Temporal Video Understanding

Multi-modal Large Language Models (MLLMs) show promise in video understanding. However, their reasoning often suffers from thinking drift and weak temporal comprehension, even when enhanced by Reinforcement Learning (RL) techniques like Group Relative Policy Optimization (GRPO). Moreover, existing RL methods usually depend on Supervised Fine-Tuning (SFT), which requires costly Chain-of-Thought (CoT) annotation and multi-stage training, and enforces fixed reasoning paths, limiting MLLMs' ability to generalize and potentially inducing bias. To overcome these limitations, we introduce Summary-Driven Reinforcement Learning (SDRL), a novel single-stage RL framework that obviates the need for SFT by utilizing a Structured CoT format: Summarize -> Think -> Answer. SDRL introduces two self-supervised mechanisms integrated into the GRPO objective: 1) Consistency of Vision Knowledge (CVK) enforces factual grounding by reducing KL divergence among generated summaries; and 2) Dynamic Variety of Reasoning (DVR) promotes exploration by dynamically modulating thinking diversity based on group accuracy. This novel integration effectively balances alignment and exploration, supervising both the final answer and the reasoning process. Our method achieves state-of-the-art performance on seven public VideoQA datasets.

Understanding tactical dynamics in badminton requires analyzing entire matches rather than isolated clips. However, existing badminton datasets mainly focus on short clips or task-specific annotations and rarely provide full-match data with dense multimodal annotations. This limitation makes it difficult to generate accurate shot captions and perform match-level analysis. To address this limitation, we introduce the first Badminton Full Match Dense (BFMD) dataset, with 19 broadcast matches (including both singles and doubles) covering over 20 hours of play, comprising 1,687 rallies and 16,751 hit events, each annotated with a shot caption. The dataset provides hierarchical annotations including match segments, rally events, and dense rally-level multimodal annotations such as shot types, shuttle trajectories, player pose keypoints, and shot captions. We develop a VideoMAE-based multimodal captioning framework with a Semantic Feedback mechanism that leverages shot semantics to guide caption generation and improve semantic consistency. Experimental results demonstrate that multimodal modeling and semantic feedback improve shot caption quality over RGB-only baselines. We further showcase the potential of BFMD by analyzing the temporal evolution of tactical patterns across full matches.

Endoscopic video analysis is essential for early gastrointestinal screening but remains hindered by limited high-quality annotations. While self-supervised video pre-training shows promise, existing methods developed for natural videos prioritize dense spatio-temporal modeling and exhibit motion bias, overlooking the static, structured semantics critical to clinical decision-making. To address this challenge, we propose Focus-to-Perceive Representation Learning (FPRL), a cognition-inspired hierarchical framework that emulates clinical examination. FPRL first focuses on intra-frame lesion-centric regions to learn static semantics, and then perceives their evolution across frames to model contextual semantics. To achieve this, FPRL employs a hierarchical semantic modeling mechanism that explicitly distinguishes and collaboratively learns both types of semantics. Specifically, it begins by capturing static semantics via teacher-prior adaptive masking (TPAM) combined with multi-view sparse sampling. This approach mitigates redundant temporal dependencies and enables the model to concentrate on lesion-related local semantics. Following this, contextual semantics are derived through cross-view masked feature completion (CVMFC) and attention-guided temporal prediction (AGTP). These processes establish cross-view correspondences and effectively model structured inter-frame evolution, thereby reinforcing temporal semantic continuity while preserving global contextual integrity. Extensive experiments on 11 endoscopic video datasets show that FPRL achieves superior performance across diverse downstream tasks, demonstrating its effectiveness in endoscopic video representation learning. The code is available at https://github.com/MLMIP/FPRL.

