Video Reasoning

Vision-Language Models (VLMs) typically rely on static initial frames for video reasoning, restricting their ability to incorporate essential dynamic information as the reasoning process evolves. Existing methods that augment Chain-of-Thought (CoT) with additional frame information often exhibit suboptimal CoT quality and lack the crucial ability to synthesize visual information for hypothetical or counterfactual scenarios. We introduce Act-to-See (Act2See), a novel framework that enables active visual perception by empowering VLMs to actively interleave video frames within text CoTs. Act2See is developed via Supervised Fine-Tuning (SFT) on a high-quality dataset of reasoning traces generated by a frontier VLM. These traces integrate active calls to either retrieve existing frames or generate new ones, and are rigorously verified against human-annotated CoTs to ensure quality. This approach cultivates an emergent capability: at inference time, the model actively determines when to search for or synthesize the necessary visual evidence. Act2See establishes new state-of-the-art results on challenging benchmarks, including VideoEspresso and ViTIB, and outperforms comparable or larger models on Video-MME, EgoNormia, and VCR-Bench, demonstrating an advancement in enabling VLMs with active visual perception for video reasoning.

Although reinforcement learning (RL) has significantly advanced reasoning capabilities in large multimodal language models (MLLMs), its efficacy remains limited for lightweight models essential for edge deployments. To address this issue, we leverage causal analysis and experiment to reveal the underlying phenomenon of perceptual bias, demonstrating that RL-based fine-tuning compels lightweight models to preferentially adopt perceptual shortcuts induced by data biases, rather than developing genuine reasoning abilities. Motivated by this insight, we propose VideoThinker, a causal-inspired framework that cultivates robust reasoning in lightweight models through a two-stage debiasing process. First, the Bias Aware Training stage forges a dedicated "bias model" to embody these shortcut behaviors. Then, the Causal Debiasing Policy Optimization (CDPO) algorithm fine-tunes the primary model, employing an innovative repulsive objective to actively push it away from the bias model's flawed logic while simultaneously pulling it toward correct, generalizable solutions. Our model, VideoThinker-R1, establishes a new state-of-the-art in video reasoning efficiency. For same-scale comparison, requiring no Supervised Fine-Tuning (SFT) and using only 1 of the training data for RL, it surpasses VideoRFT-3B with a 3.2% average gain on widely-used benchmarks and a 7% lead on VideoMME. For cross-scale comparison, it outperforms the larger Video-UTR-7B model on multiple benchmarks, including a 2.1% gain on MVBench and a 3.8% gain on TempCompass. Code is available at https://github.com/falonss703/VideoThinker.

Refereeing is vital in sports, where fair, accurate, and explainable decisions are fundamental. While intelligent assistant technologies are being widely adopted in soccer refereeing, current AI-assisted approaches remain preliminary. Existing research mostly focuses on isolated video perception tasks and lacks the ability to understand and reason about foul scenarios. To fill this gap, we propose SoccerRef-Agents, a holistic and explainable multi-agent decision-making framework for soccer refereeing. The main contributions are: (i) constructing the multimodal benchmark SoccerRefBench with over 1,200 referee theory questions and 600 foul video clips; (ii) building a vector-based knowledge base RefKnowledgeDB using the latest "Laws of the Game" and a classic case database for precise, knowledge-driven reasoning; (iii) designing a novel multi-agent architecture that collaborates via cross-modal RAG to bridge the semantic gap between visual content and regulatory texts. This work explores the technical capability of integrating MLLMs with refereeing expertise, and evaluations show our system significantly outperforms general-purpose MLLMs in decision accuracy and explanation quality. All databases, benchmarks, and code will be made available.

Visual question answering increasingly requires multi-step reasoning. Recent post-training with reinforcement learning under verifiable rewards (RLVR) and Group Relative Policy Optimization (GRPO) can improve multimodal reasoning, but most approaches rely on sparse outcome-only rewards. As a result, they struggle to tell whether an incorrect answer comes from a small mistake late in the reasoning or from an unhelpful trajectory from the start. A common solution is to train a process reward model (PRM) for step-level supervision, but this typically requires large-scale high-quality chain-of-thought annotations and additional training cost. We propose ProcessThinker, a practical post-training pipeline that provides step-level process rewards without training an explicit PRM. ProcessThinker first rewrites reasoning traces into a step-tagged format for cold-start supervised fine-tuning, then applies GRPO with a standard format reward and our rollout-based process reward. Concretely, for each intermediate step, we sample multiple continuations from that step and use the empirical success rate (final-answer verification) as the step reward. This gives dense credit assignment and encourages reasoning steps that more reliably support a correct conclusion, helping reduce inconsistent or self-contradictory progress across steps -- a key issue in logical reasoning. Across four challenging video benchmarks (Video-MMMU, MMVU, VideoMathQA, and LongVideoBench), ProcessThinker consistently improves over the baseline model Qwen3-VL-8B-Instruct

