video analytics teams

Abstract: Audio-visual instance segmentation (AVIS) segments and tracks individual sound-producing objects in videos. Existing benchmarks predominantly use monaural audio, which lacks inter-channel cues for spatially distinguishing concurrent sound sources. We introduce AVPS-360 (Audio-Visual Person Segmentation 360), a spatial audio-visual benchmark bridging AVIS and sound event localisation and detection (SELD). Built upon STARSS23, it pairs first-order Ambisonics audio with synchronised 360° video across 230 recordings totalling 10.5 hours. Recordings average 164.0 seconds. The dataset captures intermittent sound production in real-world interactions dominated by speech, alongside laughter, clapping, and footsteps. It provides 31,678 manually reviewed instance masks linked to source identities and sound-event labels. We also present a baseline that integrates spatial audio with visual features for frame-level segmentation and identity association across frames. Experiments on AVPS-360 show that existing AVIS methods struggle with detection and identity association in long recordings with intermittent sound production, while our baseline substantially improves performance in both aspects.

Abstract: Video captioning aims to preserve visual information through compact descriptions, predominantly using natural language. However, precisely expressing fine-grained spatial structures and temporal dynamics in natural language often leads to lengthy and potentially ambiguous descriptions. In this work, we explore Code as Video Caption, a paradigm where executable programs serve as a complementary representation of visual content through graphical primitives, layout rules, and functions of time. To examine whether code captions effectively preserve video information, we compare them with natural language captions on downstream question answering and video reconstruction tasks. Our results show that code captions achieve superior reasoning accuracy and reconstruction fidelity. Despite these advantages, realizing the potential of code captioning depends on whether current multimodal large language models (MLLMs) can generate faithful code captions. To systematically assess this capability, we introduce CodeCap-Bench, which comprises 323 videos spanning five categories of structured 2D content, along with fine-grained annotations and a multi-level evaluation framework. Extensive experiments reveal that while leading proprietary MLLMs demonstrate promising code captioning capabilities, they still struggle to faithfully capture complex temporal dynamics, with open-source models lagging substantially behind.

Abstract: Concept Bottleneck Models (CBMs), which route predictions through a layer of human-interpretable concepts, have substantially improved the interpretability of visual classification by exposing an explicit, language-based rationale for each prediction. Most existing work focuses on static images, and temporal dynamics in videos have been underexplored in conceptual representation. We propose a framework that extends concept-based image recognition to procedural video understanding. It represents a multi-step video as a temporal sequence of segment-level concept score vectors and feeds this sequence into a Transformer encoder to predict the class label. We also explore a way to estimate the contribution of each concept for further interpretability. We evaluated three concept sets of different granularities on cooking and household-activity videos and found that coarser-grained concepts are sufficient for classifying these videos, which implies that fine-grained textual descriptions may be unnecessary.

Abstract: Large vision-language models (VLMs), which process both visual inputs and text queries, incur high inference costs due to the large number of visual tokens, making visual token pruning a natural approach for improving efficiency. Recent work has leveraged query information to guide token selection, often through heuristics derived from attention patterns or token dynamics. However, these approaches do not explicitly exploit the underlying structure of query-vision interactions. We address this gap by observing that the query-vision interaction matrix, naturally induced by the query and key projections already present in transformer attention, is effectively low-rank, with a small number of dominant latent interaction modes capturing query-relevant semantics. Motivated by this observation, we propose Query-Vision Decomposition (QViD), a novel training-free query-aware visual token pruning method that exploits this structure. Extensive experiments on both image and video understanding benchmarks show that QViD consistently improves performance under matched token budgets, with the largest gains appearing under aggressive compression regimes.

Abstract: In video action detection, modeling the temporal dependencies in continuous video streams is crucial. Previous methods based on deep neural networks can perform frame-level video recognition tasks. However, these methods struggle to model dynamic temporal patterns, facing challenges such as redundant frame information, difficulty balancing global and local features, and insufficient fusion of visual and textual features. To address these issues, we propose a video temporal attention module based on the Multi-Scale Latent Temporal Compression Attention, which aggregates continuous temporal sequences into feature representations across diverse temporal scales while efficiently reducing redundancy. We also introduce a Video-Text Cross Attention Fusion module that leverages cross-attention at both latent and standard feature levels, achieving fine-grained fusion between video and textual features to enhance semantic alignment and feature interaction. Experimental results across multiple video recognition benchmarks demonstrate that our proposed methods achieve superior performance compared to existing methods, validating the effectiveness and generalization capability.

