Caption
-
Video-RSI: Recursive Self-Improvement of Video Understanding Agents via Harness Evolution
2026-09-29 16:24:02Bingjun Luo, Jialin Guo, Siqi LiAbstract
Video understanding agents acquire evidence through an executable harness that controls what they observe and how they use those observations. However, execution traces contain only the evidence acquired by the current harness, leaving competing explanations for failure unresolved and limiting the basis for self-improvement. We introduce Video-RSI, a framework for recursive self-improvement in which a video understanding agent uses its own language model to revise its harness. Through active video investigation, the model revisits the original training videos to test competing failure explanations with additional observations, grounding proposed changes in evidence beyond the existing trace. Cost-aware harness evolution turns these diagnoses into reusable revisions and determines which revisions to retain by considering both answer accuracy and visual cost. Across our evaluation settings on video understanding benchmarks, the evolved agent improves accuracy while processing fewer frames and achieves competitive accuracy-efficiency trade-offs against existing video understanding agents. These results demonstrate the potential for video understanding agents to improve their own evidence acquisition and use through harness evolution. Code is available at this https URL .
Abstract (translated)
URL
https://arxiv.org/abs/2609.37950
PDF
-
Does Local Video Understanding Transfer Across Encounters? The EgoGears Benchmark
2026-09-29 16:20:01Yuedong Tan, Lei Qi, Yu Liu, Di Wen, Ruiping Liu, Xiaoye Wang, Yufan Chen, Junwei Zheng, Chengzhi Wu, Chen Zhang, Zhihang Chen, Haiwen Sun, Zongwei Wu, Radu Timofte, Danda Pani Paudel, Kunyu PengAbstract
Embodied systems must make knowledge acquired during one encounter usable in another despite changes in viewpoint, motion, and illumination. Yet aggregate cross-video accuracy conflates failures of local perception with failures to preserve observation identity, establish correspondence, and compose evidence, obscuring whether local video understanding actually transfers. We introduce EgoGears, a complementary single- and multi-video benchmark designed to diagnose this transition. It contains 567 single-video and 1,487 multi-video questions derived from 126 human-collected egocentric recordings covering 39 outdoor routes. Repeated traversals across movement speeds and lighting conditions ground comparisons in shared physical environments; 531 questions require alignment across independent recordings. Single-video questions measure the local visual, spatial, and motion evidence available to a model, while multi-video questions test whether evidence remains bound to the correct observation and can be composed into consistent route relationships. We report 29 single-video and 31 multi-video MLLM configurations across six model families in the main leaderboard. Among the 20 configurations evaluated comparably on both splits, every model performs worse on multi-video questions, with a mean decrease of 22.5 percentage points, and the gap persists when answer format and scoring are held fixed. The gap is not explained simply by additional videos or recording boundaries. The central bottlenecks are observation--evidence binding and ordered route-state tracking. The code and benchmark are publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37938
PDF
-
PAIQ: Patch-Aligned Semantic Injection via Residual Rotation
2026-09-29 14:32:11Pinze Ren, Yuwei Zhang, Hao Chen, Linghao Meng, Chang Li, Qiankun LiAbstract
Language-aligned and self-supervised visual encoders offer complementary strengths in semantic abstraction and spatial detail. Harnessing this complementarity requires enriching local features while retaining distinctions between semantically related patches. We introduce PAIQ, a patch-aligned semantic injection framework that combines content-based cross-encoder matching with orthogonally constrained residual updates. Using DINOv3 patch features as the spatial base, PAIQ aggregates complementary SigLIP features through joint source allocation and injects the aggregate--base differences through a shared orthogonal transformation Q. This rotation adapts update directions while preserving residual norms and pairwise angles. For fixed projected features, we derive conditions for patch separability under similar semantic aggregates and show that rotation adds a nonnegative separation term over direct interpolation when the aggregate is shared. Only the projection and fusion parameters are trained; both visual encoders and the language model remain frozen, and fusion retains 196 visual tokens. Across diverse language backbones, PAIQ yields broad gains in judge-assessed correctness and reductions in hallucination severity over single-encoder interfaces on image description and visual question answering. On the 2B and 9B Qwen backbones, this compact interface outperforms the strongest evaluated fusion or token-compression baselines by about 2.9 correctness points on average.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37685
PDF
-