Robotic systems operating in human environments must reason about how object interactions evolve over time, which actions are currently being performed, and what manipulation step is likely to follow. Classical enriched Semantic Event Chains (eSECs) provide an interpretable relational description of manipulation, but remain primarily descriptive and do not directly support uncertainty-aware decision making. In this paper, we propose eSEC-LAM, a neuro-symbolic framework that transforms eSECs into an explicit event-level symbolic state for manipulation understanding. The proposed formulation augments classical eSECs with confidence-aware predicates, functional object roles, affordance priors, primitive-level abstraction, and saliency-guided explanation cues. These enriched symbolic states are derived from a foundation-model-based perception front-end through deterministic predicate extraction, while current-action inference and next-primitive prediction are performed using lightweight symbolic reasoning over primitive pre- and post-conditions. We evaluate the proposed framework on EPIC-KITCHENS-100, EPIC-KITCHENS VISOR, and Assembly101 across action recognition, next-primitive prediction, robustness to perception noise, and explanation consistency. Experimental results show that eSEC-LAM achieves competitive action recognition, substantially improves next-primitive prediction, remains more robust under degraded perceptual conditions than both classical symbolic and end-to-end video baselines, and provides temporally consistent explanation traces grounded in explicit relational evidence. These findings demonstrate that enriched Semantic Event Chains can serve not only as interpretable descriptors of manipulation, but also as effective internal states for neuro-symbolic action reasoning.

Existing Video Large Language Models (Video LLMs) struggle with complex video understanding, exhibiting limited reasoning capabilities and potential hallucinations. In particular, these methods tend to perform reasoning solely relying on the pretrained inherent reasoning rationales whilst lacking perception-aware adaptation to the input video content. To address this, we propose Video-ToC, a novel video reasoning framework that enhances video understanding through tree-of-cue reasoning. Specifically, our approach introduces three key innovations: (1) A tree-guided visual cue localization mechanism, which endows the model with enhanced fine-grained perceptual capabilities through structured reasoning patterns; (2) A reasoning-demand reward mechanism, which dynamically adjusts the reward value for reinforcement learning (RL) based on the estimation of reasoning demands, enabling on-demand incentives for more effective reasoning strategies; and (3) An automated annotation pipeline that constructs the Video-ToC-SFT-1k and Video-ToC-RL-2k datasets for supervised fine-tuning (SFT) and RL training, respectively. Extensive evaluations on six video understanding benchmarks and a video hallucination benchmark demonstrate the superiority of Video-ToC over baselines and recent methods. Code is available at https://github.com/qizhongtan/Video-ToC.

Fine-grained spatiotemporal reasoning on surgical videos is critical, yet the capabilities of Multi-modal Large Language Models (MLLMs) in this domain remain largely unexplored. To bridge this gap, we introduce SurgCoT, a unified benchmark for evaluating chain-of-thought (CoT) reasoning in MLLMs across 7 surgical specialties and 35 diverse procedures. SurgCoT assesses five core reasoning dimensions: Causal Action Ordering, Cue-Action Alignment, Affordance Mapping, Micro-Transition Localization, and Anomaly Onset Tracking, through a structured CoT framework with an intensive annotation protocol (Question-Option-Knowledge-Clue-Answer), where the Knowledge field provides essential background context and Clue provides definitive spatiotemporal evidence. Evaluation of 10 leading MLLMs shows: 1) commercial models outperform open-source and medical-specialized variants; 2) significant gaps exist in surgical CoT reasoning; 3) SurgCoT enables effective evaluation and enhances progressive spatiotemporal reasoning. SurgCoT provides a reproducible testbed to narrow the gap between MLLM capabilities and clinical reasoning demands. Code: https://github.com/CVI-SZU/SurgCoT.

