Video understanding has shown remarkable improvements in recent years, largely dependent on the availability of large scaled labeled datasets. Recent advancements in visual-language models, especially based on contrastive pretraining, have shown remarkable generalization in zero-shot tasks, helping to overcome this dependence on labeled datasets. Adaptations of such models for videos, typically involve modifying the architecture of vision-language models to cater to video data. However, this is not trivial, since such adaptations are mostly computationally intensive and struggle with temporal modeling. We present TP-CLIP, an adaptation of CLIP that leverages temporal visual prompting for temporal adaptation without modifying the core CLIP architecture. This preserves its generalization abilities. TP-CLIP efficiently integrates into the CLIP architecture, leveraging its pre-trained capabilities for video data. Extensive experiments across various datasets demonstrate its efficacy in zero-shot and few-shot learning, outperforming existing approaches with fewer parameters and computational efficiency. In particular, we use just 1/3 the GFLOPs and 1/28 the number of tuneable parameters in comparison to recent state-of-the-art and still outperform it by up to 15.8% depending on the task and dataset.
Video spatial reasoning, which involves inferring the underlying spatial structure from observed video frames, poses a significant challenge for existing Multimodal Large Language Models (MLLMs). This limitation stems primarily from 1) the absence of high-quality datasets for this task, and 2) the lack of effective training strategies to develop spatial reasoning capabilities. Motivated by the success of Reinforcement Learning with Verifiable Reward (RLVR) in unlocking LLM reasoning abilities, this work aims to improve MLLMs in video spatial reasoning through the RLVR paradigm. To this end, we introduce the $\textbf{SpaceR}$ framework. First, we present $\textbf{SpaceR-151k}$, a dataset with 91k questions spanning diverse spatial reasoning scenarios with verifiable answers, and 60k samples for maintaining general multimodal understanding. Second, we propose $\textbf{Spatially-Guided RLVR (SG-RLVR)}$, a novel reinforcement learning approach that extends Group Relative Policy Optimization (GRPO) with a novel map imagination mechanism, which encourages the model to infer spatial layouts in the thinking process, thereby facilitating more effective spatial reasoning. Extensive experiments demonstrate that SpaceR achieves state-of-the-art performance on spatial reasoning benchmarks (e.g., VSI-Bench, STI-Bench, and SPAR-Bench), while maintaining competitive results on video understanding benchmarks (e.g., Video-MME, TempCompass, and LongVideoBench). Remarkably, SpaceR surpasses the advanced GPT-4o by 11.6\% accuracy on VSI-Bench and is on par with the leading proprietary model Gemini-2.0-Flash, highlighting the effectiveness of our SpaceR-151k dataset and SG-RLVR in reinforcing spatial reasoning ability of MLLMs. Code, model, and dataset are available at https://github.com/OuyangKun10/SpaceR.
Large video-language models (LVLMs) have shown remarkable performance across various video-language tasks. However, they encounter significant challenges when processing long videos because of the large number of video frames involved. Downsampling long videos in either space or time can lead to visual hallucinations, making it difficult to accurately interpret long videos. Motivated by human hierarchical temporal search strategies, we propose \textbf{TimeSearch}, a novel framework enabling LVLMs to understand long videos in a human-like manner. TimeSearch integrates two human-like primitives into a unified autoregressive LVLM: 1) \textbf{Spotlight} efficiently identifies relevant temporal events through a Temporal-Augmented Frame Representation (TAFR), explicitly binding visual features with timestamps; 2) \textbf{Reflection} evaluates the correctness of the identified events, leveraging the inherent temporal self-reflection capabilities of LVLMs. TimeSearch progressively explores key events and prioritizes temporal search based on reflection confidence. Extensive experiments on challenging long-video benchmarks confirm that TimeSearch substantially surpasses previous state-of-the-art, improving the accuracy from 41.8\% to 51.5\% on the LVBench. Additionally, experiments on temporal grounding demonstrate that appropriate TAFR is adequate to effectively stimulate the surprising temporal grounding ability of LVLMs in a simpler yet versatile manner, which improves mIoU on Charades-STA by 11.8\%. The code will be released.
