loop, and amino acids do not verbalize in a reasoning language. Instead, we adapt recursive pretraining to masked pLMs, where a weight-tied Transformer core refines residue-aligned hidden states over an adjustable number of cycles. We instantiate this recipe as recAMPLIFY, a pLM with 155M unique parameters trained on the same 2T tokens as AMPLIFY-350M. Despite 56% fewer parameters, it improves CASP contact recovery and three out of four supervised tasks, notably raising fold accuracy by 22.5%. Because optimal depth for recAMPLIFY varies across tasks and assays, we then introduce Gated Mixture Recurrence (GMR), a framework that decodes each protein from a probability-weighted mixture of hidden states visited during recursion. GMR employs a learned residual gate at scalar, low-rank, or per-dimension resolution to control how much of each cycle’s proposed update is accepted into the state. This unifies halting and gating into a single adaptive computation objective while recovering PonderNet-style halting as a special case without gates. Layered on top of recAMPLIFY, GMR further reduces expected layer depth by up to 37%, although the accuracy trade-off varies by task. Mechanistic analyses show that residual gating controls recurrence stability beyond the trained depth and that halt depth tracks biological properties.

[296] Text-to-Image Generative Modeling Without Language Pretraining cs.LG | cs.CVPDF

Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas, Ehsan Adeli, Jiajun Wu

TL;DR: Halcyon 从零开始以 1 亿图文对训练早期融合自回归文生图模型，不依赖语言预训练；0.52B 模型以 18.2 倍更少预训练计算达到 gFD 48.0，1.6B 模型以 275 倍更少计算达到 gFD 35.3，并优于所比模型。

Details

Motivation: 现代文生图系统通常依赖互联网规模语言预训练。论文研究强文本条件图像生成是否必须以语言模型预训练为前提。

Result: Halcyon 是完全从头预训练的早期融合多模态自回归模型家族，使用 1 亿图文对构成的 GPIC 数据集。Halcyon-0.52B 的 gFD 为 48.0，优于所比较 LlamaGen-XL/8 的 63.3，同时预训练计算量少 18.2 倍；Halcyon-1.6B 的 gFD 为 35.3，优于 JiT-T2I-XXL/8++ 的 41.0，计算量少 275 倍。控制实验显示，额外的纯文本预训练并未提升文生图生成。用图到文本监督替换 caption dropout 后，模型还学会图像描述能力，同时保持相近文生图质量。

Insight: 在所比模型和数据设置下，图文联合从头训练已可获得强文生图结果，纯文本语言预训练并非必要前置步骤；图像描述也可通过 I2T 监督加入。摘要未说明 gFD 的计算细节或更多规模比较。

Abstract: Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraining is necessary for strong T2I generation. We introduce Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs. Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) with 18.2$\times$ less pretraining compute. Further, Halcyon-1.6B outperforms JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275$\times$ less pretraining compute. Controlled experiments show that text-only pretraining does not improve T2I generation. Halcyon also learns image captioning by replacing caption dropout with I2T supervision, maintaining competitive T2I generation quality under the same pretraining budget. We show that strong T2I systems can be trained from scratch without internet-scale language pretraining. Code and models are available at https://cs.stanford.edu/~keshik/halcyon

cs.MA [Back]

[297] Multi-LLM Systems Exhibit Robust Semantic Collapse cs.MA | cs.AI | cs.CLPDF

Weiyi Kong, Shiyang Lai, Jinghua Piao, James Evans

TL;DR: 多模型语言系统即使不更新权重，也会在 200–1,000 轮闭环交互中出现语义坍缩：表达词汇看似不同，语义表示却逐步趋同；测试的 13 类干预均未能可靠恢复多样性。

Details

Motivation: 多智能体语言模型越来越多地用于自主生成，这重新引出闭环交互能否持续产生新内容的问题。论文关注不同模型组成的系统是否会在反复交互中保持语义探索，还是逐渐收敛到狭窄表达。

Result: 作者观察到多模型系统出现系统性语义表示趋同，尽管表面词汇仍有变化，过程不需要重新训练或更新模型权重。该模式跨模型家族并持续 200 至 1,000 轮。研究测试了 13 种干预实现，覆盖解码参数、提示设计、智能体组成、激活工程和强化学习，但均不能可靠恢复语义多样性。机制分析认为，这种坍缩不完全能由研究所检验的对齐或从众偏差解释，而与自回归生成的内在性质相符。

Insight: 在封闭的多轮生成环路中，表面语言多样不代表语义探索仍然多样，简单调整提示、采样或智能体组成也未必能解决收敛。摘要未给出语义坍缩的量化度量及各干预的具体效果。

Abstract: Whether machines can originate novel content has been debated for nearly two centuries, from Lovelace’s assertion that no engine can “originate anything” to Turing’s question of whether a machine can amplify ideas brought in from outside. Interacting multi-LLM systems, increasingly deployed for autonomous generation, reopen this question empirically. Here we show that such systems exhibit semantic collapse, systematic convergence in semantic representations despite apparent lexical variation, with no retraining or weight updates. The pattern holds across model families and over 200 to 1,000 rounds. Thirteen intervention implementations tested, spanning decoding parameters, prompt design, agent composition, activation engineering, and reinforcement learning, did not reliably restore semantic diversity. Mechanistic analyses suggest that semantic collapse is not solely attributable to the alignment or conformity biases examined here, but is consistent with intrinsic properties of autoregressive generation. Our results point to persistent constraints on the ability of multi-LLM systems to sustain open-ended exploration in closed-loop settings.

cs.RO [Back]

[298] Context-Aware RGB-Thermal-Radar Obstacle Perception for Resource-Constrained ASVs cs.RO | cs.AI | cs.CVPDF

Pranav Kedia, Edwin Redhead, Heiko Hamann

TL;DR: 该系统为自主水面艇融合 RGB、长波红外、毫米波雷达和惯性上下文，并输出带传感器状态的方位扇区障碍概率；24 段、351 米闭环湖面试验覆盖昼昏夜，低照留出集融合 AP 为 0.869，优于规则融合与门控融合，但低于可用时的 RGB 单模态。

Details

Motivation: 小型自主水面艇的视觉障碍感知会受到反光和眩光干扰，船体运动还会改变水线位置，夜间则可能失去主要 RGB 传感器。论文希望让系统在多传感器质量随环境变化时仍提供可用导航输入，而不是只追求静态融合准确率。

Result: 系统结合 RGB、长波红外图像、FMCW 毫米波雷达和惯性上下文，并向导航控制器提供校准后的障碍概率、雷达距离及传感器状态，区分传感器掉线与有效无障碍观测。新多模态湖面数据上进行 24 段闭环试点，包含 16 段白天、4 段黄昏和 4 段夜间，航程 351 米；夜间 RGB 不可用时仅依靠红外和雷达行驶。八段低照留出记录中，校准融合模型 AP 为 0.869、ECE 为 0.108；规则融合 AP 为 0.673，门控融合为 0.620，RGB 单独为 0.903。结果体现权衡：RGB 可用时排名最高，RGB 失效时多模态融合仍能提供导航输入。