Accurate surgical phase recognition is essential for analyzing procedural workflows, supporting intraoperative decision-making, and enabling data-driven improvements in surgical education and performance evaluation. In this work, we present a comprehensive framework for phase recognition in pituitary tumor surgery (PTS) videos, combining self-supervised representation learning, robust temporal modeling, and scalable data annotation strategies. Our method achieves 90% accuracy on a held-out test set, outperforming current state-of-the-art approaches and demonstrating strong generalization across variable surgical cases. A central contribution of this work is the integration of a collaborative online platform designed for surgeons to upload surgical videos, receive automated phase analysis, and contribute to a growing dataset. This platform not only facilitates large-scale data collection but also fosters knowledge sharing and continuous model improvement. To address the challenge of limited labeled data, we pretrain a ResNet-50 model using the self-supervised framework on 251 unlabeled PTS videos, enabling the extraction of high-quality feature representations. Fine-tuning is performed on a labeled dataset of 81 procedures using a modified training regime that incorporates focal loss, gradual layer unfreezing, and dynamic sampling to address class imbalance and procedural variability.

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.

Video-language foundation models have proven to be highly effective in zero-shot applications across a wide range of tasks. A particularly challenging area is the intraoperative surgical procedure domain, where labeled data is scarce, and precise temporal understanding is often required for complex downstream tasks. To address this challenge, we introduce CliPPER (Contextual Video-Language Pretraining on Long-form Intraoperative Surgical Procedures for Event Recognition), a novel video-language pretraining framework trained on surgical lecture videos. Our method is designed for fine-grained temporal video-text recognition and introduces several novel pretraining strategies to improve multimodal alignment in long-form surgical videos. Specifically, we propose Contextual Video-Text Contrastive Learning (VTC_CTX) and Clip Order Prediction (COP) pretraining objectives, both of which leverage temporal and contextual dependencies to enhance local video understanding. In addition, we incorporate a Cycle-Consistency Alignment over video-text matches within the same surgical video to enforce bidirectional consistency and improve overall representation coherence. Moreover, we introduce a more refined alignment loss, Frame-Text Matching (FTM), to improve the alignment between video frames and text. As a result, our model establishes a new state-of-the-art across multiple public surgical benchmarks, including zero-shot recognition of phases, steps, instruments, and triplets. The source code and pretraining captions can be found at https://github.com/CAMMA-public/CliPPER.

Open-Vocabulary Temporal Action Detection (OV-TAD) aims to classify and localize action segments in untrimmed videos for unseen categories. Previous methods rely solely on global alignment between label-level semantics and visual features, which is insufficient to transfer temporal consistent visual knowledge from seen to unseen classes. To address this, we propose a Phase-wise Decomposition and Alignment (PDA) framework, which enables fine-grained action pattern learning for effective prior knowledge transfer. Specifically, we first introduce the CoT-Prompting Semantic Decomposition (CSD) module, which leverages the chain-of-thought (CoT) reasoning ability of large language models to automatically decompose action labels into coherent phase-level descriptions, emulating human cognitive processes. Then, Text-infused Foreground Filtering (TIF) module is introduced to adaptively filter action-relevant segments for each phase leveraging phase-wise semantic cues, producing semantically aligned visual representations. Furthermore, we propose the Adaptive Phase-wise Alignment (APA) module to perform phase-level visual-textual matching, and adaptively aggregates alignment results across phases for final prediction. This adaptive phase-wise alignment facilitates the capture of transferable action patterns and significantly enhances generalization to unseen actions. Extensive experiments on two OV-TAD benchmarks demonstrated the superiority of the proposed method.

Recent advancements in Video Large Language Models (VideoLLMs) have enabled strong performance across diverse multimodal video tasks. To reduce the high computational cost of processing dense video frames, efficiency-oriented methods such as frame selection have been widely adopted. While effective at minimizing redundancy, these methods often cause notable performance drops on tasks requiring temporal reasoning. Unlike humans, who can infer event progression from sparse visual cues, VideoLLMs frequently misinterpret temporal relations when intermediate frames are omitted. To address this limitation, we explore visual prompting (VP) as a lightweight yet effective way to enhance temporal understanding in VideoLLMs. Our analysis reveals that simply annotating each frame with explicit ordinal information helps the model perceive temporal continuity. This visual cue also supports frame-level referencing and mitigates positional ambiguity within a sparsely sampled sequence. Building on these insights, we introduce ViKey, a training-free framework that combines VP with a lightweight Keyword-Frame Mapping (KFM) module. KFM leverages frame indices as dictionary-like keys to link textual cues to the most relevant frames, providing explicit temporal anchors during inference. Despite its simplicity, our approach substantially improves temporal reasoning and, on some datasets, preserves dense-frame baseline performance with as few as 20% of frames.