Balancing temporal resolution and spatial detail under limited compute budget remains a key challenge for video-based multi-modal large language models (MLLMs). Existing methods typically compress video representations using predefined rules before feeding them into the LLM, resulting in irreversible information loss and often ignoring input instructions. To address this, we propose a novel slow-fast architecture that naturally circumvents this trade-off, enabling the use of more input frames while preserving spatial details. Inspired by how humans first skim a video before focusing on relevant parts, our slow-fast design employs a dual-token strategy: 1) "fast" visual tokens -- a compact set of compressed video features -- are fed into the LLM alongside text embeddings to provide a quick overview; 2) "slow" visual tokens -- uncompressed video features -- are cross-attended by text embeddings through specially designed hybrid decoder layers, enabling instruction-aware extraction of relevant visual details with linear complexity. We conduct systematic exploration to optimize both the overall architecture and key components. Experiments show that our model significantly outperforms self-attention-only baselines, extending the input capacity from 16 to 128 frames with just a 3% increase in computation, and achieving a 16% average performance improvement across five video understanding benchmarks. Our 7B model achieves state-of-the-art performance among models of similar size. Furthermore, our slow-fast architecture is a plug-and-play design that can be integrated into other video MLLMs to improve efficiency and scalability.
We propose Cross-Attention in Audio, Space, and Time (CA^2ST), a transformer-based method for holistic video recognition. Recognizing actions in videos requires both spatial and temporal understanding, yet most existing models lack a balanced spatio-temporal understanding of videos. To address this, we propose a novel two-stream architecture, called Cross-Attention in Space and Time (CAST), using only RGB input. In each layer of CAST, Bottleneck Cross-Attention (B-CA) enables spatial and temporal experts to exchange information and make synergistic predictions. For holistic video understanding, we extend CAST by integrating an audio expert, forming Cross-Attention in Visual and Audio (CAVA). We validate the CAST on benchmarks with different characteristics, EPIC-KITCHENS-100, Something-Something-V2, and Kinetics-400, consistently showing balanced performance. We also validate the CAVA on audio-visual action recognition benchmarks, including UCF-101, VGG-Sound, KineticsSound, and EPIC-SOUNDS. With a favorable performance of CAVA across these datasets, we demonstrate the effective information exchange among multiple experts within the B-CA module. In summary, CA^2ST combines CAST and CAVA by employing spatial, temporal, and audio experts through cross-attention, achieving balanced and holistic video understanding.
We introduce EgoLife, a project to develop an egocentric life assistant that accompanies and enhances personal efficiency through AI-powered wearable glasses. To lay the foundation for this assistant, we conducted a comprehensive data collection study where six participants lived together for one week, continuously recording their daily activities - including discussions, shopping, cooking, socializing, and entertainment - using AI glasses for multimodal egocentric video capture, along with synchronized third-person-view video references. This effort resulted in the EgoLife Dataset, a comprehensive 300-hour egocentric, interpersonal, multiview, and multimodal daily life dataset with intensive annotation. Leveraging this dataset, we introduce EgoLifeQA, a suite of long-context, life-oriented question-answering tasks designed to provide meaningful assistance in daily life by addressing practical questions such as recalling past relevant events, monitoring health habits, and offering personalized recommendations. To address the key technical challenges of (1) developing robust visual-audio models for egocentric data, (2) enabling identity recognition, and (3) facilitating long-context question answering over extensive temporal information, we introduce EgoButler, an integrated system comprising EgoGPT and EgoRAG. EgoGPT is an omni-modal model trained on egocentric datasets, achieving state-of-the-art performance on egocentric video understanding. EgoRAG is a retrieval-based component that supports answering ultra-long-context questions. Our experimental studies verify their working mechanisms and reveal critical factors and bottlenecks, guiding future improvements. By releasing our datasets, models, and benchmarks, we aim to stimulate further research in egocentric AI assistants.
The rapid advancement of multi-modal language models (MLLMs) like GPT-4o has propelled the development of Omni language models, designed to process and proactively respond to continuous streams of multi-modal data. Despite their potential, evaluating their real-world interactive capabilities in streaming video contexts remains a formidable challenge. In this work, we introduce OmniMMI, a comprehensive multi-modal interaction benchmark tailored for OmniLLMs in streaming video contexts. OmniMMI encompasses over 1,121 videos and 2,290 questions, addressing two critical yet underexplored challenges in existing video benchmarks: streaming video understanding and proactive reasoning, across six distinct subtasks. Moreover, we propose a novel framework, Multi-modal Multiplexing Modeling (M4), designed to enable an inference-efficient streaming model that can see, listen while generating.