Insight: 多模态系统的价值不仅是可用时提高精度，也在主要模态失效时维持可校准的可用感知。证据是受控规模的湖面试点，摘要未报告多日任务、误报警率或传感器失效分项。

Abstract: Reliable obstacle perception is challenging for small autonomous surface vessels (ASVs): reflections and glare confound visible-light cameras, vessel motion shifts the apparent waterline, and darkness removes the dominant sensing modality altogether. Rather than treating multimodal fusion as a static accuracy gain, we propose a context-aware architecture that combines RGB and long-wave infrared (LWIR) imagery, frequency-modulated continuous-wave (FMCW) mmWave radar, and inertial context. A bearing-sector interface provides the navigation controller with calibrated obstacle probabilities, radar ranges, and sensor status. It also distinguishes sensor dropout from a valid observation with no detected obstacle. Using new multimodal lake recordings, we deployed the pipeline in a 24-leg closed-loop pilot under uncontrolled lake conditions (16 day, 4 dusk, and 4 night; 351 m), including night legs in which RGB was unusable and the vessel drove on LWIR and radar evidence alone. On eight held-out low-light recordings, the calibrated fusion model achieved average precision (AP) 0.869 and expected calibration error 0.108, compared with AP 0.673 for rule-based fusion, 0.620 for gated fusion, and 0.903 for RGB alone. Our results expose a central tradeoff: RGB ranks best when usable, whereas fusion preserves a navigation input when RGB fails. This work is a step toward perception that remains usable and degrades predictably across multi-day ASV missions.

Siming Zhou, Bowen Ma, Ruiran Du, Pengcheng Zheng, Sheng Zheng

TL;DR: InJePA 将导航规划对象从完整动作序列改为潜在意图：目标条件 JEPA 生成意图，动作条件 JEPA 预测未来状态，预测后果再反馈到同一候选的下一步意图与动作推断；Habitat MP3D 图像目标闭环导航显示该意图引导方式有效。

Details

Motivation: 视觉导航需结合环境动态预测进行目标导向规划。代表性世界模型方法先提出完整动作候选，再预测其后果以选择或更新搜索，因此预测只负责评估动作，较少逐步指导候选本身如何构造。

Result: InJePA 提出以潜在状态为中心的意图优先规划。目标条件 JEPA 提出潜在意图，动作条件 JEPA 持续预测未来潜在状态，两者在共享空间中对齐预测表征。在每条候选计划内，预测后果会反馈到下一步意图提议和动作推断，形成“意图—动作—后果”闭环，使世界模型参与逐步构造候选，而非只在候选完成后评价。Habitat MP3D 闭环图像目标导航实验显示，意图引导的动作生成可有效完成多步导航；不同编码器实验也支持该方法适用于不同视觉表示。

Insight: 潜在意图可作为规划的中心对象，让预测后果直接引导下一步计划生成，扩展世界模型从候选评估到候选构造的作用。摘要未给出导航成功率、与基线的具体差距或复杂场景压力测试。

Abstract: Visual navigation requires goal directed planning informed by predictions of environment dynamics. Representative navigation world model methods propose complete action candidates before predicting consequences for selection or search updates. For each candidate, predictions primarily evaluate proposed actions rather than guide its stepwise construction. To this end, we propose InJepa, an intention first planning paradigm centered on latent states. A goal conditioned JEPA proposes latent intentions, and an action conditioned JEPA continually predicts future latent states. Both operate in a shared latent space to align their predictive representations. Within each candidate, predicted consequences feed into the next intention proposal and action inference, forming a closed loop of intention, action, and consequence. This extends the world model’s role beyond evaluation to leading candidate construction through latent intentions. We evaluate InJepa through closed loop image goal navigation in Habitat MP3D. The results show effective multistep navigation through intention guided action generation and consequence feedback. Experiments across encoders further support applicability across visual representations. These findings support latent intentions as the central planning object and show that world models can directly guide plan construction through coordinated reasoning about intentions and consequences. Code is available at https://github.com/zhuxun132/InJepa.

Junhua Huang, Zhizheng Liu, Honglin He, Xiang Zhang, Bolei Zhou

TL;DR: ChairNav 将通用视觉导航模型通过跨形态预训练、少量轮椅示范中训练和用户干预后训练适配为个性化电动轮椅导航系统；侧向 ADE 降低 18%，平均干预间距从 29.2 米提高到 57.4 米，公里级城市路线中自主行驶时间占 93.4%。

Details

Motivation: 自主电动轮椅的长距离人行道导航对无障碍出行很重要，但现有方法多限于室内地图，或依赖昂贵复杂传感器。视觉导航基础模型和大规模跨形态数据可能让轻量硬件也具备泛化能力。

Result: ChairNav 采用三阶段数据金字塔：先把大规模跨形态数据转换到轮椅领域进行预训练，再用少量轮椅示范中训练消除残余域差，最后根据用户干预后训练以个性化行为。系统部署在带 RGB-D 相机和 NVIDIA Jetson 的商用电动轮椅上。预训练使侧向平均位移误差降低 18%；后训练将用户干预间平均距离由 29.2 米提高到 57.4 米。公里级城市人行道路线中，系统有 93.4% 驾驶时间可自主运行。

Insight: 跨形态数据提供导航先验，小量轮椅演示补足具身域差，部署期用户干预再适配个体偏好；三阶段分别解决不同层次的数据不足。摘要未报告路线总数、用户人数或安全事件情况。

Abstract: Developing autonomous electric wheelchairs capable of long-horizon sidewalk navigation is critical for accessible personal mobility, yet existing approaches either focus on indoor map-based navigation or rely on expensive, complex sensor stacks to achieve full autonomy. In this work, we propose to tackle outdoor sidewalk navigation by leveraging visual navigation foundation models and large-scale navigation datasets, which offer strong generalization using lightweight hardware. To this end, we present ChairNav, a novel framework that adapts a general visual navigation model into a personalized electric wheelchair navigation system using a three-stage data pyramid. To address the scarcity of in-domain data, ChairNav first pre-trains on large-scale cross-embodiment datasets transformed into the wheelchair domain, followed by mid-training on a small number of wheelchair demonstrations to eliminate remaining domain gaps. Finally, a post-training stage refines the deployed model from user interventions to personalize its behavior. We deploy ChairNav on a commercial electric wheelchair equipped with an RGB-D camera and an onboard NVIDIA Jetson compute module. Experimental results validate the effectiveness of each stage in our data pyramid: pre-training reduces lateral ADE by 18%, and post-training increases the mean distance between user interventions from 29.2 m to 57.4 m. Further evaluation on kilometer-scale urban sidewalk routes shows that our system operates autonomously for 93.4% of total driving time, demonstrating its potential for everyday accessible mobility. Demo video and more details are at https://vail.cs.ucla.edu/ChairNav/.