Devils in Question Relay: Source-Conditioned Relay Steering to Mitigate Hallucinations in Audio-visual Large Language Models
2026-09-29 13:37:31Yu Zhang, Pingrui Zhang, Xuefeng Bai, Pengfei Zhang, Yang Xiang, Kehai ChenAbstract
Audio-visual large language models (AVLLMs) have made remarkable progress in multimodal understanding and reasoning through interactions among visual, auditory, and linguistic information. However, recent studies show that AVLLMs face a critical challenge: $\textbf{source-confused grounding hallucination}$, where cues from the unused modality induce responses that the required modality does not support, undermining reliability in real-world applications. Existing methods have made progress in mitigating this failure, yet how it arises from internal cross-modal interactions remains insufficiently understood. To address this gap, we conduct path-intervention and representation analyses, revealing a $\textbf{question-relay}$ mechanism: question states carry interfering cues alongside required-source evidence, undermining grounding in required-modality evidence. Cutting pathways from interfering modality to question states yields greater correct-answer logit recovery than cutting those to the generation position. Motivated by these findings, we propose $\textbf{SECRET}$ ($\textbf{S}$ourc$\textbf{E}$-$\textbf{C}$onditioned $\textbf{RE}$lay s$\textbf{T}$eering), a training-free method that mitigates cross-modal interference at the question relay. Using contrasting question representations elicited through different modality-pathway interventions, SECRET steers the original question states toward required-source evidence. Experiments on two widely adopted benchmarks CMM and AVHBench across three AVLLMs show that SECRET consistently outperforms prior training-free methods, substantially mitigating source-confused grounding hallucinations (e.g., up to +18.0 and +7.1 percentage points over base models). Modality-specific captioning further demonstrates its generalizability to open-ended generation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37568
PDF
-
LazySloth: Bounded LLM-based Lazy Tree Search for Fast Long Video Comprehension
2026-09-29 12:53:57Arka Mukherjee, Kaleen Shrestha, Larissa Zhu, Maja Matari\'cAbstract
Modern vision-language models (VLMs) have shown promising results in long-video understanding due to the rich semantic information they can capture. However, most methods focus on coarse captioning of extracted image frames that are computationally inefficient and require models with large context windows. While past work has explored efficient methods through multimodal retrieval-augmented generation (RAG), they rely on lossy embeddings that lose temporal context and fine-grained detail. Few works to date have investigated how VLM-based query-relevant information retrieval can be optimized. We introduce LazySloth, an efficient tree-based search method that speeds up video comprehension and retrieval tasks 2.9-8.3x (compared to existing agentic methods) through bounded captioning of portions of the video considered irrelevant by a VLM of the video. Compared to contemporary specialized video-understanding VLMs and RAG-based methods, LazySloth achieved similar or better final task accuracy across two recent open-source base VLMs--Gemma 4 31B and Qwen3.6 27B--across four benchmarks. LazySloth reduced the gap between the base open-source model and a closed-source model, GPT-4o. Ablations showed that replacing VLM scene understanding with CLIP-based retrieval cost 8.8-19.9% in accuracy, while lazy tree construction matches eager construction at a fraction of the captioning cost. With LazySloth, we demonstrate the possibility of faster long-video comprehension without substantial loss in performance.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37426
PDF
-
MG-Thinker: Bi-Axial Self-Reflection for Multi-Image Reasoning Grounding
2026-09-29 12:28:34Heyu Huang, Chi Chen, Zonghao Guo, Yuhua Li, Maosong Sun, Ruixuan LiAbstract
Reinforcement learning (RL) has recently delivered substantial gains in multimodal reasoning, opening a promising route for fine-grained visual perception. Yet for multi-image reasoning grounding (MRG), reasoning over real-world multi-image contexts toward pixel-precise localization, existing RL-based approaches overlook two characteristics intrinsic to this paradigm: a coarse-to-fine hierarchical reasoning pattern, and heterogeneously distributed task--sample difficulties. In this work, we present MG-Thinker, a post-training RL framework that advances a new MRG paradigm featuring such hierarchical reasoning, supported by a curated 25K MRG dataset with task-adaptive Chain-of-Thought (CoT) annotations that elicit multi-perspective evidence before conclusion. To remedy the heterogeneous task--sample difficulties, we further propose Bi-Axial DAPO (BiA-DAPO), which decomposes rollout advantages along an intra-group signal axis and an inter-group competence axis through two complementary mechanisms, both grounded on our defined candidate pool for stable group-level statistics. Extensive experiments show that MG-Thinker achieves state-of-the-art performance on multi-image reasoning grounding while consistently improving generalization across multi-image understanding and diverse multimodal benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37374
