We address the task of zero-shot fine-grained video classification, where no video examples or temporal annotations are available for unseen action classes. While contrastive vision-language models such as SigLIP demonstrate strong open-set recognition via mean-pooled image-text similarity, they fail to capture the temporal structure critical for distinguishing fine-grained activities. We introduce ActAlign, a zero-shot framework that formulates video classification as sequence alignment. For each class, a large language model generates an ordered sub-action sequence, which is aligned with video frames using Dynamic Time Warping (DTW) in a shared embedding space. Without any video-text supervision or fine-tuning, ActAlign achieves 30.5% accuracy on the extremely challenging ActionAtlas benchmark, where human accuracy is only 61.6%. ActAlign outperforms billion-parameter video-language models while using approximately 8x less parameters. These results demonstrate that structured language priors, combined with classical alignment techniques, offer a scalable and general approach to unlocking the open-set recognition potential of vision-language models for fine-grained video understanding.
Long-video understanding (LVU) remains a severe challenge for existing multimodal large language models (MLLMs), primarily due to the prohibitive computational cost. Recent approaches have explored KV compression to mitigate this issue, but they often suffer from significant information loss at high compression ratios. In this paper, we introduce Video-X^2L, which flexibly preserves critical video information for each LVU task. Video-X^2L involves two key operations. The first one is called bi-level KV compression. During the MLLM's pre-filling stage, Video-X^2L generates two types of compressed KVs: low-compression KVs (L-KVs) to capture fine-grained video details and high-compression KVs (H-KVs) to offer compact video representations. The second one is called selective KV re-loading. During the MLLM's decoding stage, Video-X^2L selectively re-loads L-KVs for the most critical video chunks while using H-KVs for other less important ones. This allows the MLLM to fully utilize task-specific information while maintaining the overall compactness. Video-X^2L is simple yet effective: it is free from additional training and directly compatible with existing KV-compressible MLLMs. We evaluate Video-X^2L with a variety of popular LVU benchmarks, including VideoMME, MLVU, LongVideoBench, and VNBench. Our experiment result shows that Video-X^2L outperforms existing KV-compression methods by a huge advantage while substantially saving the computation cost.
Video Large Language Models (VideoLLMs) have demonstrated remarkable understanding capabilities, but are found struggling to tackle multi-shot scenarios,e.g., video clips with varying camera angles or scene changes. This challenge can render failures such as instance identity forgetting and key frame negligence. In this work, we first attribute the challenge to the lack of multi-shot annotations among existing datasets and therefore we introduce a new dataset termed MultiClip-Bench, featuring dense descriptions and instruction-based question-answering pairs tailored for multi-shot scenarios. We empirically find that the training set significantly boosts the multi-shot performance, while the testing benchmark provides a reliable measure of the model capability in multi-shot scenarios. By further analyzing and discovering that current models only encode instance features in a discrete or lossy manner, at the risk of missing identity information, we then contribute a new model IPFormer-VideoLLM. Its key idea is the injection of instance-level features as instance prompts through an efficient attention-based connector. This allows for the aggregation of instance-specific information across scenes. Experiments demonstrate that our proposed dataset and model not only enhance the multi-scene video understanding significantly, but also offer distinct advantages across various video benchmarks.
Videos contain a wealth of information, and generating detailed and accurate descriptions in natural language is a key aspect of video understanding. In this paper, we present video-SALMONN 2, an advanced audio-visual large language model (LLM) with low-rank adaptation (LoRA) designed for enhanced video (with paired audio) captioning through directed preference optimisation (DPO). We propose new metrics to evaluate the completeness and accuracy of video descriptions, which are optimised using DPO. To further improve training, we propose a novel multi-round DPO (MrDPO) approach, which involves periodically updating the DPO reference model, merging and re-initialising the LoRA module as a proxy for parameter updates after each training round (1,000 steps), and incorporating guidance from ground-truth video captions to stabilise the process. Experimental results show that MrDPO significantly enhances video-SALMONN 2's captioning accuracy, reducing the captioning error rates by 28\%. The final video-SALMONN 2 model, with just 7 billion parameters, surpasses leading models such as GPT-4o and Gemini-1.5-Pro in video captioning tasks, while maintaining highly competitive performance to the state-of-the-art on widely used video question-answering benchmarks among models of similar size. Codes are available at \href{https://github.com/bytedance/video-SALMONN-2}{https://github.com/bytedance/video-SALMONN-2}.