Video understanding with multimodal large language models (MLLMs) remains challenging due to the long token sequences of videos, which contain extensive temporal dependencies and redundant frames. Existing approaches typically treat MLLMs as passive recognizers, processing entire videos or uniformly sampled frames without adaptive reasoning. Recent agent-based methods introduce external tools, yet still depend on manually designed workflows and perception-first strategies, resulting in inefficiency on long videos. We present EVA, an Efficient Reinforcement Learning framework for End-to-End Video Agent, which enables planning-before-perception through iterative summary-plan-action-reflection reasoning. EVA autonomously decides what to watch, when to watch, and how to watch, achieving query-driven and efficient video understanding. To train such agents, we design a simple yet effective three-stage learning pipeline - comprising supervised fine-tuning (SFT), Kahneman-Tversky Optimization (KTO), and Group Relative Policy Optimization (GRPO) - that bridges supervised imitation and reinforcement learning. We further construct high-quality datasets for each stage, supporting stable and reproducible training. We evaluate EVA on six video understanding benchmarks, demonstrating its comprehensive capabilities. Compared with existing baselines, EVA achieves a substantial improvement of 6-12% over general MLLM baselines and a further 1-3% gain over prior adaptive agent methods.

Long video understanding remains challenging for multimodal large language models (MLLMs) due to limited context windows, which necessitate identifying sparse query-relevant video segments. However, existing methods predominantly localize clues based solely on the query, overlooking the video's intrinsic structure and varying relevance across segments. To address this, we propose VideoDetective, a framework that integrates query-to-segment relevance and inter-segment affinity for effective clue hunting in long-video question answering. Specifically, we divide a video into various segments and represent them as a visual-temporal affinity graph built from visual similarity and temporal proximity. We then perform a Hypothesis-Verification-Refinement loop to estimate relevance scores of observed segments to the query and propagate them to unseen segments, yielding a global relevance distribution that guides the localization of the most critical segments for final answering with sparse observation. Experiments show our method consistently achieves substantial gains across a wide range of mainstream MLLMs on representative benchmarks, with accuracy improvements of up to 7.5% on VideoMME-long. Our code is available at https://videodetective.github.io/

Video large language models (Video-LLMs) face high computational costs due to large volumes of visual tokens. Existing token compression methods typically adopt a two-stage spatiotemporal compression strategy, relying on stage-specific metrics and an implicit assumption of spatiotemporal separability. Under extremely low retention ratios, however, such approaches often result in unbalanced allocation and loss of visual evidence essential for question answering. We reformulate token compression as a spatiotemporal allocation task within a global token retention pool. We propose a unified selection mechanism that integrates attention weights and semantic similarity to globally select tokens with high contribution and low redundancy. Unselected tokens are merged via clustering and refilled, preserving information integrity. Inside the LLM, we further introduce text-aware merging to perform secondary compression based on query relevance. Without requiring retraining, our method serves as a plug-and-play module compatible with existing Video-LLMs. Experiments show that retaining only about 2% of visual tokens preserves 90.1% of baseline performance across multiple benchmarks, while reducing FLOPs to roughly 2.6%. These benefits generalize across diverse backbones, decreasing end-to-end inference latency and memory consumption. Our unified spatiotemporal token compression strategy establishes the state-of-the-art in video understanding under ultra-low token retention.