PDF
-
When to Retrieve, When to Stay: Uncertainty-Aware Temporal Evidence Allocation for Streaming Video-LLMs
2026-09-29 12:13:56Xiang Hu, Jiazuo Yu, Lu Zhang, Yunzhi Zhuge, Huchuan LuAbstract
Streaming video understanding requires Video Large Language Models (Video-LLMs) to reason over continuous visual streams under causal constraints. As the visual history grows, a bounded visual?processing budget requires evidence selection that balances temporal recency with query relevance. Recent-only selection excludes potentially relevant historical evidence, whereas Semantic-only retrieval can displace useful recent context when relevance scores are ambiguous. We introduce WRWS (When to Retrieve, When to Stay), a training-free framework for uncertainty-adaptive evidence allocation. A lightweight external vision-language encoder scores query relevance across the observed history, while an adaptive allocation module uses the normalized entropy of the similarity distribution as a proxy for retrieval uncertainty. WRWS favors semantic retrieval when relevance cues are reliable and strengthens the recency prior under uncertainty. Following a retrieve-first, encode-later pipeline, WRWS selects evidence before target-model visual encoding, such that only the selected observations are processed by the costly target Video-LLM. Experiments across four Video-LLM families and multiple model scales demonstrate competitive accuracy on StreamingBench and OVO-Bench. In our efficiency evaluation, WRWS reduces average vision-to-answer time to 47.93% of the state-of-the-art method. Code will be released.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37345
PDF
-
GleanVID: Complementary Token Selection for Efficient Video Large Language Models
2026-09-29 08:50:57Shuo Yang, Changbai Li, Rui Tang, Xinyu Zhao, Linlin Yang, Baochang ZhangAbstract
Video Large Language Models (VideoLLMs) have achieved strong video understanding capabilities but incur substantial inference overhead due to the large number of visual tokens. Existing VideoLLM token compression methods largely rely on selection-independent scoring, overlooking cross-frame complementarity and consequently retaining redundant evidence across frames. Instead, we view video token selection as a progressive evidence accumulation process. It aims to retain visual evidence that is individually informative and collectively complementary under a limited token budget. Building on this insight, we introduce GleanVID, a training-free inference acceleration framework for VideoLLMs. Specifically, GleanVID first allocates the global token budget across frames according to temporal novelty and then selects tokens by jointly considering local representativeness and subspace complementarity, thereby preserving richer and less redundant visual evidence. Extensive experiments across diverse VideoLLMs and benchmarks demonstrate that GleanVID consistently achieves state-of-the-art performance. Notably, with only 25% of visual tokens, GleanVID preserves 98.6% of Qwen3-VL's original performance while reducing its prefill latency by 44.7%. On LLaVA-OV-7B, GleanVID at a 25% retention ratio even slightly surpasses the original model.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37042
PDF
-
Back2Struct: Making Structured Images Editable Again
2026-09-29 08:36:32Pengyu Yan, Yixin Wu, Yunjie Tian, David DoermannAbstract
Structured images, such as diagrams, charts, and flowcharts, are inherently symbolic and can be compactly represented in an editable format, yet in practice, they are often rendered as images, and therefore not graphically editable. This mismatch presents a significant challenge for researchers, engineers, and designers who wish to incorporate modified versions of existing graphic content into new materials without manually reconstructing it. In this study, we presentBack2Struct, which "makes structured images editable again" by directly recovering vector graphics code (SVG / XML) from image representations. Given an image of a structured graphic, Back2Struct predicts semantically object-level SVG / XML code that explicitly encodes text, shapes, topology, and layout, rather than performing low-level pixel vectorization. The generated code can be seamlessly imported into tools such as PowerPoint, allowing users to edit, refine, restyle, and reuse graphic content while preserving structural fidelity. Beyond supervised fine-tuning on ground-truth SVG token sequences, we further optimize Back2Struct with reward-based learning to better match deployment-time requirements: the output should be syntactically valid, properly concise, and visually faithful to the input diagram. Specifically, we design a composite reward that jointly encourages SVG / XML compilability, length consistency with the reference code, and structural or semantic similarity between the generated and ground-truth graphics. These complementary signals guide the model to produce SVGs that are not only closer to the training distribution, but also more complete, editable, and renderable in practice. Experiments show that Back2Struct improves accuracy, editability, validity, and user alignment over baselines. Dataset and code are available at: this http URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.37016
