Human vision is dynamic and continuous. However, in video understanding with multimodal large language models (LLMs), existing methods primarily rely on static features extracted from images sampled at a fixed low frame rate of frame-per-second (FPS) $\leqslant$2, leading to critical visual information loss. In this paper, we introduce F-16, the first multimodal LLM designed for high-frame-rate video understanding. By increasing the frame rate to 16 FPS and compressing visual tokens within each 1-second clip, F-16 efficiently captures dynamic visual features while preserving key semantic information. Experimental results demonstrate that higher frame rates considerably enhance video understanding across multiple benchmarks, providing a new approach to improving video LLMs beyond scaling model size or training data. F-16 achieves state-of-the-art performance among 7-billion-parameter video LLMs on both general and fine-grained video understanding benchmarks, such as Video-MME and TemporalBench. Furthermore, F-16 excels in complex spatiotemporal tasks, including high-speed sports analysis (\textit{e.g.}, basketball, football, gymnastics, and diving), outperforming SOTA proprietary visual models like GPT-4o and Gemini-1.5-pro. Additionally, we introduce a novel decoding method for F-16 that enables highly efficient low-frame-rate inference without requiring model retraining. Upon acceptance, we will release the source code, model checkpoints, and data.
Training robust deep video representations has proven to be computationally challenging due to substantial decoding overheads, the enormous size of raw video streams, and their inherent high temporal redundancy. Different from existing schemes, operating exclusively in the compressed video domain and exploiting all freely available modalities, i.e., I-frames, and P-frames (motion vectors and residuals) offers a compute-efficient alternative. Existing methods approach this task as a naive multi-modality problem, ignoring the temporal correlation and implicit sparsity across P-frames for modeling stronger shared representations for videos of the same action, making training and generalization easier. By revisiting the high-level design of dominant video understanding backbones, we increase inference speed by a factor of $56$ while retaining similar performance. For this, we propose a hybrid end-to-end framework that factorizes learning across three key concepts to reduce inference cost by $330\times$ versus prior art: First, a specially designed dual-encoder scheme with efficient Spiking Temporal Modulators to minimize latency while retaining cross-domain feature aggregation. Second, a unified transformer model to capture inter-modal dependencies using global self-attention to enhance I-frame -- P-frame contextual interactions. Third, a Multi-Modal Mixer Block to model rich representations from the joint spatiotemporal token embeddings. Experiments show that our method results in a lightweight architecture achieving state-of-the-art video recognition performance on UCF-101, HMDB-51, K-400, K-600 and SS-v2 datasets with favorable costs ($0.73$J/V) and fast inference ($16$V/s). Our observations bring new insights into practical design choices for efficient next-generation spatiotemporal learners. Code is available.
Long-form video understanding is essential for various applications such as video retrieval, summarizing, and question answering. Yet, traditional approaches demand substantial computing power and are often bottlenecked by GPU memory. To tackle this issue, we present Long-Video Memory Network, Long-VMNet, a novel video understanding method that employs a fixed-size memory representation to store discriminative patches sampled from the input video. Long-VMNet achieves improved efficiency by leveraging a neural sampler that identifies discriminative tokens. Additionally, Long-VMNet only needs one scan through the video, greatly boosting efficiency. Our results on the Rest-ADL dataset demonstrate an 18x -- 75x improvement in inference times for long-form video retrieval and answering questions, with a competitive predictive performance.
Understanding long video content is a complex endeavor that often relies on densely sampled frame captions or end-to-end feature selectors, yet these techniques commonly overlook the logical relationships between textual queries and visual elements. In practice, computational constraints necessitate coarse frame subsampling, a challenge analogous to "finding a needle in a haystack." To address this issue, we introduce a semantics-driven search framework that reformulates keyframe selection under the paradigm of Visual Semantic-Logical Search. Specifically, we systematically define four fundamental logical dependencies: 1) spatial co-occurrence, 2) temporal proximity, 3) attribute dependency, and 4) causal order. These relations dynamically update frame sampling distributions through an iterative refinement process, enabling context-aware identification of semantically critical frames tailored to specific query requirements. Our method establishes new SOTA performance on the manually annotated benchmark in key-frame selection metrics. Furthermore, when applied to downstream video question-answering tasks, the proposed approach demonstrates the best performance gains over existing methods on LongVideoBench and Video-MME, validating its effectiveness in bridging the logical gap between textual queries and visual-temporal reasoning. The code will be publicly available.