Despite remarkable progress toward general-purpose video models, a critical question remains unanswered: how far are these models from achieving true multimodal reasoning? Existing benchmarks fail to address this question rigorously, as they remain constrained by straightforward task designs and fragmented evaluation metrics that neglect complex multimodal reasoning. To bridge this gap, we introduce CLVG-Bench, an evaluation framework designed to probe video models' zero-shot reasoning capabilities via Context Learning in Video Generation. CLVG-Bench comprises more than 1,000 high-quality, manually annotated metadata across 6 categories and 47 subcategories, covering complex scenarios including physical simulation, logical reasoning, and interactive contexts. To enable rigorous and scalable assessment, we further propose an Adaptive Video Evaluator (AVE) that aligns with human expert perception using minimal annotations, delivering interpretable textual feedback across diverse video context tasks. Extensive experiments reveal a striking answer to our central question: while state-of-the-art (SOTA) video models, such as Seedance 2.0, demonstrate competence on certain understanding and reasoning subtasks, they fall substantially short with logically grounded and interactive generation tasks (achieving success rates <25% and ~0%, respectively), exposing multimodal reasoning and physical grounding as critical bottlenecks. By systematically quantifying these limitations, the proposed method provides actionable feedbacks and a clear roadmap toward truly robust, general-purpose video models. CLVG-Bench and code are released here.

Reinforcement learning has advanced video reasoning in large multi-modal models, yet dominant pipelines either rely on on-policy self-exploration, which plateaus at the model's knowledge boundary, or hybrid replay that mixes policies and demands careful regularization. Dynamic context methods zoom into focused evidence but often require curated pretraining and two-stage tuning, and their context remains bounded by a small model's capability. In contrast, larger models excel at instruction following and multi-modal understanding, can supply richer context to smaller models, and rapidly zoom in on target regions via simple tools. Building on this capability, we introduce an observation-level intervention: a frozen, tool-integrated teacher identifies the missing spatiotemporal dependency and provides a minimal evidence patch (e.g., timestamps, regions etc.) from the original video while the question remains unchanged. The student answers again with the added context, and training updates with a chosen-rollout scheme integrated into Group Relative Policy Optimization (GRPO). We further propose a Robust Improvement Reward (RIR) that aligns optimization with two goals: outcome validity through correct answers and dependency alignment through rationales that reflect the cited evidence. Advantages are group-normalized across the batch, preserving on-policy exploration while directing it along causally meaningful directions with minimal changes to the training stack. Experiments on various related benchmarks show consistent accuracy gains and strong generalization. Code will be available at https://jethrojames.github.io/FFR/.

While Multimodal Large Language Models (MLLMs) excel at generic video understanding, their ability to support specialized, rule-grounded decision-making remains insufficiently explored. In this paper, we introduce RefereeBench, the first large-scale benchmark for evaluating MLLMs as automatic sports referees. Spanning 11 sports with 925 curated videos and 6,475 QA pairs, RefereeBench evaluates five core officiating abilities: foul existence, foul and penalty classification, foul and penalty reasoning, entity perception, and temporal grounding. The benchmark is fully human-annotated to ensure high-quality annotations grounded in authentic officiating logic and multimodal evidence. Extensive evaluations of state-of-the-art MLLMs show that even the strongest models, such as Doubao-Seed-1.8 and Gemini-3-Pro, achieve only around 60% accuracy, while the strongest open-source model, Qwen3-VL, reaches only 47%. These results indicate that current models remain far from being reliable sports referees. Further analysis shows that while models can often identify incidents and involved entities, they struggle with rule application and temporal grounding, and frequently over-call fouls on normal clips. Our benchmark highlights the need for future MLLMs that better integrate domain knowledge and multimodal understanding, advancing trustworthy AI-assisted officiating and broader multimodal decision-making.

Video understanding requires identifying and reasoning over semantically discriminative visual objects across frames, yet existing object-agnostic solutions struggle to effectively handle substantial object variations over time. To address this, we introduce Chain-of-Glimpse, a search-guided progressive object-grounded reasoning framework that explicitly anchors each reasoning step to specific visual evidence regions, enabling compositional and multi-step decision-making. Formally, Chain-of-Glimpse formulates video reasoning as a step-by-step process that incrementally builds spatially grounded traces around task-relevant visual objects, thereby mitigating over-reliance on saliency-driven cues. Specifically, Chain-of-Glimpse features a search-guided controller, optimized via reinforcement learning with a format reward that significantly incentivizes grounding capability, to iteratively ground visual evidence regions and form reliable reasoning trajectories, yielding accurate and interpretable multi-step decisions. Extensive evaluations on both in domain NExTQA and out-of-domain Video-Holmes, CG-Bench Reasoning, and VRBench benchmarks demonstrate consistent performance gains, robustness and generalization of Chain-of-Glimpse across diverse video reasoning tasks.