Recognition of respiratory distress through visual inspection is a life saving clinical skill. Clinicians can detect early signs of respiratory deterioration, creating a valuable window for earlier intervention. In this study, we evaluate whether recent advances in video transformers can enable Artificial Intelligence systems to recognize the signs of respiratory distress from video. We collected videos of healthy volunteers recovering after strenuous exercise and used the natural recovery of each participants respiratory status to create a labeled dataset for respiratory distress. Splitting the video into short clips, with earlier clips corresponding to more shortness of breath, we designed a temporal ordering challenge to assess whether an AI system can detect respiratory distress. We found a ViViT encoder augmented with Lie Relative Encodings (LieRE) and Motion Guided Masking, combined with an embedding based comparison strategy, can achieve an F1 score of 0.81 on this task. Our findings suggest that modern video transformers can recognize subtle changes in respiratory mechanics.

Dense semantic segmentation in dynamic environments is fundamentally limited by the low-frame-rate (LFR) nature of standard cameras, which creates critical perceptual gaps between frames. To solve this, we introduce Anytime Interframe Semantic Segmentation: a new task for predicting segmentation at any arbitrary time using only a single past RGB frame and a stream of asynchronous event data. This task presents a core challenge: how to robustly propagate dense semantic features using a motion field derived from sparse and often noisy event data, all while mitigating feature degradation in highly dynamic scenes. We propose LiFR-Seg, a novel framework that directly addresses these challenges by propagating deep semantic features through time. The core of our method is an uncertainty-aware warping process, guided by an event-driven motion field and its learned, explicit confidence. A temporal memory attention module further ensures coherence in dynamic scenarios. We validate our method on the DSEC dataset and a new high-frequency synthetic benchmark (SHF-DSEC) we contribute. Remarkably, our LFR system achieves performance (73.82% mIoU on DSEC) that is statistically indistinguishable from an HFR upper-bound (within 0.09%) that has full access to the target frame. This work presents a new, efficient paradigm for achieving robust, high-frame-rate perception with low-frame-rate hardware. Project Page: https://candy-crusher.github.io/LiFR_Seg_Proj/#; Code: https://github.com/Candy-Crusher/LiFR-Seg.git.

The identification of hazardous driving behaviors from in-cabin video streams is essential for enhancing road safety and supporting the detection of traffic violations and unsafe driver actions. However, current temporal action localization techniques often struggle to balance accuracy with computational efficiency. In this work, we develop and evaluate a temporal action localization framework tailored for driver monitoring scenarios, particularly suitable for periodic inspection settings such as transportation safety checkpoints or fleet management assessment systems. Our approach follows a two-stage pipeline that combines VideoMAE-based feature extraction with an Augmented Self-Mask Attention (AMA) detector, enhanced by a Spatial Pyramid Pooling-Fast (SPPF) module to capture multi-scale temporal features. Experimental results reveal a distinct trade-off between model capacity and efficiency. At the feature extraction stage, the ViT-Giant backbone delivers higher representations with 88.09% Top-1 test accuracy, while the ViT-based variant proves to be a practical alternative, achieving 82.55% accuracy with significantly lower computational fine-tuning costs (101.85 GFLOPs/segment compared to 1584.06 GFLOPs/segment for Giant). In the downstream localization task, the integration of SPPF consistently improves performance across all configurations. Notably, the ViT-Giant + SPPF model achieves a peak mAP of 92.67%, while the lightweight ViT-based configuration maintains robust results.

A short video succeeds not simply because of what it shows, but because of how it schedules attention -- yet current multimodal models lack the structural grammar to parse or produce this organization. Existing models can describe scenes, answer event-centric questions, and read on-screen text, but they are far less reliable at identifying timeline-grounded units such as hooks, cut rationales, shot-induced tension, and platform-facing packaging cues. We propose SV6D (Structured Video in Six Dimensions), inspired by professional storyboard practice in film and television production, a representation framework that decomposes internet-native video into six complementary structural dimensions -- subject, aesthetics, camera language, editing, narrative, and dissemination -- with each label tied to physically observable evidence on the timeline. We formalize a unified optimization objective over SV6D that combines Hungarian-matched temporal alignment, dimension-wise semantic label distance, and quality regularization. Building on this framework, we present Leum-VL-8B, an 8B video-language model that realizes the SV6D objective through an expert-driven post-training pipeline, further refined through verifiable reinforcement learning on perception-oriented tasks. Leum-VL-8B achieves 70.8 on VideoMME (w/o subtitles), 70.0 on MVBench, and 61.6 on MotionBench, while remaining competitive on general multimodal evaluations such as MMBench-EN. We also construct FeedBench, a benchmark for structure-sensitive short-video understanding. Our results indicate that the missing layer in video AI is not pixel generation but structural representation: grounded on the timeline, linked to visible evidence, and directly consumable by downstream workflows such as editing, retrieval, recommendation, and generation control, including text-heavy internet video formats with overlays and image-text layouts.