PDF
-
Counterfactual Attention Policy Distillation for Temporal Video Grounding
2026-09-29 08:12:10Shaobo Ju, Haiyang Yu, Xuecheng Wu, Qiong Wu, Jiacong Wang, Fan Shi, Jun Peng, Yiyi ZhouAbstract
Temporal video grounding is a key capability of advanced Multimodal Large Language Models (MLLMs) for the thorough understanding of video events, which is however often limited by repeated actions and visually similar contexts in long videos. In this paper, we study this issue from the perspective of On-policy distillation (OPD) and propose a new training regime for MLLMs termed Counterfactual Attention Policy Distillation (CAPD). In particular, OPD is a viable solution for MLLMs via providing dense teacher supervision on student-generated trajectories. But its next-token based teacher-student distillation is hard to identify the specific video segments supporting each predicted timestamp, which is critical for temporal grounding. In this case, CAPD measures how masking each temporal group changes the teacher's output distribution. The resulting counterfactual influence calibrates the teacher's attention and weights token-level distillation, allowing the student to learn the temporal evidence that affects boundary prediction. To validate CAPD, we trained it on Qwen3-VL-8B-Instruct using only 2,500 samples for one epoch, and evaluated it on the TimeLens and multiple general video benchmarks. Experimental results show that CAPD improves average recall by 12.0% relative to GRPO on TimeLens while preserving general video understanding, achieving comparable accuracy to the base model.
Abstract (translated)
URL
https://arxiv.org/abs/2609.34581
PDF
-
Momentum-Coupled Rubric Adaptation for Detailed Image Captioning
2026-09-29 07:22:09Zhenwen Ji, Lei Jin, Shanyong Wang, Jiaming Lu, Chengqiang Lu, Yi Wu, Yao Hu, Lizhen Cui, Yanyu XuAbstract
Detailed image captioning requires accurate and comprehensive descriptions of fine-grained visual content, yet caption quality spans factual accuracy, information coverage, and clarity. Compared with conventional methods that rely mainly on high-quality supervision or holistic rewards, rubric-based reinforcement learning decomposes these requirements into explicit criteria and provides targeted, structured feedback. However, existing methods often use separate models for caption generation, rubric construction, and judging, which may lead to inconsistent interpretations across roles. Some dynamic rubric methods alternate updates between the caption policy and rubric generator while keeping the judge fixed, but staged optimization may still leave rubric construction and judging out of step with policy optimization. We propose MoCo Rubric, a two-stage framework that coordinates these roles. First, role-conditioned, shared-parameter multi-task supervised fine-tuning equips a single vision--language model to serve as the Caption Policy, Rubric Generator, and Rubric Judge. Then, the Generator constructs rubrics online from captions sampled by the current Policy, reference captions, and image evidence. The Judge provides rubric-based rewards, and only the Policy receives GRPO updates. As Policy updates change the candidates being evaluated, we use an exponential moving average of the Policy parameters to update one momentum model shared by the Generator and Judge. This gradual transfer lets both rubric roles track Policy updates without separate RL optimization while smoothing parameter changes that could disrupt their rubric capabilities under direct synchronization. Across five captioning benchmarks, MoCo Rubric achieves an average pairwise win rate of 72.83\%, the best mean rank in blind ranking, and the highest average score in caption-based question answering.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36893
PDF
-
RED: Reconstruction Evolution Dynamics for Generalizable AI-Generated Image Detection
2026-09-29 06:33:54Wenpeng Mu, Junshan Jin, Tanfeng Sun, Xinghao Jiang, Qiang XuAbstract
The rapid evolution of image generators calls for forensic cues that generalize beyond known generation mechanisms. Existing detectors often rely on static image representations or endpoint reconstruction discrepancies, leaving the evolution of intermediate reconstruction stages underexplored. We observe that the relative token predictability of real and generated images can reverse across reconstruction scales, suggesting that intermediate stages may expose forensic evidence overlooked by endpoint comparisons. Motivated by this observation, we propose RED (Reconstruction Evolution Dynamics), a framework that captures transferable forensic cues from coarse-to-fine reconstruction evolution. To our knowledge, RED is the first framework to use scale-wise token predictability to guide forensic evidence aggregation across intermediate reconstruction states. It represents the reconstruction trajectory produced by a frozen multiscale VQ-VAE in the shared feature space of a frozen CLIP encoder. To