While video large language models (Video-LLMs) excel in understanding slow-paced, real-world egocentric videos, their capabilities in high-velocity, information-dense virtual environments remain under-explored. Existing benchmarks focus on daily activities, yet lack a rigorous testbed for evaluating fast, rule-bound reasoning in virtual scenarios. To fill this gap, we introduce EgoEsportsQA, a pioneering video question-answering (QA) benchmark for grounding perception and reasoning in expert esports knowledge. We curate 1,745 high-quality QA pairs from professional matches across 3 first-person shooter games via a scalable six-stage pipeline. These questions are structured into a two-dimensional decoupled taxonomy: 11 sub-tasks in the cognitive capability dimension (covering perception and reasoning levels) and 6 sub-tasks in the esports knowledge dimension. Comprehensive evaluations of state-of-the-art Video-LLMs reveal that current models still fail to achieve satisfactory performance, with the best model only 71.58%. The results expose notable gaps across both axes: models exhibit stronger capabilities in basic visual perception than in deep tactical reasoning, and they grasp overall macro-progression better than fine-grained micro-operations. Extensive ablation experiments demonstrate the intrinsic weaknesses of current Video-LLM architectures. Further analysis suggests that our dataset not only reveals the connections between real-world and virtual egocentric domains, but also offers guidance for optimizing downstream esports applications, thereby fostering the future advancement of Video-LLMs in various egocentric environments.

We benchmark how internal reasoning traces, which we call thought streams, affect video scene understanding in vision-language models. Using four configurations of Google's Gemini 2.5 Flash and Flash Lite across scenes extracted from 100 hours of video, we ask three questions: does more thinking lead to better outputs, where do the gains stop, and what do these models actually think about? We introduce three evaluation metrics. Contentfulness measures how much of the thought stream is useful scene content versus meta-commentary. Thought-Final Coverage measures how faithfully the thought stream translates into the final output. Dominant Entity Analysis identifies which subjects, actions, and settings the model focuses on. GPT-5 serves as an independent judge. We find that quality gains from additional thinking plateau quickly, with most improvement occurring in the first few hundred tokens. Flash Lite offers the best balance between quality and token usage. Tight reasoning budgets cause the model to add content in the final output that it never reasoned about, a form of compression-step hallucination. Despite being different model tiers, Flash and Flash Lite produce similar thought streams, though they differ in style: Flash discusses its reasoning process, while Lite focuses on describing the scene.

Game engines hold what video models struggle to learn: a complete, explicit world state behind every frame. We turn one into a data instrument. GEST-Engine, our production-grade open-source system, deterministically executes Graphs of Events in Space and Time (GESTs), whether procedurally generated or derived from text, into videos of synchronized multi-actor scenarios, recording ground truth as it renders: 3D entity and camera state, pairwise spatial relations, event-to-frame mappings, instance segmentation, and long descriptions, at zero marginal annotation cost. With it we release GTASA, a 938-video sample of what the system can generate at arbitrary scale, carrying, to our knowledge, the densest spatial-relation coverage of any video dataset: a complete entity-pair relation graph at every frame, ~84x denser than the state of the art, frame-for-frame. We validate GTASA both qualitatively, through human evaluation of physical validity and semantic alignment where frontier neural generators, given the same prompts, largely fail, and quantitatively, with GTASA pretraining improving VideoLLaMA3 video captioning. Probing six frozen video encoders across 11 spatio-temporal tasks enabled by GTASA's exact 3D ground truth, a previously untestable inter-entity relational probe of frozen video features, reveals that who-is-near-whom barely rises above chance for all of them. We release the engine, the corpus, and the benchmark, making this gap a measurable, trainable target.

Video question answering (VideoQA) is a challenging task that requires integrating spatial, temporal, and semantic information to capture the complex dynamics of video sequences. Although recent advances have introduced various approaches for video understanding, most existing methods still rely on locating relevant frames to answer questions rather than reasoning through the evolving storyline as humans do. Humans naturally interpret videos through coherent storylines, an ability that is crucial for making robust and contextually grounded predictions. To address this gap, we propose SVAgent, a storyline-guided cross-modal multi-agent framework for VideoQA. The storyline agent progressively constructs a narrative representation based on frames suggested by a refinement suggestion agent that analyzes historical failures. In addition, cross-modal decision agents independently predict answers from visual and textual modalities under the guidance of the evolving storyline. Their outputs are then evaluated by a meta-agent to align cross-modal predictions and enhance reasoning robustness and answer consistency. Experimental results demonstrate that SVAgent achieves superior performance and interpretability by emulating human-like storyline reasoning in video understanding.