Capsule endoscopy event detection is challenging because diagnostically relevant findings are sparse, visually heterogeneous, and embedded in long, noisy video streams, while evaluation is performed at the event level rather than by frame accuracy alone. We therefore formulate the RARE-VISION task as a metric-aligned event detection problem instead of a purely frame-wise classification task. Our framework combines two complementary backbones, EndoFM-LV for local temporal context and DINOv3 ViT-L/16 for strong frame-level visual semantics, followed by a Diverse Head Ensemble, Validation-Guided Hierarchical Fusion, and Anatomy-Aware Temporal Event Decoding. The fusion stage uses validation-derived class-wise model weighting, backbone weighting, and probability calibration, while the decoding stage applies temporal smoothing, anatomical constraints, threshold refinement, and per-label event generation to produce stable event predictions. Validation ablations indicate that complementary backbones, validation-guided fusion, and anatomy-aware temporal decoding all contribute to event-level performance. On the official hidden test set, the proposed method achieved an overall temporal [email protected] of 0.3530 and temporal [email protected] of 0.3235.

Token pruning is essential for enhancing the computational efficiency of vision-language models (VLMs), particularly for video-based tasks where temporal redundancy is prevalent. Prior approaches typically prune tokens either (1) within the vision transformer (ViT) exclusively for unimodal perception tasks such as action recognition and object segmentation, without adapting to downstream vision-language tasks; or (2) only within the LLM while leaving the ViT output intact, often requiring complex text-conditioned token selection mechanisms. In this paper, we introduce Spatio-Temporal Token Scoring (STTS), a simple and lightweight module that prunes vision tokens across both the ViT and the LLM without text conditioning or token merging, and is fully compatible with end-to-end training. By learning how to score temporally via an auxiliary loss and spatially via LLM downstream gradients, aided by our efficient packing algorithm, STTS prunes 50% of vision tokens throughout the entire architecture, resulting in a 62% improvement in efficiency during both training and inference with only a 0.7% drop in average performance across 13 short and long video QA tasks. Efficiency gains increase with more sampled frames per video. Applying test-time scaling for long-video QA further yields performance gains of 0.5-1% compared to the baseline. Overall, STTS represents a novel, simple yet effective technique for unified, architecture-wide vision token pruning.

This work presents a multi-label temporal event detection framework for video capsule endoscopy (VCE) that addresses the extreme class imbalance inherent in the Galar dataset by combining two principal contributions: an Angular Separation Loss on class prototypes and a Biological State Machine temporal decoder. The backbone remains BiomedCLIP, a biomedical vision-language foundation model. Three consecutive frames are fused through a Local Differencing Attention module that amplifies transient pathological signals by suppressing static temporal redundancy. An Anatomy Context Head then conditions pathological predictions on soft anatomical activations, exploiting the known spatial co-occurrence structure of GI findings. Learnable text-feature prompts and prototype-based logit augmentation are trained alongside an Angular Separation Loss that penalizes off-diagonal cosine similarity between class prototypes, preventing the prototype collapse that afflicts rare classes under extreme imbalance. To counteract the skewed label distribution, the training regime combines asymmetric focal loss, inverse-frequency weighted sampling, temporal Mixup, Exponential Moving Average, and per-class threshold calibration. The Biological State Machine decoder replaces naive gap merging with a physiologically grounded forward-only state transition over anatomy labels, eliminating the fragmentation artefact that produced hundreds of spurious anatomy events per video in the prior approach and reducing per-video anatomy output to 2--3 clinically realistic events. On the held-out RARE-VISION test set comprising three NaviCam examinations (161,025 frames), the updated pipeline achieves an overall temporal [email protected] of 0.3597 and [email protected] of 0.3399, representing a relative improvement of 46% and 44% respectively over the prior submission, with total inference completed in approximately 21 minutes on a single GPU.