Abstract: Amodal video segmentation asks for the complete mask of an object in every frame, including the parts hidden behind other objects. The best current methods generate these masks with fine-tuned video diffusion models, which need many denoising steps and an extra depth network, and give a different answer for every sample. Self-supervised video world models of the JEPA family are trained on a closely related task: predicting the representation of masked space-time regions from the visible ones. We therefore treat occlusion as masking. The tokens that the object's visible mask does not cover are dropped, the pre-trained V-JEPA 2.1 predictor fills them in, and a light head decodes the observed and predicted features into amodal masks. The resulting model, OccJEPA (Occlusion-as-masking JEPA), is deterministic and 33.5x faster than the diffusion-based state of the art, and more accurate on SAIL-VOS, on TAO-Amodal, including a held-out test split (+4.00 AP50) and visible masks tracked by SAM2, and on Kubric. Controlled experiments trace the gain to the predictor's pre-trained weights rather than to its capacity, for world models from 87M to 1.85B parameters. Code is available at https://github.com/zgp110/occjepa.

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/}

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.

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.

Abstract: Segment Anything Model 3 (SAM 3) is a foundation model for image and video segmentation that sets the state of the art in visual object tracking. Its tracker, inherited from SAM 2, stores information from past frames in a memory bank, managed by a simple First-In-First-Out rule, allowing temporal consistency in complex videos. Several recent methods improve SAM 2 with hand-crafted memory update rules, each targeting a specific failure mode, such as distractors, occlusions, or object motion. However, their gains transfer only partially to the stronger SAM 3 backbone. We instead present a more general, data-driven approach that frames memory control as a sequential decision-making task. SAM3RL is an extension of the SAM 3 tracker with a lightweight memory controller, trained by reinforcement learning, that decides which frames enter the memory bank in order to maximize tracking performance. SAM3RL sets a new state of the art on 10 of 12 tracking and video object segmentation benchmarks, outperforming all hand-crafted rules without any runtime overhead. The same approach applied to SAM 2 gives SAM2RL, which outperforms hand-crafted rules on 9 of the 12 benchmarks, showing the generality of our method across model generations.

Abstract: We propose DeltaStreamer, a novel agentic framework for streaming video understanding inspired by how humans observe, remember, and investigate. Streaming video understanding requires a system to retain evidence before knowing which details future queries will need. Existing approaches often obscure event transitions and omit relevant details when compressing video history, and cannot recover missing visual evidence. DeltaStreamer addresses these limitations by combining change-aware memory with query-driven planning and targeted evidence gathering with replay. Before a query arrives, it continuously observes the stream and explicitly records visual changes in structured memory, preserving how events unfold rather than only what is present. When a query arrives, a query-driven planner identifies the required evidence and guides memory retrieval and targeted replay. By revisiting historical frames, the agent can recover visual details omitted during observation and ground its answer in evidence beyond the retained descriptions. This connects continuous observation with active investigation tailored to each question. Experiments on OVO-Bench and StreamingBench demonstrate improvements over representative streaming video understanding methods, with overall scores of 66.07% and 59.53%, respectively. Project page: https://bbbbchan.github.io/DeltaStreamer/

Abstract: LLM system capabilities depend not only on model parameters but also on the executable harnesses that organize information acquisition, tool use, reasoning, and memory. Recursive harness evolution can automate their design, but how it can reliably discover generalizable systems from limited, noisy, and repeatedly reused feedback remains unclear. We study this question in agentic video understanding, where rich multimodal evidence and coupled mechanisms make harness optimization challenging. We first introduce VideoHarness-BRSI, a recursive self-improvement framework that evolves video harnesses around a frozen vision-language model. Analyzing over 4,000 evolved candidates, we find it competitive with hand-engineered systems but identify three interconnected failure modes: validation overfitting, search-budget dilution, and the early-experience trap. We trace all three to a common root: the same imperfect validation evidence drives both harness revisions and the proposer's search strategy. We therefore propose VideoHarness-SRSI, an evidence-governed framework in which every search decision rests on evidence that is transferable, sufficient, and gathered before the search commits. It restricts the proposer to content-free feedback, informs exploration with instrumented warm-up experiments, and develops mechanisms in isolated probe-race branches before factorial composition and joint refinement. Across four video understanding benchmarks, VideoHarness-SRSI improves the equally weighted mean held-out score by 6.2% relative to VideoHarness-BRSI and reduces the mean validation-test gap by 37.9%. Reliable harness evolution thus depends on how search is organized and informed, not only on how well it optimizes validation performance.