[301] CLARO: Clutter Aware Robot Orchestrator for Open-Vocabulary Grasping using VLM Agents cs.RO | cs.CVPDF

Sungsu Kim, Vineet Bhat, Prashanth Krishnamurthy, Ramesh Karri, Farshad Khorrami

TL;DR: CLARO 用 VLM 智能体协调开放词汇检测、分割、6DoF 抓取和非抓取式推挤，以长时动作记忆在杂乱场景中定位并取回语言指定物体；仿真高杂乱成功率 89.1%，真实机械臂成功率 88.8%。

Details

Motivation: 家庭、仓库和实验室中的语言指令机器人需在杂乱环境中进行多步规划、执行并根据动态变化重规划，不能只依赖单步目标检测或抓取。

Result: CLARO（Clutter-Aware Robot Orchestrator）由 VLM 智能体协调开放词汇物体检测、分割、六自由度抓取和非抓取式推挤等感知与操作技能，以找到并取回语言指定目标。系统保存过去动作以支持长时执行，并输出抓取姿态或三维路点，可迁移到不同机械臂而无需任务专属训练。动态仿真和标准图像基准显示其对物体类别、光照及杂乱程度有较强泛化；平均成功率达 89.1%，高杂乱场景较最强对照相对提高超过 16%。真实机械臂在多种杂乱场景成功率为 88.8%。

Insight: 长时杂乱操作需要让检测、抓取和推挤技能由统一智能体根据动作历史进行组合与重规划，而非依赖单一预定义抓取管线。摘要未给不同技能消融或真实测试样本数。

Abstract: Robots that follow natural-language instructions unlock a wide range of autonomous applications in households, warehouses, and labs. Real-world environments are diverse and often heavily cluttered, requiring robots to plan over multiple steps, act, and replan as they work toward a goal. We present CLARO (CLutter-Aware Robot Orchestrator), a VLM agent that orchestrates a set of perception and manipulation skills: open-vocabulary object detection, segmentation, 6-DoF grasping, and non-prehensile pushing to locate and retrieve a language-specified target in clutter. CLARO maintains a memory of its past actions to support long-horizon execution, and emits manipulation targets (grasp poses/3D waypoints) allowing transfer across robot arms without task-specific training. We evaluate CLARO against a range of methods on a dynamic simulation environment and standard image-based benchmarks, demonstrating strong generalization across object classes, lighting, and clutter levels. CLARO attains an average success rate of 89.1%, a relative improvement of over 16% on the strongest competing method in high cluttered scenes. We further analyze performance along grounding and grasping axes to characterize where each method succeeds and fails. Finally, we deploy CLARO on a robot arm across diverse cluttered scenes, achieving a success rate of 88.8%, and demonstrating strong sim2real generalizability.

[302] Vision as Proprioception: Estimating Humanoid Robot Pose from Video cs.RO | cs.AI | cs.CVPDF

Jiaxi Jiang, Matthias Heyrman, René Zurbrügg, Marco Hutter, Christian Holz

TL;DR: HumanoidPose 将视觉当作本体感觉，用预训练人体姿态特征与机器人运动学从视频恢复世界坐标全身关节配置；通过合成迁移覆盖 16 种人形形态，并用无标签网络视频自训练提升现实泛化，在三种真实机器人上的关节角误差比各自最佳基线低 18–32%。

Details

Motivation: 与人形机器人交互时，内部状态可能因隐私、设备所有权或通信限制无法获取；网上和 AI 生成的机器人视频很多，却通常没有对应关节配置，不能直接用于机器人学习。

Result: HumanoidPose 把视觉作为本体感觉，从单目视频恢复机器人世界坐标下的全身关节配置。模型结合预训练人体姿态特征与机器人运动学，预测不同形态的关节状态。作者将人体动作重定向到 16 种人形形态，以纯合成数据规模化训练；为提高真实泛化，又使用无标签网络视频进行自训练，开发二维机器人姿态检测器。评测中，模型在关节角和关节位置准确率上超过所测基线；在三种真实人形平台上，相较各自最佳基线，关节角误差降低 18–32%，并改善全局定位和未见形态泛化。

Insight: 视频可在机器人内部状态不可访问时提供外部运动监督；合成跨形态数据和无标签网络视频自训练结合，能补充真实机器人标注稀缺。摘要未列三种平台名称、各误差绝对值或在线视频筛选办法。

Abstract: Estimating a humanoid robot’s internal state from video has a wide range of real-world applications but remains unexplored. During interactions with a robot, its onboard state information may be unavailable due to privacy, ownership, or communication constraints. Meanwhile, online and AI-generated robot videos provide rich sources of data but typically lack corresponding joint configurations, limiting their direct use in robot learning. To address these issues, we present HumanoidPose, a new framework that treats vision as proprioception to recover full-body joint configurations in world space from video alone. Specifically, we combine pretrained human-pose features with robot kinematics in a unified model that predicts each embodiment’s joint configurations. We retarget human motions to sixteen different humanoid embodiments, enabling scalable learning entirely from synthetic data. To improve real-world generalization, we further leverage unlabeled online videos through self-training to develop a 2D robot-pose detector. In our evaluation, HumanoidPose improves joint-angle and joint-position accuracy over the evaluated baselines and reduces joint-angle error by 18-32% relative to the best-performing baseline on each of the three real humanoid platforms. We further show improved global localization and generalization to unseen embodiments. HumanoidPose enables visual understanding of other robots during interaction and extraction of robot motion from online and generated videos.