Video understanding aims to enable models to perceive, reason about, and interact with the dynamic visual world. In contrast to image understanding, video understanding inherently requires modeling temporal dynamics and evolving visual context, placing stronger demands on spatiotemporal reasoning and making it a foundational problem in computer vision. In this survey, we present a structured overview of video understanding by organizing the literature into three complementary perspectives: low-level video geometry understanding, high-level semantic understanding, and unified video understanding models. We further highlight a broader shift from isolated, task-specific pipelines toward unified modeling paradigms that can be adapted to diverse downstream objectives, enabling a more systematic view of recent progress. By consolidating these perspectives, this survey provides a coherent map of the evolving video understanding landscape, summarizes key modeling trends and design principles, and outlines open challenges toward building robust, scalable, and unified video foundation models.

We developed a multi-label gastrointestinal video analysis pipeline based on a ResNet-50 frame classifier followed by anatomy-guided temporal event decoding. The system predicts 17 labels, including 5 anatomy classes and 12 pathology classes, from frames resized to 336x336. A major challenge was severe class imbalance, particularly for rare pathology labels. To address this, we used clipped class-wise positive weighting in the training loss, which improved rare-class learning while maintaining stable optimization. At the temporal stage, we found that direct frame-to-event conversion produced fragmented mismatches with the official ground truth. The final submission therefore combined GT-style framewise event composition, anatomy vote smoothing, and anatomy-based pathology gating with a conservative hysteresis decoder. This design improved the final temporal mAP from 0.3801 to 0.4303 on the challenge test set.

The transition from image to video understanding requires vision-language models (VLMs) to shift from recognizing static patterns to reasoning over temporal dynamics such as motion trajectories, speed changes, and state transitions. Yet current post-training methods fall short due to two critical limitations: (1) existing datasets often lack temporal-centricity, where answers can be inferred from isolated keyframes rather than requiring holistic temporal integration; and (2) training data generated by proprietary models contains systematic errors in fundamental temporal perception, such as confusing motion directions or misjudging speeds. We introduce SynRL, a post-training framework that teaches models temporal primitives, the fundamental building blocks of temporal understanding including direction, speed, and state tracking. Our key insight is that these abstract primitives, learned from programmatically generated synthetic videos, transfer effectively to real-world scenarios. We decompose temporal understanding into short-term perceptual primitives (speed, direction) and long-term cognitive primitives, constructing 7.7K CoT and 7K RL samples with ground-truth frame-level annotations through code-based video generation. Despite training on simple geometric shapes, SynRL achieves substantial improvements across 15 benchmarks spanning temporal grounding, complex reasoning, and general video understanding. Remarkably, our 7.7K synthetic CoT samples outperform Video-R1 with 165K real-world samples. We attribute this to fundamental temporal skills, such as tracking frame by frame changes and comparing velocity, that transfer effectively from abstract synthetic patterns to complex real-world scenarios. This establishes a new paradigm for video post-training: video temporal learning through carefully designed synthetic data provides a more cost efficient scaling path.