Abstract: Repetition counting requires identifying and enumerating action cycles over time. While state-of-the-art video language models describe actions fluently, they fail to reliably count how many times an action repeats. To investigate this limitation, we introduce RepProbe, a benchmark of 42K real and synthetic video clips containing exactly k in [1, 20] complete repetitions, with controls for clip position, temporal coverage, and duration. These controls allow us to study the counting failure directly and locate where in the model it arises. Counting accuracy collapses at high counts across the video language models we evaluate, and controlled experiments show this collapse is not explained by frame sampling alone. We compare model outputs with linear probes trained on the final-layer representations used to generate those outputs. We find a consistent representation-to-output dissociation: substantial count information remains linearly decodable even when models fail to report the correct count. Further analysis shows this failure arises primarily at the output stage, where answers concentrate on a small set of preferred numbers rather than the available count information. The dissociation holds on both real and synthetic videos and is not explained by clip duration, position, or scene content. These results indicate that models encode repetition structure but do not reliably report it in their outputs. Project page: https://arenaa.github.io/RepProbe-webpage/

Abstract: Visual agents must turn video observations into timely decisions to track progress and choose what to do next. Building on the emerging System One paradigm, we introduce VideoJev, a family of models that cast video understanding and interaction as structured decision tasks. We adapt pretrained multimodal large language models (MLLMs) by replacing autoregressive answer decoding with a shared head that scores temporal intervals and textual options against video evidence. These scores yield event spans or answer choices, unifying temporal grounding and multiple-choice video question answering. This combines the semantic flexibility of MLLMs with efficient decisions that agents can use directly. As an interaction unfolds, new goals and intermediate results prompt further queries about the same visual history. VideoJev therefore reuses visual computation through Cached inference for queries as they arrive and Joint inference for multiple queries in a single forward pass. Across three temporal grounding benchmarks, VideoJev-4B and VideoJev-35B-A3B improve mean mIoU over their backbones by 6.7 and 10.8 points, respectively, with measured inference speedups of $2.8$-$3.1\times$. On four video QA benchmarks, both achieve performance comparable to their backbones in Thinking mode, with measured inference speedups exceeding $100\times$. These results offer a unified, reusable decision interface for composing structured visual judgments into complex, open-ended workflows.

Abstract: Cross-domain few-shot action recognition (CD-FSAR) aims to recognise unseen action classes in a target domain from a handful of labelled videos. Despite advances in representation learning, CD-FSAR inference still relies on prototypes estimated from extremely limited support examples. To assess this bottleneck, we perform an oracle analysis with the representation fixed. Constructing oracle prototypes and decision boundaries from ground-truth query labels improves 5-way 1-shot accuracy by 7--20 percentage points across five target datasets. This reveals a substantial estimation gap: the representation contains useful discriminative information, but prototypes and boundaries estimated from scarce supports fail to capture each episode's true class structure. Motivated by this, we propose TRACE, a training-free transductive inference framework that improves episode-level estimation without modifying the backbone. TRACE comprises two complementary stages. Confidence-Thresholded Transductive Prototype Refinement (CTPR) refines support-based prototypes using confidently predicted queries. The refined prototypes initialise a lightweight per-episode classifier, which Transductive Classifier Refinement (TCR) further adapts by encouraging confident, class-balanced predictions over all queries. Together, the two stages improve prototype and boundary estimation, enabling better use of information in the unlabelled queries. Across five target datasets (HMDB51, UCF101, SSv2, Diving48, and RareAct) in both 1-shot and 5-shot settings, TRACE consistently improves a strong CD-FSAR baseline, including gains of more than six percentage points on HMDB51 1-shot, while recovering much of the oracle-estimated gap. These results suggest that improving test-time estimation is a powerful, broadly applicable complement to representation learning for CD-FSAR under extreme support scarcity.