[303] BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks cs.RO | cs.CV | cs.LGPDF

Yong Wang, Kecheng Zheng, Xing Zhu, Chao Xu, Yujun Shen

TL;DR: BiDex-100 是双手灵巧操作仿真基准，包含 106 个覆盖泛化、指令遵循、少量示范、长时任务和记忆的任务；每个任务配受控视觉与空间变化的域外版本，并提供两种形态、8,080 条验证成功的遥操作示范和 625 万帧。

Details

Motivation: 双手高自由度灵巧操作受到越来越多关注，但专为灵巧操作设计的基准较少，导致现有机器人动作模型的能力边界和鲁棒性缺少系统刻画。

Result: BiDex-100 将双手灵巧操作任务划分为五种能力轴：泛化、指令跟随、有限示范、长时任务与记忆，共 106 个仿真任务。每个任务还有通过受控视觉和空间变化构造的域外版本，用于系统测量鲁棒性。基准支持两种机器人形态，提供 8,080 条经成功验证的遥操作示范，共 625 万帧。作者在统一协议下评估七种近期视觉—语言—动作模型，建立各能力、鲁棒性和数据效率的基线。

Insight: 任务能力轴与受控域外变化相结合，可分辨模型是会操作，还是只适应特定外观和空间布局。摘要主要介绍基准与评测资源，没有列出七个模型的具体得分。

Abstract: Bimanual manipulation with dexterous hands has attracted growing attention in research on robotic action models. However, few studies have developed benchmarks specifically tailored for dexterous manipulation, leaving the dexterous manipulation capabilities of existing models insufficiently characterized. In this paper, we present BiDex-100, a simulation benchmark for bimanual dexterous manipulation that combines broad task coverage, multi-dimensional evaluation, and high-DoF hand embodiments. Concretely, the benchmark comprises 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes, allowing systematic assessment of robustness. In addition, the benchmark supports two embodiments and provides 8,080 success-validated teleoperated demonstrations, totaling 6.25 million frames. We evaluate seven recent Vision-Language-Action models under a unified protocol, establishing baselines for capability-specific performance, robustness, and data efficiency. The website is available at https://BiDex-100.github.io/BiDex-100/.

[304] InteractAgent: Agentic Human Motion Interaction with Memory-Augmented Reflection cs.RO | cs.CVPDF

Tao Zhang, Senhe Zhang, Zeyu Zhang, Yujia Zhang, Dong Gong

TL;DR: InteractAgent 用同一个多模态 LLM 负责长时人—场景交互的计划与评估，以冻结 MotionLLM 执行，并结合逐步观察恢复决策和记忆增强反思迭代细化计划；UniHSI 上较免训练基线提高任务成功并减少穿透。

Details

Motivation: 长时人体—场景交互需要协调空间推理、物理可行性并从累积执行错误中恢复。文本动作模型通常不感知场景且缺少闭环修正；几何条件方法则依赖成对场景—动作训练数据或显式 3D 预处理，难泛化到新布局和开放词汇目标。

Result: InteractAgent 是智能体式多模态 LLM 系统，以同一 MLLM 兼任规划器和评估器，冻结 MotionLLM 作为动作执行器，闭环进行计划、执行、观察和反思。Online Observation 模块逐步决定继续、重做或重规划；Memory-Augmented Reflection 则支持多轮计划改进和自适应停止。UniHSI 实验显示，较免训练基线，InteractAgent 提高任务成功率并降低人体与场景穿透。

Insight: 将规划、执行后的观察和记忆反思放进闭环，有助于在长时交互中修正累积误差，而无需针对新场景动作数据训练。摘要未给出任务成功和穿透的具体数值，无法量化改进幅度。

Abstract: Long-horizon human-scene interaction requires coordinated spatial reasoning, physical feasibility, and the ability to recover from compounding execution errors. Text-driven motion models are largely scene-agnostic and lack closed-loop correction, while geometry-conditioned approaches depend on paired scene-motion training or explicit 3D preprocessing that limits generalization to novel layouts and open-vocabulary goals. We present InteractAgent, an agentic multimodal LLM system for long-horizon human-scene interaction that plans, executes, observes, and reflects in a closed loop, using a single MLLM as both planner and evaluator and a frozen MotionLLM as the executor. InteractAgent integrates Online Observation for step-wise recovery decisions (continue/redo/replan) and Memory-Augmented Reflection for multi-round plan refinement with adaptive stopping. On UniHSI, InteractAgent improves task success and reduces penetration compared with training-free baselines. Code: https://github.com/AIGeeksGroup/InteractAgent.

[305] Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation cs.RO | cs.CVPDF

Lingsen Zhang, Wei Li, Jie He, Henghua Zhang, Kai Mo

TL;DR: VLDA 把机器人操作从直接学习视觉到动作映射，改为预测紧凑的可操作状态变化 Delta，并用局部转移过滤器、全局转移锚点及 Delta 驱动动作细化连接语义理解与动作生成；实验报告长时一致性、鲁棒性和泛化能力改善。

Details

Motivation: 机器人操作的核心是实现任务指定的物理状态变化，而不是单纯从观察生成动作。直接视觉—语言—动作映射要跨越语义理解与动作执行鸿沟；世界动作模型虽可预测未来，却可能把容量耗在静态外观和与动作无关的变化上。

Result: Vision-Language-Delta-Action（VLDA）将完整未来状态预测改为紧凑的可操作转移建模，以 Delta 表示潜在世界状态变化，作为视觉语言理解和动作生成的接口。Local Transition Filter 捕捉与操作相关的变化并抑制静态重建偏差；Global Transition Anchor 提供粗粒度未来约束以维持变化方向。Delta-Driven Action Refinement 再将 Delta 对齐到动作空间，作为结构化起点引导动作去噪。该框架在 LoongForge 中实现两阶段统一优化。实验显示，VLDA 提升长时一致性、鲁棒性和泛化能力。

Insight: 建模任务相关状态差分可能比重建完整未来更节省容量，并为动作生成提供中间接口。摘要未给出基准名称、定量提升或 Delta 表示误差。

Abstract: Robotic manipulation is not merely observation-to-action generation, but the realization of task-specified state changes in the physical world. Existing vision-language-action (VLA) methods directly learn observation-to-action mappings, imposing a substantial learning burden due to the gap between semantic understanding and action execution. Although world-action models (WAMs) improve control through future prediction, modeling complete future states wastes capacity on static appearance and action-irrelevant changes. We propose Vision-Language-Delta-Action (VLDA), a paradigm that reformulates full future prediction as compact actionable transition modeling. Its core is a compact Delta variable that captures latent world-state changes, serving as an interface between vision-language understanding and action generation. VLDA learns Delta through a unified objective composed of the Local Transition Filter (LT-Filter) and Global Transition Anchor (GT-Anchor). LT-Filter captures task-relevant manipulation-centric changes while suppressing static reconstruction biases. GT-Anchor provides coarse future-level regularization to preserve the overall direction of state changes. Delta-Driven Action Refinement (DDAR) further grounds Delta in the action manifold and injects it as a structured seed to guide action denoising. We implement VLDA within LoongForge, an open-source training framework for large-scale foundation and embodied models, using its modular and distributed infrastructure for unified two-stage optimization. Experiments show that VLDA improves long-horizon consistency, robustness, and generalization, validating actionable Delta as an alternative to direct action prediction and full future-state modeling.