connect the observed predictability variations with visual evidence, RED learns image-adaptive stage weights from scale-wise token negative log-likelihoods provided by a frozen VAR model. A cross-stage evidence aggregation module then jointly models the original-image representation and the weighted reconstruction features, capturing complementary forensic cues through interactions along the reconstruction trajectory. Experiments on six diverse benchmarks demonstrate that RED achieves the highest average accuracy of 92.5\% and average precision of 97.5\% among the evaluated methods. Further evaluations show strong robustness to common image degradations, supporting the value of reconstruction evolution for generalizable AI-generated image detection. The code will be made publicly available upon acceptance of this paper.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36822
PDF
-
CurvSpec: Adaptive Multi-Curvature Learning for Partial Relevant Video Retrieval
2026-09-29 06:28:17Zhen Liu, Letian Li, Jinpeng Wang, Shuzhao Xie, Yuzhi Huang, Jingyan Jiang, Zhi WangAbstract
Partially Relevant Video Retrieval (PRVR) seeks to retrieve untrim-med videos containing a moment that matches a text query, without temporal annotations. The relevant moment may last only seconds within a video spanning several minutes, creating an extremely low signal-to-noise ratio that makes PRVR more challenging than standard full-video retrieval. This task presents two intertwined challenges: (1) signal dilution, where coarse global representations blur the brief relevant signal into the dominant irrelevant surroundings;(2) curvature rigidity, where embedding all videos in the same fixed-geometry space distorts representations for videos that range from flat atomic events to deep compositional hierarchies. Existing PRVR methods have improved moment selection and cross-modal matching, but they still typically encode all videos in a single fixed-curvature retrieval space, limiting their ability to model diverse video structures. To address both challenges, we propose CurvSpec, a framework that learns content-adaptive curvature for video retrieval representations rather than imposing a fixed geometric prior. CurvSpec processes features through parallel Euclidean and hyperbolic attention layers, with independently learned curvatures assigned to the hyperbolic layers, and a content-aware fusion mechanism routes each input to its most suitable geometric regime. To further suppress signal dilution, CurvSpec represents each video with semantic centroids whose number is determined by the video's content complexity, projects them onto the learned manifold, and matches each query against its nearest centroid by geodesic distance. Experiments on ActivityNet Captions, TVR, and Charades-STA demonstrate state-of-the-art retrieval performance.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36815
PDF
-
EGSD: Event-Grounded Self-Distillation for Streaming Video Understanding
2026-09-29 06:15:16Yuwei Miao, Xuesheng Zhang, Wenhao Zou, Jixia Zhang, Jianwei Lv, Bo Yuan, Junfeng Wang, Shiao XieAbstract
Real-time video understanding requires incrementally maintaining a memory of streaming content, and optimizing this requires dense process signals. On-Policy Self-Distillation (OPSD), which lets one model serve as both teacher and student with the teacher receiving additional privileged information such as the question and ground-truth (GT) answer, can supply such token-level signals. However, applying it directly to streaming video raises two problems. (1) The student cannot be optimized end-to-end, where memory is written before the question arrives, yet the teacher scores it with the question-and-GT privilege, misaligning their preferences. (2) Effective-entity memory collapses, where the question-and-GT privilege makes the teacher favor only question-relevant entities, and token-mean averaging over a memory renders its signal invariant to how many entities that memory covers, both driving memory against the streaming need for diversity. To address these issues, we propose Event-Grounded Self-Distillation (EGSD), which characterizes streaming memory as an incremental update over verifiable Events (key visual entities, actions, and details) and targets the two problems on this basis. For problem (1), we adapt the OPSD signal into a multiplicative weight combined with the outcome reward; for problem (2), we re-weight the teacher with Events as privileged information to counter its question-relevance bias, and add an entity-coverage reward to supply the coverage preference the token-mean teacher lacks. Extensive experiments on mainstream online and offline benchmarks show EGSD achieves strong performance, reaching 79.8% on StreamingBench and 73.4% on the OVO-Bench Real-Time track, while memory analysis shows effective-entity recall rises 17.4% at only 6.8% more memory length.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36803
PDF
-
Causal-EVC: Breaking Emotional Spurious Causality via Spatiotemporal Grounding and Counterfactual Intervention
2026-09-29 05:49:15Cheng Ye, Weidong Chen, Peipei Song, Zhendong MaoAbstract