Abstract: Reliable video reasoning requires models to produce correct answers and expose supporting visual evidence. Open-o3-Video enables verifiable reasoning by embedding timestamps and object-level bounding boxes into reasoning traces, but is restricted to frames from an initial sparse sampler. We introduce Open-o3-Video-v2, which formulates video reasoning as a unified agentic evidencegrounded process. The model identifies missing information, acquires additional observations, grounds the moments and objects supporting its reasoning, and writes this evidence into the trajectory before continuing or answering. To support this behavior, we construct STGR-v2-CoT135K, including 52K multi-turn agentic grounded reasoning trajectories, and STGR-v2-RL-13K for reinforcement learning. We further propose Clue-Seek GRPO (CS-GRPO), which tracks how tool calls change accumulated evidence, rewards clue coverage and step-wise gains, and penalizes excessive, repetitive, or uninformative retrieval. On V-STaR, Open-o3-Video-v2-4B achieves 40.4% mAM and 56.3% mLGM, outperforming its Qwen3-VL-4B backbone by 8.5% and 12.6%, respectively. It also delivers consistent gains across general video understanding, long-video reasoning, and temporal grounding benchmarks. These results show that active evidence acquisition and verifiable object-level grounding are complementary for reliable video reasoning. The code and models are released at https://msalab-pku.github.io/projects/Open-o3-Video-v2.

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

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

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

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

Abstract: Koopman operator theory provides a linear framework for analyzing nonlinear dynamical systems and has become a major tool for data-driven modeling. A central challenge, however, is that finite-dimensional approximations computed by methods such as extended dynamic mode decomposition (EDMD) require the dictionary to be specified a priori. Recent machine-learning approaches address this limitation by learning the dictionary from data, predominantly using artificial neural network (ANN) autoencoder architectures. Although kernel methods offer an alternative with greater interpretability and tractability for theoretical analysis, they have received little attention in this setting. We introduce extended dynamic mode decomposition with kernel-based dictionary learning (EDMD-kDL), a kernel-based method for learning finite-dimensional Koopman embeddings directly from data. The method combines ideas from collocation methods and bilevel optimization to simultaneously learn a kernel dictionary and the corresponding Koopman approximation. We evaluate EDMD-kDL against state-of-the-art ANN-based approaches on a range of numerical experiments, including global sea-surface-temperature forecasting and learning directly from video data. Across all tested settings, EDMD-kDL achieves performance comparable to or better than the ANN-based methods. Moreover, in contrast to standard kernel methods, the proposed approach is scalable to large datasets by design since the size of the required kernel matrices depends on the number of collocation points rather than the size of the training dataset.

Abstract: Long video understanding increasingly relies on external memory to organize massive visual streams into compact representations. However, most memory-based methods dynamically adapt how information is retrieved for different questions, while largely fixing what is remembered. This mismatch makes missing details costly to recover, whereas stored information is valuable only when it can be reliably retrieved. To address this issue, we propose VideoEvolve, a novel self-evolving framework that jointly evolves memory and retrieval for long video understanding. Specifically, starting from a coarse low-frame-rate overview, VideoEvolve couples a Memory Evolver for selective memory augmentation with a Retrieval Evolver for adaptive retrieval over the evolving memory. We then co-evolve the two Evolvers through alternating agentic reinforcement learning (Agentic RL), updating one while freezing the other. To steer this alternating evolution, Bottleneck-Aware Evolution Feedback (BEF) identifies whether the current bottleneck lies in memory or retrieval and directs optimization toward the more limiting side. Furthermore, VideoEvolve introduces Capability-Aware Evolution Feedback (CEF) to alleviate downstream feedback from over-specializing memory to a fixed set of training questions, shifting training toward underdeveloped yet learnable video capabilities. By integrating Agentic RL with BEF and CEF, VideoEvolve transforms downstream reasoning experience into transferable capability updates, providing a concrete path from static long-video systems toward experience-driven, self-improving multimodal intelligence. Extensive experiments on multiple long video understanding benchmarks demonstrate the effectiveness of VideoEvolve.