[306] Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy cs.RO | cs.AI | cs.CVPDF

Zekai Li, Jiayi Yu, Jinlu Zhang, Luyang Niu, Zekun Qi

TL;DR: PhysRSI-eval 用首次表现、效率、成功、预算和失败后改进五轴衡量视觉语言模型的物理递归自我改进，并在 18 项操作任务与固定交互预算下评估；分析发现移除历史会降低后期成功，冻结工作区仍能帮助未见布局，提示场景适配和可复用经验可并存。

Details

Motivation: 物理环境中的递归自我改进需要模型在动作尝试后根据反馈调整策略，但现有公开基准平均成绩难区分首次能力、交互效率和失败后学习速度。

Result: 论文提出 PhysRSI-eval，以首次尝试表现、效率、成功、预算消耗和从失败中改进五个维度量化物理 RSI，在固定交互预算下评估多个 VLM 的 18 项操作任务。不同模型公共基准平均接近时，五轴指标仍能显示差异，也表明高平均分不保证早期或高效进步。作者从两个角度做观察性缩放分析：用公开能力预测 VLM 智能体相对表现，以及用早期物理反馈预测后续改进。工作区干预发现，删除历史会降低后期成功；冻结工作区仍有助于未见布局；三个跨任务案例实现更早成功。结果提示场景特定适配与可复用经验可以共存。

Insight: 衡量物理环境自我改进不能只看最终成功率，还要记录首次能力、交互预算和失败后学习；历史与场景信息既可能影响局部适配，也可能支持跨任务复用。摘要未提供缩放规律数值或各 VLM 表现排名。

Abstract: In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.

cs.SD [Back]

Jaeyeon Kim, Chao-Han Huck Yang, Luoyi Zhang, Chan-Jan Hsu, Fernando Ruiloba Portilla

TL;DR: LLT 构建结合细致听觉感知与显式认知行为的长音频推理训练数据：合成 160 万条推理轨迹并用于 Qwen2.5-Omni-7B 的 SFT+DPO，在 MMAU Test-mini、MMAR 和 MMSU 上分别将准确率从 71.5 提至 78.2、56.7 提至 63.7、60.6 提至 66.0。

Details

Motivation: 大型音频语言模型在音频理解和推理基准上的主要瓶颈往往是感知错误，但现有后训练数据中的音频描述类型有限，推理过程也较少显式包含核验、回溯等认知行为。论文希望同时提升音频证据覆盖和长链推理质量。

Result: Long Listening Thoughts（LLT）通过四阶段流程构建数据：生成覆盖语义、声学和时间信息的多角度密集描述；提出更具挑战的问题；合成错误推理轨迹；再用思路续写策略把短轨迹扩展成长轨迹，诱发核验和回溯等行为。作者合成 160 万条推理轨迹，并以先监督微调、后直接偏好优化训练 Qwen2.5-Omni-7B。MMAU Test-mini 准确率从 71.5 提至 78.2，MMAR 从 56.7 提至 63.7，MMSU 从 60.6 提至 66.0。

Insight: 音频推理训练数据需要明确覆盖细粒度听觉线索和可检查的思维行为，而不只是增加音频描述样本。结果来自所列模型与三个基准，摘要未拆分四阶段组件的独立贡献。

Abstract: We introduce Long Listening Thoughts (LLT), a framework for constructing post-training datasets that improves audio reasoning by pairing deliberate perceptual grounding with explicit cognitive behaviors. On audio understanding and reasoning benchmarks, large audio-language models (LALMs) are limited primarily by perceptual errors, yet existing post-training datasets rely on audio descriptions of limited diversity. Through a four-stage pipeline, LLT (1) builds multi-aspect dense audio descriptions spanning semantic, acoustic, and temporal information, (2) generates more challenging questions, (3) synthesizes incorrect reasoning traces, and (4) extends short reasoning traces into long ones via a thought-continuation strategy that elicits cognitive behaviors such as verification and backtracking. Using LLT, we synthesize 1.6M reasoning traces and train Qwen2.5-Omni-7B with a two-stage pipeline of supervised fine-tuning (SFT) followed by direct preference optimization (DPO). This improves accuracy on MMAU Test-mini (71.5 to 78.2), MMAR (56.7 to 63.7), and MMSU (60.6 to 66.0).

[308] AudioScorer: Learning Question-Aware Token Relevance for Large Audio Language Models cs.SD | cs.CL | eess.ASPDF

Marwan Farag, Bin Yang

TL;DR: AudioScorer 根据用户问题预测大型音频语言模型的音频 token 相关性，以冻结 LALM 的答案梯度显著性排序作为蒸馏监督，并结合答案监督训练；仅保留 10% token 时，在 10 个语义与声学任务中的 7 个达到最佳或并列最佳。

Details

Motivation: 大型音频语言模型需要处理语音、环境声、音乐等密集音频 token 序列，长音频计算成本高。现有剪枝方法依赖注意力、表示相似度或时间采样等启发式，没有直接学习哪些 token 对当前问题最有用。

Result: AudioScorer 是轻量问题条件打分模型，根据用户问题为音频 token 评分。训练时蒸馏冻结 LALM 对答案的梯度显著性排序，并加入下游答案监督。与其他 token 剪枝方法比较，在仅保留 10% token 时，AudioScorer 在 10 个语义和声学任务中的 7 个取得最佳或并列最佳，包括语音识别、情绪识别和音频场景分类。

Insight: 音频 token 的价值依赖当前问题，问题条件的相关性排序比固定时间采样或通用注意力分数更适合压缩长音频输入。摘要未报告其余三项任务的差距、编码器速度或显存节省。

Abstract: Large Audio Language Models (LALMs) reason over speech, environmental sounds, music, and other acoustic information. However, their audio encoders produce dense token sequences that make long-form audio costly to process. Existing audio pruning methods reduce this cost using heuristics such as attention, representation similarity, or temporal sampling, but do not explicitly learn which tokens are most useful for a given question. We propose AudioScorer, a lightweight model that scores audio tokens according to the user question. It is trained by distilling the ranking of audio tokens under the frozen LALM’s answer-gradient saliency, together with downstream answer supervision. Compared with other pruning methods, AudioScorer achieves the best or tied-best result at 10% token retention on 7 of 10 semantic and acoustic tasks, including speech recognition, emotion recognition, and audio-scene classification.