Audio Description is a narrated commentary designed to aid vision-impaired audiences in perceiving key visual elements in a video. While short-form video understanding has advanced rapidly, a solution for maintaining coherent long-term visual storytelling remains unresolved. Existing methods rely solely on frame-level embeddings, effectively describing object-based content but lacking contextual information across scenes. We introduce DANTE-AD, an enhanced video description model leveraging a dual-vision Transformer-based architecture to address this gap. DANTE-AD sequentially fuses both frame and scene level embeddings to improve long-term contextual understanding. We propose a novel, state-of-the-art method for sequential cross-attention to achieve contextual grounding for fine-grained audio description generation. Evaluated on a broad range of key scenes from well-known movie clips, DANTE-AD outperforms existing methods across traditional NLP metrics and LLM-based evaluations.
With the rapid development of multimodal models, the demand for assessing video understanding capabilities has been steadily increasing. However, existing benchmarks for evaluating video understanding exhibit significant limitations in coverage, task diversity, and scene adaptability. These shortcomings hinder the accurate assessment of models' comprehensive video understanding capabilities. To tackle this challenge, we propose a hierarchical and holistic video understanding (H2VU) benchmark designed to evaluate both general video and online streaming video comprehension. This benchmark contributes three key features: Extended video duration: Spanning videos from brief 3-second clips to comprehensive 1.5-hour recordings, thereby bridging the temporal gaps found in current benchmarks. Comprehensive assessment tasks: Beyond traditional perceptual and reasoning tasks, we have introduced modules for countercommonsense comprehension and trajectory state tracking. These additions test the models' deep understanding capabilities beyond mere prior knowledge. Enriched video data: To keep pace with the rapid evolution of current AI agents, we have expanded first-person streaming video datasets. This expansion allows for the exploration of multimodal models' performance in understanding streaming videos from a first-person perspective. Extensive results from H2VU reveal that existing multimodal large language models (MLLMs) possess substantial potential for improvement in our newly proposed evaluation tasks. We expect that H2VU will facilitate advancements in video understanding research by offering a comprehensive and in-depth analysis of MLLMs.
Recent advancements in Chain of Thought (COT) generation have significantly improved the reasoning capabilities of Large Language Models (LLMs), with reinforcement learning (RL) emerging as an effective post-training approach. Multimodal Large Language Models (MLLMs) inherit this reasoning potential but remain underexplored in tasks requiring both perception and logical reasoning. To address this, we introduce SEED-Bench-R1, a benchmark designed to systematically evaluate post-training methods for MLLMs in video understanding. It includes intricate real-world videos and complex everyday planning tasks in the format of multiple-choice questions, requiring sophisticated perception and reasoning. SEED-Bench-R1 assesses generalization through a three-level hierarchy: in-distribution, cross-environment, and cross-environment-task scenarios, equipped with a large-scale training dataset with easily verifiable ground-truth answers. Using Qwen2-VL-Instruct-7B as a base model, we compare RL with supervised fine-tuning (SFT), demonstrating RL's data efficiency and superior performance on both in-distribution and out-of-distribution tasks, even outperforming SFT on general video understanding benchmarks like LongVideoBench. Our detailed analysis reveals that RL enhances visual perception but often produces less logically coherent reasoning chains. We identify key limitations such as inconsistent reasoning and overlooked visual cues, and suggest future improvements in base model reasoning, reward modeling, and RL robustness against noisy signals.
Large video-language models (VLMs) have demonstrated promising progress in various video understanding tasks. However, their effectiveness in long-form video analysis is constrained by limited context windows. Traditional approaches, such as uniform frame sampling, often inevitably allocate resources to irrelevant content, diminishing their effectiveness in real-world scenarios. In this paper, we introduce BOLT, a method to BOost Large VLMs without additional Training through a comprehensive study of frame selection strategies. First, to enable a more realistic evaluation of VLMs in long-form video understanding, we propose a multi-source retrieval evaluation setting. Our findings reveal that uniform sampling performs poorly in noisy contexts, underscoring the importance of selecting the right frames. Second, we explore several frame selection strategies based on query-frame similarity and analyze their effectiveness at inference time. Our results show that inverse transform sampling yields the most significant performance improvement, increasing accuracy on the Video-MME benchmark from 53.8% to 56.1% and MLVU benchmark from 58.9% to 63.4%. Our code is available at https://github.com/sming256/BOLT.