Multimodal large language models (MLLMs) are typically trained in multiple stages, with video-based supervised fine-tuning (Video-SFT) serving as a key step for improving visual understanding. Yet its effect on the fine-grained evolution of visual capabilities, particularly the balance between spatial and temporal understanding, remains poorly understood. In this paper, we systematically study how Video-SFT reshapes visual capabilities in MLLMs. Across architectures, parameter scales, and frame sampling settings, we observe a consistent pattern: Video-SFT reliably improves video performance, but often yields limited gains or even degradation on static image benchmarks. We further show that this trade-off is closely tied to temporal budget: increasing the number of sampled frames generally improves video performance, but does not reliably improve static image performance. Motivated by this finding, we study an instruction-aware Hybrid-Frame strategy that adaptively allocates frame counts and partially mitigates the image-video trade-off. Our results indicate that Video-SFT is not a free lunch for MLLMs, and preserving spatial understanding remains a central challenge in joint image-video training.

Recognizing and localizing student confusion from video is an important yet challenging problem in educational AI. Existing confusion datasets suffer from noisy labels, coarse temporal annotations, and limited expert validation, which hinder reliable fine-grained recognition and temporally grounded analysis. To address these limitations, we propose a practical multi-stage filtering pipeline that integrates two stages of model-assisted screening, researcher curation, and expert validation to build a higher-quality benchmark for confusion understanding. Based on this pipeline, we introduce ConfusionBench, a new benchmark for educational videos consisting of a balanced confusion recognition dataset and a video localization dataset. We further provide zero-shot baseline evaluations of a representative open-source model and a proprietary model on clip-level confusion recognition, long-video confusion localization tasks. Experimental results show that the proprietary model performs better overall but tends to over-predict transitional segments, while the open-source model is more conservative and more prone to missed detections. In addition, the proposed student confusion report visualization can support educational experts in making intervention decisions and adapting learning plans accordingly. All datasets and related materials will be made publicly available on our project page.

In the context of long-term video understanding with large multimodal models, many frameworks have been proposed. Although transformer-based visual compressors and memory-augmented approaches are often used to process long videos, they usually compress each frame independently and therefore fail to achieve strong performance on tasks that require understanding complete events, such as temporal ordering tasks in MLVU and VNBench. This motivates us to rethink the conventional one-way scheme from perception to memory, and instead establish a feedbackdriven process in which past visual contexts stored in the context memory can benefit ongoing perception. To this end, we propose Question-guided Visual Compression with Memory Feedback (QViC-MF), a framework for long-term video understanding. At its core is a Question-guided Multimodal Selective Attention (QMSA), which learns to preserve visual information related to the given question from both the current clip and the past related frames from the memory. The compressor and memory feedback work iteratively for each clip of the entire video. This simple yet effective design yields large performance gains on longterm video understanding tasks. Extensive experiments show that our method achieves significant improvement over current state-of-the-art methods by 6.1% on MLVU test, 8.3% on LVBench, 18.3% on VNBench Long, and 3.7% on VideoMME Long. The code will be released publicly.

Despite advances in the application of MLLMs for various video tasks, video event prediction (VEP) remains relatively underexplored. VEP requires the model to perform fine-grained temporal modeling of videos and establish logical relationships between videos and future events, which current MLLMs still struggle with. In this work, we first present a comprehensive evaluation of current leading MLLMs on the VEP task, revealing the reasons behind their inaccurate predictions, including lack of logical reasoning ability for future events prediction and insufficient utilization of visual information. To address these challenges, we propose Chain of Events (CoE) paradigm, which constructs temporal event chains to implicitly enforce MLLM focusing on the visual content and the logical connections between videos and future events, incentivizing model's reasoning capability with multiple training protocols. Experimental results on public benchmarks demonstrate that our method outperforms both leading open-source and commercial MLLMs, establishing a new state-of-the-art on the VEP task. Codes and models will be released soon.

Existing text-to-video retrieval benchmarks often permit matching from static appearance, making it difficult to determine whether models capture state changes and final outcomes. We introduce GenState-AI, a benchmark comprising 956 AI-generated videos and 195 verified state-change queries across Toy, Cartoon, Home, and Manipulation domains. Its construction uses prompt-controlled variants that target either conflicting temporal/state outcomes or semantic substitutions. Following human verification, all valid videos contribute to subset-level retrieval pools, while complete query-specific triplets support diagnostic hard-negative ordering. We evaluate VCF-Lik, Qwen3-VL, and LanguageBind using both standard Recall@K and triplet-based consistency metrics. Qwen3-VL (8B) achieves the best macro-average R@1 and triplet consistency, while LanguageBind obtains the highest semantic discrimination accuracy. However, stronger retrieval does not consistently yield correct hard-negative ordering, and Qwen3-VL and LanguageBind remain weak on Manipulation triplets, where VCF-Lik performs substantially better. GenState-AI therefore provides a controlled diagnostic testbed for distinguishing global retrieval quality from fine-grained temporal and semantic grounding in AI-generated videos.