Abstract: Reinforcement learning with verifiable rewards has substantially advanced multimodal reasoning, yet it remains fundamentally limited by ambiguous token-level credit assignment. While high-entropy token heuristics encourage possibility exploration, naively extending them to video reasoning tends to induce lengthy reasoning, as the model becomes overly reliant on high-entropy visual activations. Alternative approaches that rely on counterfactual-based visual token localization for credit assignment also tend to over-prioritize visual exploration at the expense of decisive reasoning cues for answer derivation, thereby exacerbating the interference from spurious visual nuances. Moreover, these methods employ static counterfactual strategies that fail to co-evolve with the policy during training. In this paper, we introduce DyCPO, a co-evolutionary framework that jointly optimizes reliable token selection and adaptive counterfactual intervention. It constructs a multi-role dependence metric to balance visual exploration and answer-relevance mining in token-wise contrastive learning, while suppressing exploration-only filler tokens and spurious visual noise. Rather than relying on static counterfactual priors, DyCPO dynamically derives counterfactual signals from the model's own successful and failed rollouts, enabling self-diagnostic analysis and co-evolution of the optimization objective with the policy. Extensive experiments on complex video reasoning and general video understanding benchmarks demonstrate consistent performance improvements, establishing DyCPO as a robust token-level credit assignment paradigm for multimodal reinforcement learning.

Abstract: Streaming vision-language models must process continuously growing video streams under a bounded compute budget, creating a persistent tension between real-time perception and long-term memory. Retrieving historical information provides a natural remedy, yet historical recall is not uniformly beneficial: unnecessary history may introduce irrelevant context into current reasoning and interfere with native real-time perception. Effective streaming memory should therefore address not only what to remember, but also when and how to access it. To this end, we introduce FlashBack, a training-free framework for selective, multi-level memory in streaming vision-language models. Before retrieving history, FlashBack draws on the semantic understanding of the frozen streaming VLM to infer whether a query calls for historical evidence. This assessment determines whether inference remains on the Native trajectory or invokes an isolated Recall trajectory. The Recall trajectory combines recent context with retrieved long-term memory through a query-local Side-KV pathway, preserving local temporal continuity without modifying the persistent Native state. We instantiate FlashBack on StreamingVLM and Mage-VL-4B and evaluate it on OVO-Bench and StreamingBench. The results show improvements on several long-horizon and memory-dependent tasks while largely preserving real-time perception, with performance competitive with strong training-based streaming methods despite requiring no additional training. Our code will be announced later.

Abstract: Query-aware keyframe selection enables multimodal large language models (MLLMs) to process long videos using only a small set of question-relevant frames. Existing score-based methods, however, typically search within a fixed, uniformly sampled candidate pool, preventing evidence outside this pool from ever being selected. Given a limited relevance-scoring budget, the key challenge is to allocate evaluations adaptively to promising frames while continuing to explore underrepresented temporal regions. We introduce FORTE, a training-free framework that addresses this challenge through two stages: adaptive relevance scoring and global keyframe optimization. Starting from sparse, uniformly distributed observations, our efficient Gaussian-process relevance predictor estimates relevance for unscored frames, exploiting temporal locality and the approximately banded kernel structure to reduce the core computation from cubic to linear time in the number of frames for fixed bandwidth. The scoring stage then selects which frames to score next by balancing predicted relevance with temporal coverage, prioritizing promising regions while also exploring less-represented parts of the video. The optimization stage selects the final keyframes by maximizing an objective that jointly captures measured relevance and temporal coverage. We derive an exact algorithm that leverages the logarithmic coverage structure to identify the optimal subset of the scored candidate pool in time linear in the pool size, for a fixed final-frame budget. Experiments on four long-video question-answering benchmarks show that FORTE achieves the highest observed mean accuracy among the compared selectors under every tested scoring budget. Further evaluations demonstrate its consistent effectiveness across different relevance scorers and downstream MLLMs.

Abstract: We introduce VideoMSN, a Masked Siamese Network framework for efficient self-supervised spatio-temporal representation learning in videos. Instead of relying on heavy 3D architectures or reconstruction-based autoencoders for learning with unlabeled data, we repurpose standard image Vision Transformers by representing videos as super images which are grids composed of frames sampled from videos. From each super image, we construct two views: one with spatial patch masking and the other with temporal frame masking, ensuring no information leakage across frames. A shared Vision Transformer (ViT) encoder aligns their embeddings using a masked Siamese loss, capturing both motion and appearance cues without reconstruction. Our decoder-free formulation leverages an image foundation model towards efficient video representation learning. Starting from pretrained DINO-v3 and DeiT-v3 image encoders, VideoMSN achieves state-of-the-art performance on Kinetics-400, UCF101, and HMDB51 while requiring up to $32\times$ fewer and $160\times$ fewer video pretraining epochs compared to prior video self-supervised learning methods. Our proposed approach also shows strong performance in low-shot classification, confirming the transferability of the learned representations in a label-scarce scenario. Project Page: https://cvir.github.io/projects/videomsn.