Vision-Language Models (VLMs) have demonstrated strong performance in video-language tasks, yet their application to long video understanding remains constrained by the quadratic complexity of standard attention mechanisms. In this paper, we propose \textbf{PEVLM}, a parallel encoding strategy specifically designed to improve the prefill efficiency of VLMs without requiring model finetuning. PEVLM partitions the input into block-wise segments with a shared sink, preserves full-attention positional embeddings, and aligns attention weights to mimic full-attention distributions. This design reduces attention computation from $O((T \times N)^2)$ to $O(T \times N)$ while maintaining high accuracy. Extensive experiments on the LongVideoBench benchmark show that PEVLM achieves up to 8.37\% accuracy improvement over existing inference-efficient methods and delivers up to 7.47x speedup in attention computation and 40\% reduction in end-to-end latency. Under strict latency constraints, PEVLM significantly outperforms baselines, raising accuracy from 23.26\% to 61.03\%. These results highlight PEVLM's effectiveness for low-latency, long-context video understanding, making it well-suited for real-world applications such as autonomous driving.
Multimodal Large Language Models (MLLMs) struggle with long videos due to fixed context windows and weak long-term dependency modeling. Existing Retrieval-Augmented Generation (RAG) methods for videos use static retrieval strategies, leading to inefficiencies for simple queries and information loss for complex tasks. To address this, we propose AdaVideoRAG, a novel framework that dynamically adapts retrieval granularity based on query complexity using a lightweight intent classifier. Our framework employs an Omni-Knowledge Indexing module to build hierarchical databases from text (captions, ASR, OCR), visual features, and semantic graphs, enabling optimal resource allocation across tasks. We also introduce the HiVU benchmark for comprehensive evaluation. Experiments demonstrate improved efficiency and accuracy for long-video understanding, with seamless integration into existing MLLMs. AdaVideoRAG establishes a new paradigm for adaptive retrieval in video analysis. Codes will be open-sourced at https://github.com/xzc-zju/AdaVideoRAG.
We propose an efficient framework to compress multiple video-frame features before feeding them into large multimodal models, thereby mitigating the severe token explosion arising from long or dense videos. Our design leverages a bidirectional state-space-based block equipped with a gated skip connection and a learnable weighted-average pooling mechanism applied to periodically inserted learned queries. This structure enables hierarchical downsampling across both spatial and temporal dimensions, preserving performance in a cost-effective manner. Across challenging long and dense video understanding tasks, our approach demonstrates competitive results against state-of-the-art models, while significantly reducing overall token budget. Notably, replacing our proposed state-space block with a conventional Transformer results in substantial performance degradation, highlighting the advantages of state-space modeling for effectively compressing multi-frame video data. Our framework emphasizes resource-conscious efficiency, making it practical for real-world deployments. We validate its scalability and generality across multiple benchmarks, achieving the dual objectives of efficient resource usage and comprehensive video understanding.
Modern multimodal large language models (MLLMs) can reason over hour-long video, yet their key-value (KV) cache grows linearly with time--quickly exceeding the fixed memory of phones, AR glasses, and edge robots. Prior compression schemes either assume the whole video and user query are available offline or must first build the full cache, so memory still scales with stream length. InfiniPot-V is the first training-free, query-agnostic framework that enforces a hard, length-independent memory cap for streaming video understanding. During video encoding it monitors the cache and, once a user-set threshold is reached, runs a lightweight compression pass that (i) removes temporally redundant tokens via Temporal-axis Redundancy (TaR) metric and (ii) keeps semantically significant tokens via Value-Norm (VaN) ranking. Across four open-source MLLMs and four long-video and two streaming-video benchmarks, InfiniPot-V cuts peak GPU memory by up to 94%, sustains real-time generation, and matches or surpasses full-cache accuracy--even in multi-turn dialogues. By dissolving the KV cache bottleneck without retraining or query knowledge, InfiniPot-V closes the gap for on-device streaming video assistants.