[309] Label Authority in Speech Emotion Recognition: Intended and Perceived Ground Truth in Emozionalmente cs.SD | cs.CL | eess.ASPDF

Marco Giordano, Stefano Giacomelli, Claudia Rinaldi, Fabio Graziosi

TL;DR: 在意大利 Emozionalmente 语料上比较三种语音情绪标签权威后，作者发现识别表演者表达的情绪比预测听众多数票容易；完整听众投票分布虽不明显改变模型硬分类，却能更好保留模糊语音中的感知不确定性。

Details

Motivation: 语音情绪识别常将真值当作单一类别，但表演者意图表达的情绪与听众实际听到的情绪是同一录音的不同方面。若标签权威不明确，模型之间的性能比较可能把任务定义差异误认为识别能力差异。

Result: 研究使用 Emozionalmente 意大利表演语音语料，每段录音均有诱发的表演情绪标签和五位听众判断；以冻结 XLS-R、HuBERT 和 WavLM 表征及匹配探针，比较表达条件、听众硬多数票和完整投票分布。三种表征得到一致结论：恢复表演情绪比预测打破平票后的听众多数类别容易，但二者是不同问题，不能据此断言一般语音情绪识别能力下降。听众一致性越弱，多数票硬标签越脆弱。使用完整投票分布训练不会明显改变硬分类胜出类别，却会使概率分配更贴近听众选择的备选情绪，尤其对模糊片段如此。

Insight: 表达标签说明说话者如何表演，听众多数票提供单类操作决策，投票分布保留感知不确定性，三者不能互相替代。面向用户的评测应说明预测哪种权威并保留聚合和一致性信息。

Abstract: Ground truth in Speech Emotion Recognition (SER) is often presented as a single categorical fact, yet an acted emotion and the emotion heard by listeners describe different aspects of the same recording. We examine this distinction in Emozionalmente, an acted Italian corpus that provides both an elicited production label and five listener judgements for every clip. Using frozen XLS-R, HuBERT, and WavLM representations with matched probes, we compare three label authorities: the expressed condition, the listeners’ hard plurality, and their complete vote distribution. The results tell a consistent story. Recovering the expressed condition is easier than predicting a tie-broken listener plurality, but that gap compares two different questions rather than revealing a general loss of SER ability. As listener agreement weakens, the plurality becomes an increasingly brittle hard target. Training on the full vote distribution does not substantially change which class wins the model’s hard decision, but it produces probability assignments that more closely follow the alternatives selected by listeners, especially for ambiguous clips. No authority therefore replaces the others. Expressed labels describe how speech was produced, pluralities provide an operational single-class decision, and vote distributions preserve perceptual uncertainty. For human-facing Internet of Sounds applications, evaluations should state which authority is being predicted, retain its aggregation rule and agreement information, and use distributional scores when disagreement is part of the target.

eess.IV [Back]

[310] Severity-Channel Conditioning: An Injection-Level Study for Consumer-Device Deblurring eess.IV | cs.AI | cs.CV | cs.LG | cs.MMPDF

Abdullah Al Shafi, Kazi Saeed Alam, Sumaiya Rahim Suma, Sk. Imran Hossain

TL;DR: 这项消费者设备去模糊研究比较四种严重程度信息注入方式；在允许推理时使用部分由配对清晰目标导出的 oracle 严重程度描述符的诊断设置中，输入拼接使 PSNR 从 31.23 提至 31.30 dB，但作者明确指出结果不证明注入位置是唯一原因，也不构成可靠部署优势。

Details

Motivation: 运动去模糊模型从专业相机迁移到手机图像时准确率下降。论文研究从 iPhoneBlur 配对数据导出的模糊严重程度描述符是否能改善适配，并在相同更新预算下比较输入拼接、全局元数据调制、空间特征调制和分层对比输出目标。

Result: 实验在同一个域适配潜扩散基线上微调所有方案，并纳入无条件频谱损失控制。主要实验是 oracle 诊断：推理时提供部分基于配对清晰目标得到的严重程度信息。被测模型对中，拼接严重程度图将整体 PSNR 从 31.23 提至 31.30 dB；特征调制、对比目标和频谱损失控制均未超过基线。增益在 87.2% 测试图像及 15 个留出源视频中的 14 个为正，但视频级结果只描述该检查点对，不证明重训鲁棒性；报告三个训练种子。输入级增益也在一个 Transformer 骨干上出现。一次探索实验中，推理时将该通道置零仍保留大部分 PSNR 增益，但 SSIM 降至基线以下；归一化步长诊断不支持预期的梯度更大解释。

Insight: 结果支持被测配置中输入级严重程度通道带来有限 PSNR 改善，但 oracle 目标信息、指标差异和检查点范围限制了部署结论；作者明确表示这些结果不能证明注入位置是唯一原因、存在空间局部机制或有可靠部署优势。

Abstract: Motion-deblurring models lose substantial accuracy when transferred from professional cameras to smartphone-derived imagery. This paper studies whether a severity descriptor, partly derived from the paired sharp target, helps adaptation on the iPhoneBlur consumer-device benchmark. Under a shared update budget, input concatenation of a severity map is compared with global feature modulation by capture metadata, spatial feature modulation by the severity map, and a tier-contrastive output objective, together with an unconditioned spectral-loss control, all fine-tuned from the same domain-adapted latent-diffusion baseline. The primary experiment is an oracle diagnostic: target-derived information is supplied at inference. For the evaluated model pair, input concatenation improves overall peak signal-to-noise ratio from 31.23 to 31.30 dB, whereas feature modulation, the contrastive objective, and the spectral-loss control do not exceed the baseline. The gain is positive on 87.2% of test images and on 14 of 15 held-out source videos; this video-level analysis describes the evaluated checkpoint pair and does not establish robustness to retraining, and three training seeds are reported separately. The input-level gain recurs on a transformer backbone. In one exploratory run, zeroing the channel at inference retains most of the gain in peak signal-to-noise ratio but lowers structural similarity below the baseline. A normalized-step diagnostic does not support the expected larger-gradient explanation. The results characterize the evaluated configurations and do not establish injection location as the sole cause, a spatial-locality mechanism, or a reliable deployment advantage.

[311] Progressive multi-scale tumor-state memory for incomplete multimodal brain tumor segmentation eess.IV | cs.CVPDF

Milad Rostamian, Mohammadreza Mohammadi, Mohsen Soryani

TL;DR: PMTM 在编码器各阶段维护当前受试者的肿瘤状态记忆，将可用 MRI 序列的空间特征汇聚为有序证据并门控融合到状态，再用于引导后续尺度融合；在 BraTS2018、2020、2023 GLI 的全部 15 种非空序列子集上，平均 Dice 分别为 75.31%、77.54% 和 80.47%。

Details

Motivation: 脑肿瘤分割常依赖互补 MRI 序列，但实际中某些序列可能不可用，不同肿瘤区域可用的信息也随缺失模态改变。仅支持多种输入组合的融合结构，并未明确说明肿瘤证据如何跨尺度保留并更新。

Result: Progressive Multi-scale Tumor-state Memory（PMTM）面向 T1ce、T1、FLAIR、T2 任意非空子集分割。在每个编码阶段，分配图将空间特征汇聚为各可用模态的有序证据向量，跨模态取均值后由学习门控与上一阶段状态合并；更新状态随后引导该阶段残差融合，并传递给下一阶段。模态 token 混合器还在每一阶段建模可用模态间交互，所有阶段的分配图均接受语义监督。三组 BraTS 数据集的 15 种模态子集上，平均 Dice 分别为 75.31%、77.54%、80.47%，均高于作者实验中 RobustSeg、RFNet 和 mmFormer。BraTS2020 另进行组件消融和内部状态分析。