Human perception interprets video dynamics by constructing structured mental representations of entities, actions, and temporal boundaries before engaging in high-level reasoning. In contrast, current Video-LLMs typically rely on unstructured Chain-of-Thought (CoT), where crucial visual cues are obscured by verbose narration. While recent efforts ground reasoning in intermediate visual evidence, they rely on task-specific schemas that fail to generalize across diverse video understanding tasks. To address this, we introduce Structured Event Evidence, a unified, temporally ordered schema designed for general video reasoning. However, post-training a model to extract structured evidence while reasoning introduces a severe multi-objective optimization bottleneck, as candidate rollouts must simultaneously satisfy schema compliance, task accuracy, and reasoning budget. Standard reinforcement learning (RL) algorithms attempt to balance these objectives via static scalarization, enforcing a rigid accuracy-length trade-off across training and obscuring fine-grained reward signals. We resolve this by proposing Group-Relative Advantage Balancing (GRAB), which dynamically computes Pareto-optimal objective weights within each rollout group via a minimum-norm criterion over centered rewards. Operating strictly in reward space, GRAB incurs virtually zero computational overhead (<0.002% relative runtime to GRPO) and eliminates brittle weight tuning. Supported by our new PARSE-60K dataset (60K annotations across 35K videos) and a three-stage progressive curriculum, we train PARSE-4B. Across temporal grounding and video question-answering benchmarks, GRAB generates 13% to 39% fewer tokens while maintaining competitive accuracy with exhaustively hand-tuned static weights -- without requiring hyperparameter weight sweeps.

Video reasoning has advanced with large multimodal models (LMMs), yet their inference is often a single pass that returns an answer without verifying whether the reasoning is evidence-aligned. We introduce Reinforce to Learn, Elect to Reason (RLER), a dual paradigm that decouples learning to produce evidence from obtaining a reliable answer. In RLER-Training, we optimize the policy with group-relative reinforcement learning (RL) and 3 novel task-driven rewards: Frame-sensitive reward grounds reasoning on explicit key frames, Think-transparency reward shapes readable and parsable reasoning traces, and Anti-repetition reward boosts information density. These signals teach the model to emit structured, machine-checkable evidence and potentiate reasoning capabilities. In RLER-Inference, we apply a train-free orchestrator that generates a small set of diverse candidates, parses their answers and cited frames, scores them by evidence consistency, confidence, transparency, and non-redundancy, and then performs a robust evidence-weighted election. This closes the loop between producing and using evidence, improving reliability and interpretability without enlarging the model. We comprehensively evaluate RLER against various open-source and RL-based LMMs on 8 representative benchmarks. RLER achieves state of the art across all benchmarks and delivers an average improvement of 6.3% over base models, while using on average 3.1 candidates per question, indicating a favorable balance between compute and quality. The results support a simple thesis: making evidence explicit during learning and electing by evidence during inference is a robust path to trustworthy video reasoning.

When video reasoning requires external knowledge, many systems with large multimodal models (LMMs) adopt retrieval augmentation to supply the missing context. Appending textual or multi-clip evidence, however, forces heterogeneous signals into a single attention space. We observe diluted attention and higher cognitive load even on non-long videos. The bottleneck is not only what to retrieve but how to represent and fuse external knowledge with the video backbone.We present Graph-to-Frame RAG (G2F-RAG), a training free and auditable paradigm that delivers knowledge in the visual space. On the offline stage, an agent builds a problem-agnostic video knowledge graph that integrates entities, events, spatial relations, and linked world knowledge. On the online stage, a hierarchical multi-agent controller decides whether external knowledge is needed, retrieves a minimal sufficient subgraph, and renders it as a single reasoning frame appended to the video. LMMs then perform joint reasoning in a unified visual domain. This design reduces cognitive load and leaves an explicit, inspectable evidence trail.G2F-RAG is plug-and-play across backbones and scales. It yields consistent gains on diverse public benchmarks, with larger improvements in knowledge-intensive settings. Ablations further confirm that knowledge representation and delivery matter. G2F-RAG reframes retrieval as visual space knowledge fusion for robust and interpretable video reasoning.