Video temporal understanding is crucial for multimodal large language models (MLLMs) to reason over events in videos. Despite recent advances in general video understanding, current MLLMs still struggle with fine-grained temporal reasoning. While reinforcement learning (RL) has been explored to address this issue recently, existing RL approaches remain limited in effectiveness. In this work, we propose MUSEG, a novel RL-based method that enhances temporal understanding by introducing timestamp-aware multi-segment grounding. MUSEG enables MLLMs to align queries with multiple relevant video segments, promoting more comprehensive temporal reasoning. To facilitate effective learning, we design a customized RL training recipe with phased rewards that progressively guides the model toward temporally grounded reasoning. Extensive experiments on temporal grounding and time-sensitive video QA tasks demonstrate that MUSEG significantly outperforms existing methods and generalizes well across diverse temporal understanding scenarios. View our project at https://github.com/THUNLP-MT/MUSEG.
Egocentric video-language understanding demands both high efficiency and accurate spatial-temporal modeling. Existing approaches face three key challenges: 1) Excessive pre-training cost arising from multi-stage pre-training pipelines, 2) Ineffective spatial-temporal encoding due to manually split 3D rotary positional embeddings that hinder feature interactions, and 3) Imprecise learning objectives in soft-label multi-instance retrieval, which neglect negative pair correlations. In this paper, we introduce EVA02-AT, a suite of EVA02-based video-language foundation models tailored to egocentric video understanding tasks. EVA02-AT first efficiently transfers an image-based CLIP model into a unified video encoder via a single-stage pretraining. Second, instead of applying rotary positional embeddings to isolated dimensions, we introduce spatial-temporal rotary positional embeddings along with joint attention, which can effectively encode both spatial and temporal information on the entire hidden dimension. This joint encoding of spatial-temporal features enables the model to learn cross-axis relationships, which are crucial for accurately modeling motion and interaction in videos. Third, focusing on multi-instance video-language retrieval tasks, we introduce the Symmetric Multi-Similarity (SMS) loss and a novel training framework that advances all soft labels for both positive and negative pairs, providing a more precise learning objective. Extensive experiments on Ego4D, EPIC-Kitchens-100, and Charades-Ego under zero-shot and fine-tuning settings demonstrate that EVA02-AT achieves state-of-the-art performance across diverse egocentric video-language tasks with fewer parameters. Models with our SMS loss also show significant performance gains on multi-instance retrieval benchmarks. Our code and models are publicly available at https://github.com/xqwang14/EVA02-AT .
Self-supervised learning (SSL) has achieved major advances in natural images and video understanding, but challenges remain in domains like echocardiography (heart ultrasound) due to subtle anatomical structures, complex temporal dynamics, and the current lack of domain-specific pre-trained models. Existing SSL approaches such as contrastive, masked modeling, and clustering-based methods struggle with high intersample similarity, sensitivity to low PSNR inputs common in ultrasound, or aggressive augmentations that distort clinically relevant features. We present DISCOVR (Distilled Image Supervision for Cross Modal Video Representation), a self-supervised dual branch framework for cardiac ultrasound video representation learning. DISCOVR combines a clustering-based video encoder that models temporal dynamics with an online image encoder that extracts fine-grained spatial semantics. These branches are connected through a semantic cluster distillation loss that transfers anatomical knowledge from the evolving image encoder to the video encoder, enabling temporally coherent representations enriched with fine-grained semantic understanding. Evaluated on six echocardiography datasets spanning fetal, pediatric, and adult populations, DISCOVR outperforms both specialized video anomaly detection methods and state-of-the-art video-SSL baselines in zero-shot and linear probing setups, and achieves superior segmentation transfer.
Audiovisual segmentation (AVS) aims to identify visual regions corresponding to sound sources, playing a vital role in video understanding, surveillance, and human-computer interaction. Traditional AVS methods depend on large-scale pixel-level annotations, which are costly and time-consuming to obtain. To address this, we propose a novel zero-shot AVS framework that eliminates task-specific training by leveraging multiple pretrained models. Our approach integrates audio, vision, and text representations to bridge modality gaps, enabling precise sound source segmentation without AVS-specific annotations. We systematically explore different strategies for connecting pretrained models and evaluate their efficacy across multiple datasets. Experimental results demonstrate that our framework achieves state-of-the-art zero-shot AVS performance, highlighting the effectiveness of multimodal model integration for finegrained audiovisual segmentation.