Abstract: Ultra-long video temporal grounding requires balancing long-range evidence search with fine-grained event understanding under a limited visual budget, yet existing agentic methods still rely largely on predefined policies and tool capabilities. Motivated by this, we propose a novel policy-tool coevolution framework that jointly evolves high-level policies and executable media tools from the agentic reasoning trajectories of a VLM, forming a reusable skill without updating model parameters. During evolution, an external skill updater distills transferable task experience in long-video temporal grounding, accordingly refining the orchestration of long-range image-based and fine-grained video-based observations. Alongside these policy updates, the updater employs its coding capabilities to upgrade existing tools or create new ones, adapting the tools to long-video evidence acquisition. Equipped with the evolved skill, the VLM autonomously orchestrates tools under the guidance of the evolved policy, coordinating image and video observations for agentic inference without relying on a separate, stronger planning model. Extensive experiments spanning five benchmarks and three VLMs show that policy-tool coevolution consistently improves temporal grounding accuracy in ultra-long videos while reducing visual token cost at inference, and that the evolved skill yields substantial performance gains on general long-video QA without additional task-specific evolution, demonstrating the effectiveness and generalizability of our framework for long-video understanding.

Abstract: 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 https://github.com/bingjunluo/Video-RSI .

Abstract: Continuous video counting requires distinguishing new observations from new objects or completed events. We introduce StaMina (State Maintenance), which learns to maintain counting state through state-conditioned updates. Recurrent visual context supports recognition; learned transitions maintain visibility, persistent identities, and completed-event records. A differentiable recurrence trains event transitions over legal paths constrained by count endpoints; visibility and association objectives train the object branch. A multi-source pipeline organizes 39.8K spatial queries and complementary event annotations into counting trajectories. On SVCBench, we evaluate counting adaptation with partial video overlap and held-out groups of linked annotations. Under prefix replay (Full) and persistent streaming (Stream), 4B and 8B models reach 41.9/36.4 and 44.9/38.2 Gaussian Precision Accuracy, respectively. The 8B model gains 10.9/3.2 points over Counting-SFT on the same queries. Matched-graph comparisons isolate phase conditioning and trajectory supervision, assessing training objectives alongside hard decisions. Online video benchmarks and count-conditioned decisions assess online understanding and task eligibility. Project Page: https://PLACEHOLDER.github.io/StaMina/

Abstract: Video Large Language Models (Video LLMs) have achieved remarkable progress in video understanding, but their inference efficiency is constrained by the large number of visual tokens produced by long videos. Recent video token compression methods increasingly introduce sophisticated strategies for token selection, pruning, and merging. This raises a fundamental question: how much of compression performance can be obtained by simply preserving the structure encoded in the visual representations? We investigate this question with SimpleCluster, a simple and training-free baseline that performs position-aware cross-frame clustering in the visual feature space and represents each cluster using the mean of its original visual features. Extensive experiments across four video understanding benchmarks and three representative Video LLMs show that SimpleCluster achieves competitive or superior performance over recent compression methods across a wide range of token retention ratios, with particularly strong robustness under extremely low retention rates (e.g., 1%). To understand this behavior, we analyze the feature space preserved by different compression methods in terms of local approximation fidelity and global coverage. The results show that stronger downstream performance is consistently associated with better preservation of the original visual feature distribution, especially its global coverage. These findings highlight feature-space preservation as an important consideration for video token compression under highly constrained token budgets. Our code is available at https://github.com/xiaozhang79/SimpleCluster.

Abstract: Multimodal large language models (MLLMs) have demonstrated impressive performance in multimodal understanding, but processing large numbers of visual tokens results in high computational costs. While many methods have been proposed to reduce the number of visual tokens, most of them rely on heuristics and are prone to discarding substantial visual information during pruning, leading to degradation in model performance. In this work, by using a semantic erasure model, we derive a general mutual information coverage objective from task log-loss and propose MiCo, a training-free two-stage pruning method. MiCo first uses visual signals to select a representative candidate pool before visual tokens enter the language model, then performs task-aware subset selection within it. At each stage, suitable observable proxies instantiate the derived objective as a monotone submodular coverage function, which MiCo greedily optimizes under the token budget. MiCo is evaluated on diverse MLLMs ranging from 7B to 13B parameters across a broad range of image and video benchmarks spanning general visual reasoning, fine-grained OCR and grounding, hallucination detection, and long-video understanding. MiCo consistently achieves the best performance across nearly all evaluated models under all pruning ratios. On LLaVA-NEXT-13B, MiCo uses only 5.6% visual tokens, retains 97.5% of baseline performance, and achieves a 3.8-fold inference speedup. Our experiments demonstrate the effectiveness of MiCo and our mutual information coverage objective for visual token pruning.