Video-grounded numerical reasoning requires Vision-Language Models (VLMs) to identify, track, and combine quantitative evidence across frames, actions, and scene changes. Existing benchmarks provide fragmented coverage: general VideoQA includes counting among broader tasks, while dedicated benchmarks focus on repetition counting, ultra-long-video enumeration, or instructional mathematics. We introduce VidNum, a manually curated and independently verified benchmark containing 1,167 multiple-choice questions. Its three task groups distinguish Direct and Distinct Enumeration, Conditioned and Structured Enumeration, and Compositional Quantitative Reasoning. Question-level annotations further identify the evidence target, counting structure, and required reasoning operation. The best evaluated VLM reaches 59.8% accuracy, compared with 98.2% for human annotators, and no evaluated open-weight model exceeds 45%. Stratified analyses reveal that failures are not uniformly distributed: structured target construction and action-grounded compositional reasoning form recurring bottlenecks across models. Zero-shot chain-of-thought prompting is not a reliable remedy: it recovers some errors but breaks previously correct answers, with effects that vary across models and task structures. VidNum therefore supports diagnostic analysis beyond a single aggregate score.

We present Scene-Graph Based Multi-Modal Traffic Agent (SGTA), a modular framework for traffic video understanding that combines structured scene graphs with multi-modal reasoning. It constructs a traffic scene graph from roadside videos using detection, tracking, and lane extraction, followed by tool-based reasoning over both symbolic graph queries and visual inputs. SGTA adopts ReAct to process interleaved reasoning traces from large language models with tool invocations, enabling interpretable decision-making for complex video questions. Experiments on selected TUMTraffic VideoQA dataset sample demonstrate that SGTA achieves competitive accuracy across multiple question types while providing transparent reasoning steps. These results highlight the potential of integrating structured scene representations with multi-modal agents for traffic video understanding.

We introduce STRIVE (SpatioTemporal Reinforcement with Importance-aware Variant Exploration), a structured reinforcement learning framework for video question answering. While group-based policy optimization methods have shown promise in large multimodal models, they often suffer from low reward variance when responses exhibit similar correctness, leading to weak or unstable advantage estimates. STRIVE addresses this limitation by constructing multiple spatiotemporal variants of each input video and performing joint normalization across both textual generations and visual variants. By expanding group comparisons beyond linguistic diversity to structured visual perturbations, STRIVE enriches reward signals and promotes more stable and informative policy updates. To ensure exploration remains semantically grounded, we introduce an importance-aware sampling mechanism that prioritizes frames most relevant to the input question while preserving temporal coverage. This design encourages robust reasoning across complementary visual perspectives rather than overfitting to a single spatiotemporal configuration. Experiments on six challenging video reasoning benchmarks including VideoMME, TempCompass, VideoMMMU, MMVU, VSI-Bench, and PerceptionTest demonstrate consistent improvements over strong reinforcement learning baselines across multiple large multimodal models. Our results highlight the role of structured spatiotemporal exploration as a principled mechanism for stabilizing multimodal reinforcement learning and improving video reasoning performance.

Recent video reasoning models have shown strong results on temporal and multimodal understanding, yet they depend on large-scale supervised data and multi-stage training pipelines, making them costly to train and difficult to adapt to new domains. In this work, we leverage the paradigm of Test-Time Reinforcement Learning on video-language data to allow for adapting a pretrained model to incoming video samples at test-time without explicit labels. The proposed test-time adaptation for video approach (TTA-Vid) combines two components that work simultaneously: (1) a test-time adaptation that performs step-by-step reasoning at inference time on multiple frame subsets. We then use a batch-aware frequency-based reward computed across different frame subsets as pseudo ground truth to update the model. It shows that the resulting model trained on a single batch or even a single sample from a dataset, is able to generalize at test-time to the whole dataset and even across datasets. Because the adaptation occurs entirely at test time, our method requires no ground-truth annotations or dedicated training splits. Additionally, we propose a multi-armed bandit strategy for adaptive frame selection that learns to prioritize informative frames, guided by the same reward formulation. Our evaluation shows that TTA-Vid yields consistent improvements across various video reasoning tasks and is able to outperform current state-of-the-art methods trained on large-scale data. This highlights the potential of test-time reinforcement learning for temporal multimodal understanding.

Video diffusion models exhibit emergent reasoning capabilities like solving mazes and puzzles, yet little is understood about how they reason during generation. We take a first step towards understanding this and study the internal planning dynamics of video models using 2D maze solving as a controlled testbed. Our investigations reveal two findings. Our first finding is early plan commitment: video diffusion models commit to a high-level motion plan within the first few denoising steps, after which further denoising alters visual details but not the underlying trajectory. Our second finding is that path length, not obstacle density, is the dominant predictor of maze difficulty, with a sharp failure threshold at 12 steps. This means video models can only reason over long mazes by chaining together multiple sequential generations. To demonstrate the practical benefits of our findings, we introduce Chaining with Early Planning, or ChEaP, which only spends compute on seeds with promising early plans and chains them together to tackle complex mazes. This improves accuracy from 7% to 67% on long-horizon mazes and by 2.5x overall on hard tasks in Frozen Lake and VR-Bench across Wan2.2-14B and HunyuanVideo-1.5. Our analysis reveals that current video models possess deeper reasoning capabilities than previously recognized, which can be elicited more reliably with better inference-time scaling.