Action segmentation is a core challenge in high-level video understanding, aiming to partition untrimmed videos into segments and assign each a label from a predefined action set. Existing methods primarily address single-person activities with fixed action sequences, overlooking multi-person scenarios. In this work, we pioneer textual reference-guided human action segmentation in multi-person settings, where a textual description specifies the target person for segmentation. We introduce the first dataset for Referring Human Action Segmentation, i.e., RHAS133, built from 133 movies and annotated with 137 fine-grained actions with 33h video data, together with textual descriptions for this new task. Benchmarking existing action recognition methods on RHAS133 using VLM-based feature extractors reveals limited performance and poor aggregation of visual cues for the target person. To address this, we propose a holistic-partial aware Fourier-conditioned diffusion framework, i.e., HopaDIFF, leveraging a novel cross-input gate attentional xLSTM to enhance holistic-partial long-range reasoning and a novel Fourier condition to introduce more fine-grained control to improve the action segmentation generation. HopaDIFF achieves state-of-the-art results on RHAS133 in diverse evaluation settings. The code is available at https://github.com/KPeng9510/HopaDIFF.git.
Recent advances in Video Large Language Models (VLLMs) have significantly enhanced their ability to understand video content. Nonetheless, processing long videos remains challenging due to high computational demands and the redundancy present in the visual data. In this work, we propose METok, a training-free, Multi-stage Event-based Token compression framework designed to accelerate VLLMs' inference while preserving accuracy. METok progressively eliminates redundant visual tokens across three critical stages: (1) event-aware compression during vision encoding, (2) hierarchical token pruning in the prefilling stage based on semantic alignment and event importance, and (3) a decoding-stage KV Cache optimization that further reduces memory consumption. Our experiments on diverse video benchmarks demonstrate that METok achieves an optimal trade-off between efficiency and accuracy by dynamically selecting informative visual tokens. For instance, equipping LongVA-7B with METok realizes an 80.6% FLOPs reduction and 93.5% KV Cache memory savings, all while maintaining comparable or even superior accuracy.
Long video understanding (LVU) presents a significant challenge for current multi-modal large language models (MLLMs) due to the task's inherent complexity and context window constraint. It is widely assumed that addressing LVU tasks requires foundation MLLMs with extended context windows, strong visual perception capabilities, and proficient domain expertise. In this work, we challenge this common belief by introducing VideoDeepResearch, a novel agentic framework for long video understanding. Our approach relies solely on a text-only large reasoning model (LRM) combined with a modular multi-modal toolkit, including multimodal retrievers and visual perceivers, all of which are readily available in practice. For each LVU task, the system formulates a problem-solving strategy through reasoning, while selectively accessing and utilizing essential video content via tool using. We conduct extensive experiments on popular LVU benchmarks, including MLVU, Video-MME, and LVBench. Our results demonstrate that VideoDeepResearch achieves substantial improvements over existing MLLM baselines, surpassing the previous state-of-the-art by 9.6%, 6.6%, and 3.9% on MLVU (test), LVBench, and LongVideoBench, respectively. These findings highlight the promise of agentic systems in overcoming key challenges in LVU problems.
Current Multimodal Large Language Models (MLLMs) may struggle with understanding long or complex videos due to computational demands at test time, lack of robustness, and limited accuracy, primarily stemming from their feed-forward processing nature. These limitations could be more severe for models with fewer parameters. To address these limitations, we propose a novel framework inspired by cybernetic principles, redesigning video MLLMs as adaptive systems capable of self-monitoring, self-correction, and dynamic resource allocation during inference. Our approach, CyberV, introduces a cybernetic loop consisting of an MLLM Inference System, a Sensor, and a Controller. Specifically, the sensor monitors forward processes of the MLLM and collects intermediate interpretations, such as attention drift, then the controller determines when and how to trigger self-correction and generate feedback to guide the next round. This test-time adaptive scaling framework enhances frozen MLLMs without requiring retraining or additional components. Experiments demonstrate significant improvements: CyberV boosts Qwen2.5-VL-7B by 8.3% and InternVL3-8B by 5.5% on VideoMMMU, surpassing the competitive proprietary model GPT-4o. When applied to Qwen2.5-VL-72B, it yields a 10.0% improvement, achieving performance even comparable to human experts. Furthermore, our method demonstrates consistent gains on general-purpose benchmarks, such as VideoMME and WorldSense, highlighting its effectiveness and generalization capabilities in making MLLMs more robust and accurate for dynamic video understanding. The code is released at https://github.com/marinero4972/CyberV.