Multimodal Large Language Models (MLLMs) have revolutionized video understanding, yet are still limited by context length when processing long videos. Recent methods compress videos by leveraging visual redundancy uniformly, yielding promising results. Nevertheless, our quantitative analysis shows that redundancy varies significantly across time and model layers, necessitating a more flexible compression strategy. We propose AdaReTaKe, a training-free method that flexibly reduces visual redundancy by allocating compression ratios among time and layers with theoretical guarantees. Integrated into state-of-the-art MLLMs, AdaReTaKe improves processing capacity from 256 to 2048 frames while preserving critical information. Experiments on VideoMME, MLVU, LongVideoBench, and LVBench datasets demonstrate that AdaReTaKe outperforms existing methods by 2.3% and 2.8% for 7B and 72B models, respectively, with even greater improvements of 5.9% and 6.0% on the longest LVBench. Our code is available at https://github.com/SCZwangxiao/video-FlexReduc.git.
State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames (640$\times$360) on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least 50% reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by 4.3% on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks.
Human processes video reasoning in a sequential spatio-temporal reasoning logic, we first identify the relevant frames ("when") and then analyse the spatial relationships ("where") between key objects, and finally leverage these relationships to draw inferences ("what"). However, can Video Large Language Models (Video-LLMs) also "reason through a sequential spatio-temporal logic" in videos? Existing Video-LLM benchmarks primarily focus on assessing object presence, neglecting relational reasoning. Consequently, it is difficult to measure whether a model truly comprehends object interactions (actions/events) in videos or merely relies on pre-trained "memory" of co-occurrences as biases in generating answers. In this work, we introduce a Video Spatio-Temporal Reasoning (V-STaR) benchmark to address these shortcomings. The key idea is to decompose video understanding into a Reverse Spatio-Temporal Reasoning (RSTR) task that simultaneously evaluates what objects are present, when events occur, and where they are located while capturing the underlying Chain-of-thought (CoT) logic. To support this evaluation, we construct a dataset to elicit the spatial-temporal reasoning process of Video-LLMs. It contains coarse-to-fine CoT questions generated by a semi-automated GPT-4-powered pipeline, embedding explicit reasoning chains to mimic human cognition. Experiments from 14 Video-LLMs on our V-STaR reveal significant gaps between current Video-LLMs and the needs for robust and consistent spatio-temporal reasoning.
Automated surgical workflow analysis is crucial for education, research, and clinical decision-making, but the lack of annotated datasets hinders the development of accurate and comprehensive workflow analysis solutions. We introduce a novel approach for addressing the sparsity and heterogeneity of annotated training data inspired by the human learning procedure of watching experts and understanding their explanations. Our method leverages a video-language model trained on alignment, denoising, and generative tasks to learn short-term spatio-temporal and multimodal representations. A task-specific temporal model is then used to capture relationships across entire videos. To achieve comprehensive video-language understanding in the surgical domain, we introduce a data collection and filtering strategy to construct a large-scale pretraining dataset from educational YouTube videos. We then utilize parameter-efficient fine-tuning by projecting downstream task annotations from publicly available surgical datasets into the language domain. Extensive experiments in two surgical domains demonstrate the effectiveness of our approach, with performance improvements of up to 7% in phase segmentation tasks, 8% in zero-shot phase segmentation, and comparable capabilities to fully-supervised models in few-shot settings. Harnessing our model's capabilities for long-range temporal localization and text generation, we present the first comprehensive solution for dense video captioning (DVC) of surgical videos, addressing this task despite the absence of existing DVC datasets in the surgical domain. We introduce a novel approach to surgical workflow understanding that leverages video-language pretraining, large-scale video pretraining, and optimized fine-tuning. Our method improves performance over state-of-the-art techniques and enables new downstream tasks for surgical video understanding.