Soccer commentary plays a crucial role in enhancing the soccer game viewing experience for audiences. Previous studies in automatic soccer commentary generation typically adopt an end-to-end method to generate anonymous live text commentary. Such generated commentary is insufficient in the context of real-world live televised commentary, as it contains anonymous entities, context-dependent errors and lacks statistical insights of the game events. To bridge the gap, we propose GameSight, a two-stage model to address soccer commentary generation as a knowledge-enhanced visual reasoning task, enabling live-televised-like knowledgeable commentary with accurate reference to entities (players and teams). GameSight starts by performing visual reasoning to align anonymous entities with fine-grained visual and contextual analysis. Subsequently, the entity-aligned commentary is refined with knowledge by incorporating external historical statistics and iteratively updated internal game state information. Consequently, GameSight improves the player alignment accuracy by 18.5% on SN-Caption-test-align dataset compared to Gemini 2.5-pro. Combined with further knowledge enhancement, GameSight outperforms in segment-level accuracy and commentary quality, as well as game-level contextual relevance and structural composition. We believe that our work paves the way for a more informative and engaging human-centric experience with the AI sports application. Demo Page: https://gamesight2025.github.io/gamesight2025

Vision-language models (VLMs) have shown impressive capabilities across diverse tasks, motivating efforts to leverage these models to supervise robot learning. However, when used as evaluators in reinforcement learning (RL), today's strongest models often fail under partial observability and distribution shift, enabling policies to exploit perceptual errors rather than solve the task. We introduce SOLE-R1 (Self-Observing LEarner), a video-language reasoning model explicitly designed to serve as the sole reward signal for online RL. Given only raw video observations and a natural-language goal, SOLE-R1 performs per-timestep spatiotemporal chain-of-thought (CoT) reasoning and produces dense estimates of task progress that can be used directly as rewards. To train SOLE-R1, we develop a large-scale video trajectory and reasoning synthesis pipeline that generates temporally grounded CoT traces aligned with continuous progress supervision. This data is combined with foundational spatial and multi-frame temporal reasoning, and used to train the model with a hybrid framework that couples supervised fine-tuning with RL from verifiable rewards. Across four different simulation environments and a real-robot setting, SOLE-R1 enables zero-shot online RL from random initialization: robots learn previously unseen manipulation tasks without ground-truth rewards, success indicators, demonstrations, or task-specific tuning. SOLE-R1 succeeds on 24 unseen tasks and substantially outperforms strong vision-language rewarders, including Robometer, RoboReward, ReWiND, GPT-5, and Gemini-3-Pro, while exhibiting markedly greater robustness to reward hacking. We release all models, data, code, and demos at the anonymous page: https://philip-mit.github.io/sole-r1/

Video generation models produce visually coherent content but struggle with tasks requiring spatial reasoning and multi-step planning. Reinforcement learning (RL) offers a path to improve generalization, but its effectiveness in video reasoning hinges on reward design -- a challenge that has received little systematic study. We investigate this problem by adapting Group Relative Policy Optimization (GRPO) to flow-based video models and training them on maze-solving and robotic navigation tasks. We first show that multimodal reward models fail catastrophically in this setting. To address this, we design verifiable reward functions grounded in objective task metrics. For structured game environments, we introduce a multi-component trajectory reward. For robotic navigation, we propose an embedding-level verifiable reward. Our experiments show that RL fine-tuning with verifiable rewards improves generalization. For example, on complex 3D mazes, our model improves exact match accuracy by 29.1% over the SFT baseline, and on trap-avoidance tasks by 51.4%. Our systematic reward analysis reveals that verifiable rewards are critical for stable training, while multimodal reward models could lead to degenerate solutions. These findings establish verifiable reward design as a key enabler for robust video reasoning. Code will be publicly available.