We introduce the task of early mistake detection in video, where the goal is to determine whether a keystep in a procedural activity is performed correctly while observing as little of the streaming video as possible. To tackle this problem, we propose a method comprising a mistake detector and a reinforcement learning policy. At each timestep, the detector processes recently observed frames to estimate the keystep's correctness while anticipating future visual features, enabling reliable early mistake estimates. Meanwhile, the policy aggregates the detector outputs and visual observations over time and adaptively decides when to exit (i.e., stop processing incoming frames) while producing the final prediction. Using diverse real-world procedural video datasets, we demonstrate that our MistExit model achieves superior mistake detection accuracy while reducing the fraction of video observed compared to state-of-the-art models. Project: https://vision.cs.utexas.edu/projects/mist_exit.

Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\textbf{benchmarking}$, $\textbf{diagnosis}$, and $\textbf{injection}$. We curate $\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark--$\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at https://hf.co/datasets/fengyee/camera-motion-dataset-and-benchmark.

Humans inhabit a physical 4D world where geometric structure and semantic content evolve over time, constituting a dynamic 4D reality (spatial with temporal dimension). While current Multimodal Large Language Models (MLLMs) excel in static visual understanding, can they also be adept at "thinking in dynamics", i.e., perceive, track and reason about spatio-temporal dynamics in evolving scenes? To systematically assess their spatio-temporal reasoning and localized dynamics perception capabilities, we introduce Dyn-Bench, a large-scale benchmark built from diverse real-world and synthetic video datasets, enabling robust and scalable evaluation of spatio-temporal understanding. Through multi-stage filtering from massive 2D and 4D data sources, Dyn-Bench provides a high-quality collection of dynamic scenes, comprising 1k videos, 7k visual question answering (VQA) pairs, and 3k dynamic object grounding pairs. We probe general, spatial and region-level MLLMs to express how they think in dynamics both linguistically and visually, and find that existing models cannot simultaneously maintain strong performance in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent interpretations of motion and interaction. Notably, conventional prompting strategies (e.g., chain-of-thought or caption-based hints) provide limited improvement, whereas structured integration approaches, including Mask-Guided Fusion and Spatio-Temporal Textual Cognitive Map (ST-TCM), significantly enhance MLLMs' dynamics perception and spatio-temporal reasoning in the physical 4D world. Code and benchmark are available at https://dyn-bench.github.io/.

Multi-modal large language models (MLLMs) have advanced general-purpose video understanding but struggle with long, high-resolution videos -- they process every pixel equally in their vision transformers (ViTs) or LLMs despite significant spatiotemporal redundancy. We introduce AutoGaze, a lightweight module that removes redundant patches before processed by a ViT or an MLLM. Trained with next-token prediction and reinforcement learning, AutoGaze autoregressively selects a minimal set of multi-scale patches that can reconstruct the video within a user-specified error threshold, eliminating redundancy while preserving information. Empirically, AutoGaze reduces visual tokens by 4x-100x and accelerates ViTs and MLLMs by up to 19x, enabling scaling MLLMs to 1K-frame 4K-resolution videos and achieving superior results on video benchmarks (e.g., 67.0% on VideoMME). Furthermore, we introduce HLVid: the first high-resolution, long-form video QA benchmark with 5-minute 4K-resolution videos, where an MLLM scaled with AutoGaze improves over the baseline by 10.1% and outperforms the previous best MLLM by 4.5%. Project page: https://autogaze.github.io/.