Abstract: Frame selection is an important component of long-video question answering (VQA) with Multimodal Large Language Models (MLLMs). Existing frame-selection methods improve over simple top-$k$ embedding retrieval and uniform sampling, but are typically applied under a fixed global selection budget. We introduce \textbf{MetaSampling}, a training-free, plug-and-play sampling strategy that can be applied on top of existing frame selectors. MetaSampling improves downstream VQA efficiency by dynamically reducing the number of frames passed to the MLLM while preserving, and in some cases improving, answer accuracy. We evaluate MetaSampling across 36 paired frame-selector--MLLM-backbone--VQA-benchmark configurations. MetaSampling reduces the number of selected frames in all 36 configurations and improves accuracy in 25 of them, yielding an average frame reduction of $8.9\%$ while slightly improving accuracy overall.

Abstract: Action boundary detection partitions untrimmed video into intervals without assigning action classes. We present a boundary-detection adapter operating on pre-extracted video features, learning temporal representations via groupwise selective scans. Learned group fusion and temporal modeling convert these into transition scores, which are decoded into boundary timestamps. Trained with boundary-time supervision, the class-agnostic model is evaluated on Breakfast, GTEA, and 50Salads using temporal tolerances and bipartite matching, achieving boundary $F_1$ scores of 0.457, 0.622, and 0.611. A stateful variant enables feature-streaming inference with zero neural look-ahead, one-sample peak confirmation, and bounded memory. Downstream systems can subsequently assign s

Abstract: Most keyframe selection studies focus on offline settings, assuming access to the full video and query in advance. In contrast, real-world streaming scenarios require online frame selection under unknown video duration, without access to either the query or future frames during selection. To address this, we introduce Streaming Video Memory (SVMem), a compact and representative memory of previously observed content, updated continuously as the video stream unfolds. Building on this setting, we propose the Streaming Video Memory Agent (SVMemAgent), which dynamically maintains a memory by deciding at each timestep whether to replace an existing memory frame with the incoming frame or discard it. SVMemAgent is trained using Group Relative Policy Optimization (GRPO) with task-driven rewards derived from diverse question-answer pairs, implicitly exposing the policy to a distribution of queries during training so that SVMem retains generally informative frames at inference, when queries are unavailable. Experiments on both online and offline video benchmarks show that SVMemAgent consistently outperforms online frame selection baselines and achieves competitive performance with offline methods that assume access to the full video and query. Through task-driven rewards, SVMemAgent learns an emergent keyframe selection policy that prefers frames containing textual information, which may benefit downstream VideoQA tasks.

Abstract: Wearable assistants require episodic memory over egocentric video, yet current vision-language models face bounded frame budgets, growing visual-token costs, and long-context retrieval failures. Under these practical constraints, we study whether textual captions can serve as reusable episodic memory. We define the Episodic Memory Video Caption QA task and introduce CapMem, a human-annotated benchmark with 75 videos totaling 33.7 hours, and 1,000 multiple-choice questions across 16 scenarios. On long videos (>20 min), full-coverage CaptionQA with 30s and 60s caption windows outperforms direct VideoQA for 10/12 and 8/12 models, respectively. On the same video subset, a matched-frame control across six Qwen models retains mean accuracy gains of 3.22 and 2.55 points, respectively. Our caption-guided retrieve-and-verify harness further improves accuracy by up to 5.3 points. These results support the effectiveness of caption memory for episodic reasoning over long egocentric video.