Training-free video large language models (LLMs) leverage pretrained Image LLMs to process video content without the need for further training. A key challenge in such approaches is the difficulty of retaining essential visual and temporal information, constrained by the token limits in Image LLMs. To address this, we propose a two-stage method for selecting query-relevant tokens based on the LLM attention scores: compressing the video sequence and then expanding the sequence. However, during the compression stage, Image LLMs often exhibit a positional attention bias in video sequences, where attention is overly concentrated on later frames, causing early-frame information to be underutilized. To alleviate this attention bias during sequence compression, we propose Gridded Attention Pooling for preserving spatiotemporal structure. Additionally, we introduce Visual Summarization Tail to effectively utilize this bias, facilitating overall video understanding during sequence expansion. In this way, our method effectively Mitigates and Leverages attention Bias (LLaVA-MLB), enabling the frozen Image LLM for detailed video understanding. Experiments on several benchmarks demonstrate that our approach outperforms state-of-the-art methods, achieving superior performance in both efficiency and accuracy. Our code will be released.
Counterfactual reasoning is crucial for robust video understanding but remains underexplored in existing multimodal benchmarks. In this paper, we introduce \textbf{COVER} (\textbf{\underline{CO}}unterfactual \textbf{\underline{V}}id\textbf{\underline{E}}o \textbf{\underline{R}}easoning), a multidimensional multimodal benchmark that systematically evaluates MLLMs across the abstract-concrete and perception-cognition dimensions. Beyond prior multimodal benchmarks, COVER decomposes complex queries into structured sub-questions, enabling fine-grained reasoning analysis. Experiments on commercial and open-source models reveal a strong correlation between sub-question accuracy and counterfactual reasoning performance, highlighting the role of structured inference in video understanding. Furthermore, our results suggest a key insight: enhancing the reasoning capability of models is essential for improving the robustness of video understanding. COVER establishes a new standard for assessing MLLMs' logical reasoning abilities in dynamic environments.
Vision language models (VLMs) demonstrate strong capabilities in jointly processing visual and textual data. However, they often incur substantial computational overhead due to redundant visual information, particularly in long-form video scenarios. Existing approaches predominantly focus on either vision token pruning, which may overlook spatio-temporal dependencies, or keyframe selection, which identifies informative frames but discards others, thus disrupting contextual continuity. In this work, we propose KVTP (Keyframe-oriented Vision Token Pruning), a novel framework that overcomes the drawbacks of token pruning and keyframe selection. By adaptively assigning pruning rates based on frame relevance to the query, KVTP effectively retains essential contextual information while significantly reducing redundant computation. To thoroughly evaluate the long-form video understanding capacities of VLMs, we curated and reorganized subsets from VideoMME, EgoSchema, and NextQA into a unified benchmark named SparseKV-QA that highlights real-world scenarios with sparse but crucial events. Our experiments with VLMs of various scales show that KVTP can reduce token usage by 80% without compromising spatiotemporal and contextual consistency, significantly cutting computation while maintaining the performance. These results demonstrate our approach's effectiveness in efficient long-video processing, facilitating more scalable VLM deployment.
Existing Multimodal Large Language Models (MLLMs) encounter significant challenges in modeling the temporal context within long videos. Currently, mainstream Agent-based methods use external tools (e.g., search engine, memory banks, OCR, retrieval models) to assist a single MLLM in answering long video questions. Despite such tool-based support, a solitary MLLM still offers only a partial understanding of long videos, resulting in limited performance. In order to better address long video tasks, we introduce LVAgent, the first framework enabling multi-round dynamic collaboration of MLLM agents in long video understanding. Our methodology consists of four key steps: 1. Selection: We pre-select appropriate agents from the model library to form optimal agent teams based on different tasks. 2. Perception: We design an effective retrieval scheme for long videos, improving the coverage of critical temporal segments while maintaining computational efficiency. 3. Action: Agents answer long video-related questions and exchange reasons. 4. Reflection: We evaluate the performance of each agent in each round of discussion and optimize the agent team for dynamic collaboration. The agents iteratively refine their answers by multi-round dynamical collaboration of MLLM agents. LVAgent is the first agent system method that outperforms all closed-source models (including GPT-4o) and open-source models (including InternVL-2.5 and Qwen2-VL) in the long video understanding tasks. Our LVAgent achieves an accuracy of 80% on four mainstream long video understanding tasks. Notably, on the LongVideoBench dataset, LVAgent improves accuracy by up to 13.3% compared with SOTA.