Emotional Video Captioning aims to generate factually accurate and emotionally empathetic descriptions. While recent methods have recognized the importance of visual causes to guide emotion perception and caption generation, they fundamentally rely on simple attention matching, which inevitably suffers from {causal redundancy and spurious correlations} in co-occurrence bias (e.g., misclassifying ``sadness'' as ``joy'' on a sunny beach), leading to severe shortcut learning from confusing backgrounds. Furthermore, existing evaluations fail to verify whether models have genuinely mastered causal reasoning or merely exploited background confounders. To address these limitations, we first construct {EVC-CauseGround}, a comprehensive benchmark with dense spatio-temporal causal annotations. Crucially, it introduces a carefully selected {Causal-Faithfulness Subset} to explicitly quantify genuine emotion-cause attribution. Second, we propose {Causal-EVC}, an emotion-grounding captioning framework, which introduces a Motion-guided Causal Spatiotemporal Localization module to precisely decouple causal triggers from background confounders. Besides, we introduce an Interpretable Sparse Emotion Routing module. By synthesizing counterfactual representations and formulating a novel counterfactual contrastive objective, we enforce the model to anchor its emotion predictions strictly on authentic causal triggers instead of confusing background. Extensive experiments show that Causal-EVC not only achieves the best performance on semantic metrics but also exhibits significant advantages in the causal-faithfulness subset, which demonstrates that our model could mine emotional cues from genuine visual causes and mitigate co-occurrence bias for interpretable multimodal emotion understanding.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36776
PDF
-
Beyond Binary Preferences: Graded Preference Optimization for Limb-Motion Captioning
2026-09-29 03:37:06Yanan Wang, Tingsong Li, Kaixun Jiang, Chongyang Zhong, Chenwei Xoe, Zhaohe LiaoAbstract
Vision-Language Models (VLMs) can generate rich video captions, yet often misidentify which person performs an action or which limb is involved, particularly across camera cuts. Improving these details requires evaluation and training that distinguish missing information from incorrect assertions. We introduce FlexBench, a benchmark spanning 3,105 shots and 18,161 evaluation queries, with human-verified identities and systematic per-person coverage of fine-grained limb actions and states. Its reference-derived checklists support automated assessment of complete captions in their person and shot contexts. Our Graded Physical Alignment score (GPA) awards credit for correct content and deducts points for incorrect or fabricated actions, making these errors explicit in the aggregate score. Building on this rubric, we propose Graded Margin Direct Preference Optimization (GM-DPO), which assigns stronger preference margins and greater training weight to more severe action errors. Across three VLM backbones, GM-DPO achieves the highest substantive-action and GPA scores among the evaluated preference objectives, improving GPA over DPO by 2.02-3.40 points. On Qwen3-8B, it reduces the weighted hallucination rate by 21.3% relative to DPO. These gains accompany sustained long-form output, improved shot structure, and competitive performance on three additional multimodal benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36628
PDF
-
MemEvo: Automatic Discovery of Streaming Video Memory Mechanisms
2026-09-29 02:59:42Guohong Liu, Jialei Ye, Shanhui Zhao, Yunxin Liu, Yuanchun LiAbstract
Query-agnostic streaming video understanding requires vision-language models to continuously compress an indefinitely growing visual stream into a bounded memory before future queries are known. The performance depends critically on the memory mechanism--what observations to preserve, how to represent and consolidate them, and what information to retrieve when a query eventually arrives. Rather than designing a single memory architecture by hand, we formulate memory design as a search problem over executable memory programs. We introduce a lightweight domain-specific language that expresses memory mechanisms through structured primitives for representation, admission, retention, consolidation, budgeting, and retrieval, while enforcing causal and bounded-memory constraints. Although structured, the derived program space remains large and contains heterogeneous, conditionally dependent design choices whose effects can only be assessed via downstream execution. We therefore propose MemEvo, an LLM-driven auto-research framework that uses pretrained LLM as a semantics-aware proposal model to iteratively generate and refine candidate memory programs based on accumulated experimental feedback. At runtime, a deterministic evaluation pipeline validates and evaluates each candidate, while the underlying vision-language model remains frozen throughout discovery. We finally produce a training-free, bounded-memory mechanism. Extensive experiments on StreamingBench and OVO-Bench demonstrate strong streaming video understanding performance together with substantial context and inference efficiency.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36581