Abstract: As a key model compression technique, knowledge distillation aims to transfer knowledge from a high-capacity teacher model to a lightweight student model for enhancing the latter's performance. In this work, we reviewed the feature knowledge distillation for 3D-CNNs and observed that most feature distillation methods in video analysis are simple adaptations of those used in image analysis, often neglecting the differences of video features in the temporal dimension. To address this issue, we proposed Label-Guided Knowledge Distillation (LGKD) to guide the distillation of student model features using ground truth labels. Our method entails two components: sample-wise distillation and class-wise distillation, enabling the student model to learn feature representation of the teacher model at two levels. Sample-wise distillation utilizes label information and the teacher's probability distribution to guide the learning of features that significantly impact temporal accuracy while mitigating noise. Meanwhile, class-wise feature distillation employs a prototype network to further capture the relational knowledge among samples within the same category, enhancing the student's ability to learn higher-dimensional semantic information and improving model generalization. To demonstrate the effectiveness and superiority of our method, we conducted comprehensive experiments on two benchmark action recognition datasets, UCF101 and HMDB51, achieving competitive results.

Abstract: Long-Video Question Answering (LVQA) requires locating decisive evidence in hour-scale videos under a limited frame budget. Most training-free methods apply the same frame-selection strategy to all questions, despite substantial variation in the evidence required by different question types. Our analysis shows that the relative effectiveness of frame-selection strategies varies across semantic categories and benchmarks, motivating adaptive evidence acquisition. In this paper, we introduce AutoSkill, a source-supervised framework for automatically discovering and routing executable frame-selection skills. Starting from a small labelled source pool, LLM agents iteratively propose, implement, evaluate, and refine candidate skills. For a target benchmark, AutoSkill uses only unlabelled question and option text to induce a shared semantic taxonomy, rewrite labelled source examples into the target style, and estimate a category-to-skill mapping. Neither target videos nor target answers are used in this process. At inference time, each question is assigned one skill, which selects the frames used in a single inference of the frozen video MLLM. Across five long-video benchmark splits, AutoSkill improves Qwen2.5-VL-7B and Qwen3.5-4B by 2.4% and 1.2%, respectively, demonstrating the effectiveness of our AutoSkill.

Abstract: Real-world video understanding requires integrating visual, audio, textual, and temporal evidence distributed across a video. Yet many pipelines use a fixed video context and single-pass inference, limiting adaptive evidence acquisition when observations are incomplete, ambiguous, or conflicting. We present VLX-VR, an agentic-aware video reasoning model trained within a video reasoning framework defined by a Think--Memory--Observation loop. At each step, VLX-VR determines the needed evidence, invokes read_memory or write_memory, incorporates the returned Observation, and decides whether to continue or produce the task output. We train VLX-VR with multimodal data, including videos and agent trajectories, using reinforcement learning to learn evidence acquisition, memory use, and termination. On MINERVA, VLX-VR achieves state-of-the-art performance among the models included in our comparison, with 78.79% accuracy. Under the original three duration groups, its accuracies are 76.70%, 78.73%, and 80.92%, with a cross-duration accuracy variance of 2.97~$\mathrm{pp}^2$. On correctly answered samples, 96.20% of VLX-VR's reasoning traces are consistent with the MINERVA reference reasoning traces and the evidence described by them, while approximately 75.80% of all evaluated samples satisfy both answer correctness and this evidence-grounded trace criterion. These results show strong performance and broadly stable behavior across durations, while counting, state changes, causal reasoning, and spatial perception remain challenging.

Abstract: Video Large Language Models (VideoLLMs) are increasingly deployed in safety-critical applications such as content moderation and video analytics. To process long videos efficiently, VideoLLMs rely on frame sampling, token compression, and modality fusion, which together form an observation pipeline that reduces the raw video to a compact internal representation. Recent observation-level attacks exploit this pipeline to prevent the model from perceiving harmful content, yet no defense has been explicitly designed for this threat. We introduce DefTEval, a controlled evaluation framework that systematically assesses whether input-level adversarial defenses, which operate on the pixel content of already-sampled frames, can mitigate observation-level attacks. Across five VideoLLMs, eleven representative defenses, and five attack types, we find that input-level defenses offer limited and inconsistent protection, with harmful detection rates frequently near zero. Critically, defenses fail even against attacks that embed harmful signals in every sampled frame, indicating that the bottleneck extends beyond sampling omission to the suppression of signals that do enter the model. Token compression discards localized features, and modality fusion systematically down-weights weakened visual signals. Furthermore, defense effectiveness is dominated by model architecture rather than by the defense method itself, and detection rates vary drastically across content categories, exposing structural weaknesses in temporal reasoning. These findings demonstrate that securing VideoLLMs requires system-level robustness mechanisms spanning sampling-aware coverage guarantees, token-level preservation of safety-relevant features, and modality-balanced fusion.