Video large language models have achieved remarkable performance in tasks such as video question answering, however, their temporal understanding remains suboptimal. To address this limitation, we curate a dedicated instruction fine-tuning dataset that focuses on enhancing temporal comprehension across five key dimensions. In order to reduce reliance on costly temporal annotations, we introduce a multi-task prompt fine-tuning approach that seamlessly integrates temporal-sensitive tasks into existing instruction datasets without requiring additional annotations. Furthermore, we develop a novel benchmark for temporal-sensitive video understanding that not only fills the gaps in dimension coverage left by existing benchmarks but also rigorously filters out potential shortcuts, ensuring a more accurate evaluation. Extensive experimental results demonstrate that our approach significantly enhances the temporal understanding of video-LLMs while avoiding reliance on shortcuts.
Vision Language Models (VLMs) have demonstrated significant potential in various downstream tasks, including Image/Video Generation, Visual Question Answering, Multimodal Chatbots, and Video Understanding. However, these models often struggle with basic image transformations. This paper investigates the image-level understanding of VLMs, specifically CLIP by OpenAI and SigLIP by Google. Our findings reveal that these models lack comprehension of multiple image-level augmentations. To facilitate this study, we created an augmented version of the Flickr8k dataset, pairing each image with a detailed description of the applied transformation. We further explore how this deficiency impacts downstream tasks, particularly in image editing, and evaluate the performance of state-of-the-art Image2Image models on simple transformations.
Human daily activities can be concisely narrated as sequences of routine events (e.g., turning off an alarm) in video streams, forming an event vocabulary. Motivated by this, we introduce VLog, a novel video understanding framework that define video narrations as vocabulary, going beyond the typical subword vocabularies in existing generative video-language models. Built on the lightweight language model GPT-2, VLog feature three key innovations: (i) A generative retrieval model, marrying language model's complex reasoning capabilities with contrastive retrieval's efficient similarity search. (ii) A hierarchical vocabulary derived from large-scale video narrations using our narration pair encoding algorithm, enabling efficient indexing of specific events (e.g., cutting a tomato) by identifying broader scenarios (e.g., kitchen) with expressive postfixes (e.g., by the left hand). (iii) A vocabulary update strategy leveraging generative models to extend the vocabulary for novel events encountered during inference. To validate our approach, we introduce VidCap-Eval, a development set requiring concise narrations with reasoning relationships (e.g., before and after). Experiments on EgoSchema, COIN, and HiREST further demonstrate the effectiveness of VLog, highlighting its ability to generate concise, contextually accurate, and efficient narrations, offering a novel perspective on video understanding. Codes are released at https://github.com/showlab/VLog.
This paper introduces VideoScan, an efficient vision-language model (VLM) inference framework designed for real-time video interaction that effectively comprehends and retains streamed video inputs while delivering rapid and accurate responses. A longstanding challenge in video understanding--particularly for long-term or real-time applications--stems from the substantial computational overhead caused by the extensive length of visual tokens. To address this, VideoScan employs a single semantic carrier token to represent each frame, progressively reducing computational and memory overhead during its two-phase inference process: prefilling and decoding. The embedding of the semantic carrier token is derived from an optimized aggregation of frame-level visual features, ensuring compact yet semantically rich representations. Critically, the corresponding key-value pairs are trained to retain contextual semantics from prior frames, enabling efficient memory management without sacrificing temporal coherence. During inference, the visual tokens of each frame are processed only once during the prefilling phase and subsequently discarded in the decoding stage, eliminating redundant computations. This design ensures efficient VLM inference even under stringent real-time constraints. Comprehensive experiments on diverse offline and online benchmarks demonstrate that LLaVA-Video, supported by our method, achieves up to $\sim 5\times$ and $1.29\times$ speedups compared to its original version and previous efficient streaming video understanding approaches, respectively. Crucially, these improvements are attained while maintaining competitive performance and ensuring stable GPU memory consumption (consistently $\sim 18$GB, independent of video duration).