We introduce PerceptionComp, a manually annotated benchmark for complex, long-horizon, perception-centric video reasoning. PerceptionComp is designed so that no single moment is sufficient: answering each question requires multiple temporally separated pieces of visual evidence and compositional constraints under conjunctive and sequential logic, spanning perceptual subtasks such as objects, attributes, relations, locations, actions, and events, and requiring skills including semantic recognition, visual correspondence, temporal reasoning, and spatial reasoning. The benchmark contains 1,114 highly complex questions on 279 videos from diverse domains including city walk tours, indoor villa tours, video games, and extreme outdoor sports, with 100% manual annotation. Human studies show that PerceptionComp requires substantial test-time thinking and repeated perception steps: participants take much longer than on prior benchmarks, and accuracy drops to near chance (18.97%) when rewatching is disallowed. State-of-the-art MLLMs also perform substantially worse on PerceptionComp than on existing benchmarks: the best model in our evaluation, Gemini-3-Flash, reaches only 45.96% accuracy in the five-choice setting, while open-source models remain below 40%. These results suggest that perception-centric long-horizon video reasoning remains a major bottleneck, and we hope PerceptionComp will help drive progress in perceptual reasoning.

Most vision-language systems are static observers: they describe pixels, do not act, and cannot safely improve under shift. This passivity limits generalizable, physically grounded visual intelligence. Learning through action, not static description, is essential beyond curated data. We present Pixelis, a pixel-space agent that operates directly on images and videos via a compact set of executable operations (zoom/crop, segment, track, OCR, temporal localization) and learns from its consequences. Pixelis trains in three phases: (1) Supervised Fine-Tuning learns a pixel-tool grammar from Chain-of-Thought-Action traces with a masked imitation loss that upweights operation/argument tokens and auxiliary heads to stabilize pixel-grounded arguments; (2) Curiosity-Coherence Reward Fine-Tuning optimizes a dual-drive objective marrying prediction-error curiosity with adjacent-step coherence and a mild efficiency prior under a KL anchor, yielding short, valid, structured toolchains; (3) Pixel Test-Time RL performs label-free adaptation by retrieving neighbors, voting over complete trajectories rather than answers, and updating toward short, high-fidelity exemplars while constraining drift with a KL-to-EMA safety control. Across six public image and video benchmarks, Pixelis yields consistent improvements: the average relative gain is +4.08% over the same 8B baseline (peaking at +6.03% on VSI-Bench), computed as (ours-baseline)/baseline, while producing shorter, auditable toolchains and maintaining in-corridor KL during test-time learning. Acting within pixels, rather than abstract tokens, grounds multimodal perception in the physical world, linking visual reasoning with actionable outcomes, and enables embodied adaptation without external feedback.

The dense, temporal nature of video presents a profound challenge for automated analysis. Despite the use of powerful Vision-Language Models, prevailing methods for video understanding are limited by the inherent disconnect between reasoning and perception: they rely on static, pre-processed information and cannot actively seek raw evidence from video as their understanding evolves. To address this, we introduce LensWalk, a flexible agentic framework that empowers a Large Language Model reasoner to control its own visual observation actively. LensWalk establishes a tight reason-plan-observe loop where the agent dynamically specifies, at each step, the temporal scope and sampling density of the video it observes. Using a suite of versatile, Vision-Language Model based tools parameterized by these specifications, the agent can perform broad scans for cues, focus on specific segments for fact extraction, and stitch evidence from multiple moments for holistic verification. This design allows for progressive, on-demand evidence gathering that directly serves the agent's evolving chain of thought. Without requiring any model fine-tuning, LensWalk delivers substantial, plug-and-play performance gains on multiple model recipes, boosting their accuracy by over 5% on challenging long-video benchmarks like LVBench and Video-MME. Our analysis reveals that enabling an agent to control how it sees is key to unlocking more accurate, robust, and interpretable video reasoning.

Referring Video Object Segmentation (RVOS) aims to segment a target object throughout a video given a natural language query. Training-free methods for this task follow a common pipeline: a MLLM selects keyframes, grounds the referred object within those frames, and a video segmentation model propagates the results. While intuitive, this design asks the MLLM to make temporal decisions before any object-level evidence is available, limiting both reasoning quality and spatio-temporal coverage. To overcome this, we propose AgentRVOS, a training-free agentic pipeline built on the complementary strengths of SAM3 and a MLLM. Given a concept derived from the query, SAM3 provides reliable perception over the full spatio-temporal extent through generated mask tracks. The MLLM then identifies the target through query-grounded reasoning over this object-level evidence, iteratively pruning guided by SAM3's temporal existence information. Extensive experiments show that AgentRVOS achieves state-of-the-art performance among training-free methods across multiple benchmarks, with consistent results across diverse MLLM backbones. Our project page is available at: https://cvlab-kaist.github.io/AgentRVOS/.