PDF
-
How Medical VLMs Underutilize Their Vision Encoders: A Dermatology Perspective
2026-09-29 02:42:07Janet Wang, Yunbei Zhang, Xiao Wang, Jihun HammAbstract
Medical Vision-Language Models (VLMs) show significant promise for clinical image understanding, offering accurate diagnosis with interpretable reasoning. However, a critical performance gap exists between their strong vision encoders and the full multimodal model: in dermatology, the MedSigLIP encoder outperforms MedGemma by an average of 10.26 percentage points even when both use zero target-task labels; few-shot linear probing provides further evidence of strong visual representations. This gap motivates an investigation of how visual information is used in end-to-end diagnosis and why plausible-sounding predictions can lack grounding in image evidence. Using dermatology as our primary testbed, we systematically investigate three hypotheses for this phenomenon. We further provide a mechanistic analysis of the model's internal attention patterns, showing that a simple describe-then-decide prompting strategy increases vision attention by 30-40% during generation. Task-specific fine-tuning improves dermatology classification but reduces cross-domain medical question-answering performance in our evaluation. To address these challenges, we combine label-free prompting with low-label encoder-assisted reranking while keeping the VLM frozen. We validate the interventions across five VLM backbones in dermatology and provide supporting representation and attention analyses across additional medical modalities.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36557
PDF
-
EyeVQA: Benchmarking Ophthalmic Vision-Language Models from Recognition to Spatial Grounding
2026-09-29 01:27:10Gujie Shao, Zixun Xie, Xuechun Xing, Ruixiang Wang, Ziyun Lan, Yanlin Qi, Gangyi Zhang, Yuxin Yang, Dawei Li, Haiming TangAbstract
Vision-language models (VLMs) have shown increasing potential for medical image understanding, yet their capabilities in ophthalmic imaging remain insufficiently characterized. Existing ophthalmic datasets are typically designed for individual diseases or specialized tasks, making it difficult to systematically evaluate whether VLMs can move beyond disease recognition toward comparative reasoning and fine-grained spatial grounding. We introduce EyeVQA, a unified visual question answering benchmark for comprehensive evaluation of ophthalmic VLMs. EyeVQA is constructed from 21 available ophthalmic datasets and contains 20,000 question-answer pairs spanning six disease groups and seven question types: Single-Choice, Multi-Select, Variable-Select, True-False, Ranking, Point Location, and Bounding Box. Gold answers are deterministically derived from source-provided diagnoses, severity grades, clinical findings, segmentation masks, bounding boxes, and anatomical landmarks, enabling reproducible evaluation without relying on model-generated annotations. Notably, 44.5% of the questions require reasoning across multiple images, extending evaluation beyond conventional single-image medical VQA. We benchmark fourteen representative general-purpose, scientific, and medically specialized VLMs under a unified zero-shot protocol. The best-performing model only achieves an overall score of 62.8, while substantial gaps remain in spatial grounding and cross-task generalization. These results highlight the limitations of current VLMs in comprehensive ophthalmic visual understanding and establish EyeVQA as a diagnostic benchmark for developing more reliable and spatially grounded ophthalmic multimodal models. The project page is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32352
PDF
-
FigAct: Turning Scientific Figures into Active Canvases for Explanation
2026-09-28 19:55:22Shishi Xiao, Zichao Wang, Alexa Siu, David H. Laidlaw, Jennifer HealeyAbstract
Scientific figures are designed to communicate information visually, yet MLLMs typically explain them by translating their visual content back into text. This requires readers to manually map the resulting explanations back to the figure. Inspired by how people present visual information, we introduce FigAct, a framework that transforms static scientific figures into question-conditioned visual presentations by acting directly on their existing graphical elements. Like a human presenter, FigAct generates a sequence of short narrations, grounds each narration in the corresponding visual evidence, and applies visual actions to guide the viewer's attention. We develop a hierarchical search strategy for efficient element localization, reducing token usage by approximately 40$\times$. We further train FigAct-8B using three task-specific rewards for grounding accuracy, search efficiency, and rendering quality. We further build a human-verified benchmark from figures in real-world scientific papers to evaluate the ability of MLLMs to generate grounded visual explanations. Our results demonstrate the effectiveness of FigAct and show that treating scientific figures as presentation canvases makes explanations clearer and easier to follow.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36190
PDF