Insight: 显式跨阶段保存和更新受试者级肿瘤证据，可使后续多尺度融合利用前阶段信息，而不仅仅堆叠模态特征。摘要未列各肿瘤区域分数或不同缺失模式的逐项结果。

Abstract: Brain tumor segmentation uses complementary magnetic resonance imaging (MRI) sequences, but some sequences may be unavailable in practice. Their absence changes the information available for segmenting different tumor regions. Fusion methods can support different input subsets. However, fusion at several scales does not by itself specify how tumor evidence is retained and updated to guide fusion at later stages. We propose a progressive multi-scale tumor-state memory network (PMTM) for segmentation from any non-empty subset of T1ce, T1, FLAIR, and T2. PMTM maintains a compact state for the current subject across the encoder stages. At each stage, we use assignment maps to pool spatial features into ordered evidence vectors for each available modality. We average these vectors across the available modalities, and a learned gate combines this evidence with the previous state. The up- dated state guides residual fusion at that stage and is used in the state update at the next stage. A modality-token mixer also models interactions among the available modalities at each stage. We apply semantic supervision to the assignment maps at all four stages. We evaluate PMTM on BraTS2018, BraTS2020, and BraTS2023 GLI across all 15 non-empty modality subsets. Averaged over the three tumor regions and the 15 subsets, PMTM achieves Dice scores of 75.31%, 77.54%, and 80.47% on BraTS2018, BraTS2020, and BraTS2023 GLI, respectively. These scores are higher than those of RobustSeg, RFNet, and mmFormer in our experiments on each dataset. Component ablations and internal state analysis are performed on BraTS2020.

stat.ML [Back]

[312] Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models stat.ML | cs.CL | cs.LG | q-fin.RMPDF

Agus Sudjianto, Denis Burakov

TL;DR: 理论分析表明，学生会继承软标签教师偏差，增加学生数据或匹配温度校准都不能自动纠正错误方向；在 231,521 条 CFPB 投诉上，结果标签训练的决策成本比逻辑教师概率训练低 8%，说明候选评分模型的概率需与真实结果核验。

Details

Motivation: 候选评分决策模型对给定选项输出概率分布。已有系统采用大模型软标签、冻结预训练骨干和温度校准，但这些训练环节究竟决定概率代表什么，以及它们何时降低风险或隐藏偏差，需要精确分析。

Result: 作者在高斯代理模型中精确计算有限样本风险：教师软标签训练的学生会继承教师偏差，更多学生数据无法消除，师生一致也不能证明无偏；固定表示仅在明确样本量阈值以下降低风险；probit 模型的精确再校准可用温度缩放表达，只有教师系数方向正确时才能消除全部额外风险，否则留下常规校准诊断未发现的残差。Luce 选择公理下，候选分数对事件细分保持一致，当且仅当通过 log-sum-exp 聚合。CFPB 的 231,521 条信用卡投诉中，以是否获得金钱补偿为结果：逻辑教师概率训练的学生达到教师决策成本，结果标签训练则比教师低 8%；30B 模型零样本结果即使再校准仍差于教师，且候选分数偏离一致性条件 17.5 decibans。

Insight: 冻结表示可以节省标注，校准可修正概率尺度，但二者都不能消除教师偏差；概率是否可信需要用实际结果检验。实际投诉数据的结论对应特定决策任务，摘要未说明其他业务场景的决策成本。

Abstract: Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft targets from larger models, a frozen pretrained backbone and temperature calibration. We analyze each ingredient in Gaussian surrogates where the finite-sample risk can be computed exactly. A student trained on teacher targets inherits the teacher’s bias, which more student data does not remove and agreement with the teacher does not reveal. A fixed representation lowers the risk only below an explicit sample-size threshold. In a probit model the exact recalibration map is a temperature scaling, which removes all excess risk only when the teacher’s coefficient has the correct direction and otherwise leaves a residual that calibration diagnostics do not detect. Under Luce’s choice axiom, candidate scores are coherent across refinements of an event if and only if they aggregate by log-sum-exp. On 231,521 credit card complaints from the Consumer Financial Protection Bureau, where the decision is whether a complaint ends in monetary relief, a student trained on a logistic teacher’s probabilities ends at the teacher’s decision cost while a student trained on outcomes goes 8% below it, a 30B model read zero-shot remains worse than the teacher after recalibration, and that model departs from the coherence condition by 17.5 decibans. Pretrained representations reduce the labeled data required and calibration corrects the scale of the probabilities, but neither corrects a biased teacher, so the probabilities have to be checked against realized outcomes.

Table of Contents

cs.CL [Back]

[1] The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections cs.CL | cs.AI | cs.CRPDF

[2] Preservation and loss of algorithmic structure in experimental diffusion of joint-action norms cs.CL | cs.MAPDF

[3] Reasoning with Context Management and External Memory cs.CLPDF

[4] Preserving Reasoning Diversity in Reinforcement Learning with Verifiable Rewards via Adaptive Replay cs.CL | cs.LGPDF

[5] Encoder-Free Multimodal LLMs Learn Implicit Encoders cs.CL | cs.AI | cs.CV | cs.LG | eess.ASPDF

[6] Toward a Universal Local Speech Feature Extractor Through Distillation cs.CLPDF

[7] When Opposite Premises Translate Identically: A Source-Conditioned Diagnostic for Multilingual Commonsense Items cs.CL | cs.AIPDF

[8] Modeling Discourse Profiles in Climate Misinformation Discourse cs.CLPDF

[9] EGIP: An Evidence-Grounded Interaction Protocol for Clinical Agents cs.CLPDF

[10] From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability cs.CL | cs.AI | cs.LGPDF

[11] Muse Glimmer 30B: An On-Device Agentic Model cs.CL | cs.AIPDF

[12] What RL Brings to Reasoning: A Few Decisions the Base Model Would Not Make cs.CL | cs.LGPDF

[13] DiffuPlan: Editable Planning State for Diffusion Language Model Agents cs.CL | cs.AIPDF

[14] ECHO: Efficient Causal Self-Recursion Beyond Auxiliary Multi-Token Prediction cs.CLPDF

[15] Ask, Don’t Guess: Targeted Clarification for Speech Recognition Errors cs.CLPDF

[16] Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning cs.CL | cs.IRPDF

[17] Epi: Towards a Persistent, Self-Improving Agent for Autonomous Research cs.CL | cs.AIPDF