Video understanding models often struggle with high computational requirements, extensive parameter counts, and slow inference speed, making them inefficient for practical use. To tackle these challenges, we propose Mobile-VideoGPT, an efficient multimodal framework designed to operate with fewer than a billion parameters. Unlike traditional video large multimodal models (LMMs), Mobile-VideoGPT consists of lightweight dual visual encoders, efficient projectors, and a small language model (SLM), enabling real-time throughput. To further improve efficiency, we present an Attention-Based Frame Scoring mechanism to select the key-frames, along with an efficient token projector that prunes redundant visual tokens and preserves essential contextual cues. We evaluate our model across well-established six video understanding benchmarks (e.g., MVBench, EgoSchema, NextQA, and PercepTest). Our results show that Mobile-VideoGPT-0.5B can generate up to 46 tokens per second while outperforming existing state-of-the-art 0.5B-parameter models by 6 points on average with 40% fewer parameters and more than 2x higher throughput. Our code and models are publicly available at: https://github.com/Amshaker/Mobile-VideoGPT.
View-invariant representation learning from egocentric (first-person, ego) and exocentric (third-person, exo) videos is a promising approach toward generalizing video understanding systems across multiple viewpoints. However, this area has been underexplored due to the substantial differences in perspective, motion patterns, and context between ego and exo views. In this paper, we propose a novel masked ego-exo modeling that promotes both causal temporal dynamics and cross-view alignment, called Bootstrap Your Own Views (BYOV), for fine-grained view-invariant video representation learning from unpaired ego-exo videos. We highlight the importance of capturing the compositional nature of human actions as a basis for robust cross-view understanding. Specifically, self-view masking and cross-view masking predictions are designed to learn view-invariant and powerful representations concurrently. Experimental results demonstrate that our BYOV significantly surpasses existing approaches with notable gains across all metrics in four downstream ego-exo video tasks. The code is available at https://github.com/park-jungin/byov.
The rise of Large Vision-Language Models (LVLMs) has significantly advanced video understanding. However, efficiently processing long videos remains a challenge due to the ``Sampling Dilemma'': low-density sampling risks missing critical information, while high-density sampling introduces redundancy. To address this issue, we introduce LSDBench, the first benchmark designed to evaluate LVLMs on long-video tasks by constructing high Necessary Sampling Density (NSD) questions, where NSD represents the minimum sampling density required to accurately answer a given question. LSDBench focuses on dense, short-duration actions to rigorously assess the sampling strategies employed by LVLMs. To tackle the challenges posed by high-NSD questions, we propose a novel Reasoning-Driven Hierarchical Sampling (RHS) framework, which combines global localization of question-relevant cues with local dense sampling for precise inference. Additionally, we develop a lightweight Semantic-Guided Frame Selector to prioritize informative frames, enabling RHS to achieve comparable or superior performance with significantly fewer sampled frames. Together, our LSDBench and RHS framework address the unique challenges of high-NSD long-video tasks, setting a new standard for evaluating and improving LVLMs in this domain. Our benchmark and evaluation codes has been released at: https://github.com/dvlab-research/LSDBench
Multi-modal Large language models (MLLMs) show remarkable ability in video understanding. Nevertheless, understanding long videos remains challenging as the models can only process a finite number of frames in a single inference, potentially omitting crucial visual information. To address the challenge, we propose generating multiple predictions through visual context sampling, followed by a scoring mechanism to select the final prediction. Specifically, we devise a bin-wise sampling strategy that enables MLLMs to generate diverse answers based on various combinations of keyframes, thereby enriching the visual context. To determine the final prediction from the sampled answers, we employ a self-reward by linearly combining three scores: (1) a frequency score indicating the prevalence of each option, (2) a marginal confidence score reflecting the inter-intra sample certainty of MLLM predictions, and (3) a reasoning score for different question types, including clue-guided answering for global questions and temporal self-refocusing for local questions. The frequency score ensures robustness through majority correctness, the confidence-aligned score reflects prediction certainty, and the typed-reasoning score addresses cases with sparse key visual information using tailored strategies. Experiments show that this approach covers the correct answer for a high percentage of long video questions, on seven datasets show that our method improves the performance of three MLLMs.
Large Vision-Language Models (LVLMs) demonstrate remarkable performance in short-video tasks such as video question answering, but struggle in long-video understanding. The linear frame sampling strategy, conventionally used by LVLMs, fails to account for the non-linear distribution of key events in video data, often introducing redundant or irrelevant information in longer contexts while risking the omission of critical events in shorter ones. To address this, we propose SelfReS, a non-linear spatiotemporal self-reflective sampling method that dynamically selects key video fragments based on user prompts. Unlike prior approaches, SelfReS leverages the inherently sparse attention maps of LVLMs to define reflection tokens, enabling relevance-aware token selection without requiring additional training or external modules. Experiments demonstrate that SelfReS can be seamlessly integrated into strong base LVLMs, improving long-video task accuracy and achieving up to 46% faster inference speed within the same GPU memory budget.