Surgical video understanding is pivotal for enabling automated intraoperative decision-making, skill assessment, and postoperative quality improvement. However, progress in developing surgical video foundation models (FMs) remains hindered by the scarcity of large-scale, diverse datasets for pretraining and systematic evaluation. In this paper, we introduce \textbf{SurgBench}, a unified surgical video benchmarking framework comprising a pretraining dataset, \textbf{SurgBench-P}, and an evaluation benchmark, \textbf{SurgBench-E}. SurgBench offers extensive coverage of diverse surgical scenarios, with SurgBench-P encompassing 53 million frames across 22 surgical procedures and 11 specialties, and SurgBench-E providing robust evaluation across six categories (phase classification, camera motion, tool recognition, disease diagnosis, action classification, and organ detection) spanning 72 fine-grained tasks. Extensive experiments reveal that existing video FMs struggle to generalize across varied surgical video analysis tasks, whereas pretraining on SurgBench-P yields substantial performance improvements and superior cross-domain generalization to unseen procedures and modalities. Our dataset and code are available upon request.
Video understanding has been considered as one critical step towards world modeling, which is an important long-term problem in AI research. Recently, multi-modal foundation models have shown such potential via large-scale pretraining. However, these models simply align encoders of different modalities via contrastive learning, while lacking deeper multi-modal interactions, which is critical for understanding complex target movements with diversified video scenes. To fill this gap, we propose a unified Super Encoding Network (SEN) for video understanding, which builds up such distinct interactions through recursive association of multi-modal encoders in the foundation models. Specifically, we creatively treat those well-trained encoders as "super neurons" in our SEN. Via designing a Recursive Association (RA) block, we progressively fuse multi-modalities with the input video, based on knowledge integrating, distributing, and prompting of super neurons in a recursive manner. In this way, our SEN can effectively encode deeper multi-modal interactions, for prompting various video understanding tasks in downstream. Extensive experiments show that, our SEN can remarkably boost the four most representative video tasks, including tracking, recognition, chatting, and editing, e.g., for pixel-level tracking, the average jaccard index improves 2.7%, temporal coherence(TC) drops 8.8% compared to the popular CaDeX++ approach. For one-shot video editing, textual alignment improves 6.4%, and frame consistency increases 4.1% compared to the popular TuneA-Video approach.
Perceiving the world from both egocentric (first-person) and exocentric (third-person) perspectives is fundamental to human cognition, enabling rich and complementary understanding of dynamic environments. In recent years, allowing the machines to leverage the synergistic potential of these dual perspectives has emerged as a compelling research direction in video understanding. In this survey, we provide a comprehensive review of video understanding from both exocentric and egocentric viewpoints. We begin by highlighting the practical applications of integrating egocentric and exocentric techniques, envisioning their potential collaboration across domains. We then identify key research tasks to realize these applications. Next, we systematically organize and review recent advancements into three main research directions: (1) leveraging egocentric data to enhance exocentric understanding, (2) utilizing exocentric data to improve egocentric analysis, and (3) joint learning frameworks that unify both perspectives. For each direction, we analyze a diverse set of tasks and relevant works. Additionally, we discuss benchmark datasets that support research in both perspectives, evaluating their scope, diversity, and applicability. Finally, we discuss limitations in current works and propose promising future research directions. By synthesizing insights from both perspectives, our goal is to inspire advancements in video understanding and artificial intelligence, bringing machines closer to perceiving the world in a human-like manner. A GitHub repo of related works can be found at https://github.com/ayiyayi/Awesome-Egocentric-and-Exocentric-Vision.
Despite recent advances in retrieval-augmented generation (RAG) for video understanding, effectively understanding long-form video content remains underexplored due to the vast scale and high complexity of video data. Current RAG approaches typically segment videos into fixed-length chunks, which often disrupts the continuity of contextual information and fails to capture authentic scene boundaries. Inspired by the human ability to naturally organize continuous experiences into coherent scenes, we present SceneRAG, a unified framework that leverages large language models to segment videos into narrative-consistent scenes by processing ASR transcripts alongside temporal metadata. SceneRAG further sharpens these initial boundaries through lightweight heuristics and iterative correction. For each scene, the framework fuses information from both visual and textual modalities to extract entity relations and dynamically builds a knowledge graph, enabling robust multi-hop retrieval and generation that account for long-range dependencies. Experiments on the LongerVideos benchmark, featuring over 134 hours of diverse content, confirm that SceneRAG substantially outperforms prior baselines, achieving a win rate of up to 72.5 percent on generation tasks.