[18] DeGMatch: Target-Aware Reasoning Data Selection via Deconfounded Gradient Matching cs.CLPDF

[19] Rubrics Carry the Standard: Shapley Attribution for Checklist-Based LLM Judges cs.CLPDF

[20] P2N: Reusing Deep Representations for Greater Effective Depth cs.CLPDF

[21] Beyond Factual Accuracy: A Benchmark for Evaluating Perspective Alignment in LLMs cs.CLPDF

[22] Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models cs.CLPDF

[23] TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis cs.CL | cs.AI | cs.LGPDF

[24] A Thousand Faces: A Persuasion Benchmark with Experience-Grounded Persuadees Informed by Psychology and Pragmatics cs.CLPDF

[25] Calibration-Free Early Stopping for Chain-of-Thought Reasoning via E-processes cs.CL | cs.AI | cs.LGPDF

[26] Atoms, Not Knowledge: What Production-Scale Addressed-Memory Writes Deposit cs.CLPDF

[27] Where and Why Search-Augmented LLMs Fail When Reasoning over Rapidly Evolving Knowledge cs.CL | cs.AI | cs.IR | cs.LGPDF

[28] From Unverifiable Domain Text to Cross-Lingual RLVR via Code-Switched Task Synthesis cs.CLPDF

[29] RARE: Reinforcement Learning with Adaptive Rubric Evolution for Open-Ended Generation cs.CL | cs.LGPDF

[30] Design Considerations in Low-Resource Archival OCR: A Case Study in Lakota cs.CLPDF

[31] Quality-Controlled Reply and Summary Data for Post-Training Small Language Models in AI Call-Screening Agents cs.CLPDF

[32] Self-Distillation for Supervised Fine-Tuning cs.CLPDF

[33] EmotiQ: A Quantum Residual Specialist Framework for Text Emotion Classification cs.CLPDF

[34] Flattening Fantasies: Analyzing Stereotypes in Pornographic Discourse cs.CLPDF

[35] Assertion and Belief: Moore’s Paradox in Language Models cs.CLPDF

[36] Cognitive Map in Large Language Models cs.CL | cs.LGPDF

[37] MapSteal: Map-based Autonomous Knowledge Stealing Attack Against Large Language Models cs.CL | cs.AIPDF

[38] A Calibrated Instrument for Detecting Disposition Drift in Language Models cs.CL | cs.AI | cs.LGPDF

[39] Truncated Reasoning Under a Budget: Evaluating System-1 Models as Decision-makers cs.CLPDF

[40] Diffusion Language Models Change Speed Across Tasks and Languages cs.CL | cs.AI | cs.LGPDF

[41] Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining cs.CL | cs.AIPDF

[42] Measurement Validity and Representation Dynamics of Confidence Signals in Large Language Models cs.CL | cs.AI | cs.LGPDF

[43] Learning to Reason Efficiently with Interventional Step Credit cs.CL | cs.LGPDF

[44] CALM: Consistency-Aware Logit Masking in Decoding Space for Mitigating Overthinking of Large Reasoning Models cs.CLPDF

[45] TempoMoE: Identifying and Steering Fast-Slow Thinking Experts for Efficient Multimodal Reasoning cs.CL | cs.CV | cs.LGPDF

[46] Scaling Synthetic Pre-training Data and Evaluation for Indic LLMs cs.CLPDF

[47] Balance-Aware Pivot Selection for Parallel Diffusion LLM Decoding cs.CLPDF

[48] Honeyed Words, Hidden Daggers: Benchmarking and Training LLMs to See Through Mock Politeness cs.CL | cs.AIPDF

[49] DeepSEA: A Multilingual Evaluation Benchmark for Southeast Asia cs.CLPDF

[50] Conditional information Retrieval for Longitudinal Survey Harmonisation cs.CL | cs.CY | cs.IRPDF

[51] Heartbeat Thinking: Speculative Reasoning and Tool Execution While Listening in Realtime Voice Agents cs.CLPDF

[52] Bridged Self-Distillation for On-Policy Learning cs.CLPDF

[53] How Benchmark Graph Construction Shapes Multi-Hop KG Navigation: A Controlled Audit of PathQuestion and PathQuestion-Large cs.CLPDF

[54] Co-Evolving Policy and Reward for Code Generation: A Mechanistic Study of Single-Adapter Drift cs.CL | cs.LG | cs.SEPDF

[55] HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL cs.CLPDF

[56] WenSight-Decision-1 for Direct Decisions under E-commerce Rules cs.CLPDF

[57] Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties cs.CL | cs.AIPDF

[58] Hierarchical Token Merging with Recursive Autoencoders for Unsupervised Grammar Induction cs.CL | cs.AIPDF

[59] HoloCosm: Beyond Single-Metric Benchmark Compression Across Multiple Metrics cs.CLPDF

[60] RESIN: Multi-Chain Reasoning with Entity-Structured Indexing cs.CLPDF

[61] Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability cs.CLPDF

[62] RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development cs.CL | cs.AI | cs.CVPDF

[63] Competing Learning in Hybrid Architectures cs.CLPDF

[64] OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory cs.CL | cs.AI | cs.CV | cs.LG | cs.MMPDF

cs.CV [Back]

[65] I$^2$-Med: Interpretable Medical Inference through Visual-Guided Dynamic Logits Calibration cs.CV | cs.CLPDF

[66] Latent Multimodal Memory Reasoning cs.CV | cs.AI | cs.CLPDF

[67] EMMA: Event-Driven Multimodal Memory System with Multi-Agent Retrieval cs.CV | cs.AIPDF

[68] Path-Conditioned Attention Mobility: From Depth to Execution Hubs in SegFormer cs.CVPDF

[69] EAGLE: Entropy-Aware Geometric Learning Enhancer for Partially Polygon-Supervised Object Localization cs.CV | cs.LGPDF

[70] AI-Experiment: an augmented reality art project co-created with artificial intelligence cs.CVPDF

[71] LVMH: A Meta-Optimization Harness for Long-Video Agent Memory cs.CV | cs.MAPDF

[72] Beyond Generator-Specific Shortcuts: Open-World AI-Generated Video Detection via Adversarial Training cs.CVPDF

[73] EgoVine: Post-Training with Branching Experience Compiled from Real-World Videos cs.CV | cs.AI | cs.ROPDF

[74] ContraSafe-VL: A Contrastive Safety Preference Dataset for Vision-Language Models cs.CVPDF

[75] When the Editing Intent Is Split: A Cross-Modal Jailbreak Attack for Large Image Editing Models cs.CV | cs.AIPDF

[76] Grounding Confusable Movie Moments cs.CVPDF

[77] Pseudo-Label Foreground Quality Weighting for Domain-Generalized Fundus Image Segmentation cs.CVPDF