Pre-trained video large language models (Video LLMs) exhibit remarkable reasoning capabilities, yet adapting these models to new tasks involving additional modalities or data types (e.g., audio or 3D information) remains challenging. In this paper, we present PAVE, a flexible framework for adapting pre-trained Video LLMs to downstream tasks with side-channel signals, such as audio, 3D cues, or multi-view videos. PAVE introduces lightweight adapters, referred to as "patches," which add a small number of parameters and operations to a base model without modifying its architecture or pre-trained weights. In doing so, PAVE can effectively adapt the pre-trained base model to support diverse downstream tasks, including audio-visual question answering, 3D reasoning, multi-view video recognition, and high frame rate video understanding. Across these tasks, PAVE significantly enhances the performance of the base model, surpassing state-of-the-art task-specific models while incurring a minor cost of ~0.1% additional FLOPs and parameters. Further, PAVE supports multi-task learning and generalizes well across different Video LLMs. Our code is available at https://github.com/dragonlzm/PAVE.
The hallucination of large multimodal models (LMMs), providing responses that appear correct but are actually incorrect, limits their reliability and applicability. This paper aims to study the hallucination problem of LMMs in video modality, which is dynamic and more challenging compared to static modalities like images and text. From this motivation, we first present a comprehensive benchmark termed HAVEN for evaluating hallucinations of LMMs in video understanding tasks. It is built upon three dimensions, i.e., hallucination causes, hallucination aspects, and question formats, resulting in 6K questions. Then, we quantitatively study 7 influential factors on hallucinations, e.g., duration time of videos, model sizes, and model reasoning, via experiments of 16 LMMs on the presented benchmark. In addition, inspired by recent thinking models like OpenAI o1, we propose a video-thinking model to mitigate the hallucinations of LMMs via supervised reasoning fine-tuning (SRFT) and direct preference optimization (TDPO)-- where SRFT enhances reasoning capabilities while TDPO reduces hallucinations in the thinking process. Extensive experiments and analyses demonstrate the effectiveness. Remarkably, it improves the baseline by 7.65% in accuracy on hallucination evaluation and reduces the bias score by 4.5%. The code and data are public at https://github.com/Hongcheng-Gao/HAVEN.
Audio often serves as an auxiliary modality in video understanding tasks of audio-visual large language models (LLMs), merely assisting in the comprehension of visual information. However, a thorough understanding of videos significantly depends on auditory information, as audio offers critical context, emotional cues, and semantic meaning that visual data alone often lacks. This paper proposes an audio-centric video understanding benchmark (ACVUBench) to evaluate the video comprehension capabilities of multimodal LLMs with a particular focus on auditory information. Specifically, ACVUBench incorporates 2,662 videos spanning 18 different domains with rich auditory information, together with over 13k high-quality human annotated or validated question-answer pairs. Moreover, ACVUBench introduces a suite of carefully designed audio-centric tasks, holistically testing the understanding of both audio content and audio-visual interactions in videos. A thorough evaluation across a diverse range of open-source and proprietary multimodal LLMs is performed, followed by the analyses of deficiencies in audio-visual LLMs. Demos are available at https://github.com/lark-png/ACVUBench.
Temporal action detection (TAD) aims to identify and localize action instances in untrimmed videos, which is essential for various video understanding tasks. However, recent improvements in model performance, driven by larger feature extractors and datasets, have led to increased computational demands. This presents a challenge for applications like autonomous driving and robotics, which rely on limited computational resources. While existing channel pruning methods can compress these models, reducing the number of channels often hinders the parallelization efficiency of GPU, due to the inefficient multiplication between small matrices. Instead of pruning channels, we propose a Progressive Block Drop method that reduces model depth while retaining layer width. In this way, we still use large matrices for computation but reduce the number of multiplications. Our approach iteratively removes redundant blocks in two steps: first, we drop blocks with minimal impact on model performance; and second, we employ a parameter-efficient cross-depth alignment technique, fine-tuning the pruned model to restore model accuracy. Our method achieves a 25% reduction in computational overhead on two TAD benchmarks (THUMOS14 and ActivityNet-1.3) to achieve lossless compression. More critically, we empirically show that our method is orthogonal to channel pruning methods and can be combined with it to yield further efficiency gains.