Video-based multimodal large language models (VMLLMs) have demonstrated remarkable potential in cross-modal video understanding. However, their abilities in fine-grained face comprehension remain largely underexplored. Given its pivotal role in human-centric intelligence, developing VMLLMs for facial understanding holds a fundamental problem. To address this gap, we propose FaVChat, the first VMLLM specifically designed for fine-grained facial video understanding. To facilitate its training, we construct a large-scale facial video dataset comprising over 60k videos, with the majority annotated with 83 fine-grained facial attributes. These attributes are incorporated to enrich GPT-4o-generated captions, yielding 60k high-quality video-summary pairs and an additional 170k fine-grained question-answering (QA) pairs. To effectively capture rich facial clues, we propose a hybrid model architecture composed of a general visual encoder, a dedicated facial encoder, and a mixture-of-experts-enhanced adapter for adaptive fusion of multi-source visual features. To mitigate information loss during feature transformation, we extract multi-granularity representations from the facial encoder and integrate them into the subsequent LLM. This design enhances the model's ability to comprehend and respond to questions involving diverse levels of visual details. We employ a progressive training paradigm, transitioning from video summarization to a high-quality subset of video QA, gradually increasing task complexity to enhance the model's fine-grained visual perception. We conduct extensive zero-shot evaluation on a couple of public benchmarks, demonstrating that FaVChat consistently surpasses existing VMLLMs across multiple tasks.
AI personal assistants, deployed through robots or wearables, require embodied understanding to collaborate effectively with humans. Current Multimodal Large Language Models (MLLMs) primarily focus on third-person (exocentric) vision, overlooking the unique aspects of first-person (egocentric) videos. Additionally, high acquisition costs limit data size, impairing MLLM performance. To address these challenges, we propose learning the mapping between exocentric and egocentric domains, leveraging the extensive exocentric knowledge within existing MLLMs to enhance egocentric video understanding. To this end, we introduce Ego-ExoClip, a pre-training dataset comprising 1.1M synchronized ego-exo clip-text pairs derived from Ego-Exo4D. Our approach features a progressive training pipeline with three stages: Teacher Self-Preparation, Teacher-Student Guidance, and Student Self-Practice. Additionally, we propose an instruction-tuning data EgoIT from multiple sources to strengthen the model's instruction-following capabilities, along with the EgoBench benchmark comprising eight different tasks for thorough evaluation. Extensive experiments across diverse egocentric tasks reveal that existing MLLMs perform inadequately in egocentric video understanding, while our model significantly outperforms these leading models.
Long Video Question Answering (LVQA) is challenging due to the need for temporal reasoning and large-scale multimodal data processing. Existing methods struggle with retrieving cross-modal information from long videos, especially when relevant details are sparsely distributed. We introduce UMaT (Unified Multi-modal as Text), a retrieval-augmented generation (RAG) framework that efficiently processes extremely long videos while maintaining cross-modal coherence. UMaT converts visual and auditory data into a unified textual representation, ensuring semantic and temporal alignment. Short video clips are analyzed using a vision-language model, while automatic speech recognition (ASR) transcribes dialogue. These text-based representations are structured into temporally aligned segments, with adaptive filtering to remove redundancy and retain salient details. The processed data is embedded into a vector database, enabling precise retrieval of dispersed yet relevant content. Experiments on a benchmark LVQA dataset show that UMaT outperforms existing methods in multimodal integration, long-form video understanding, and sparse information retrieval. Its scalability and interpretability allow it to process videos over an hour long while maintaining semantic and temporal coherence. These findings underscore the importance of structured retrieval and multimodal synchronization for advancing LVQA and long-form AI systems.
Localizing user-queried events through natural language is crucial for video understanding models. Recent methods predominantly adapt Video LLMs to generate event boundary timestamps to handle temporal localization tasks, which struggle to leverage LLMs' powerful semantic understanding. In this work, we introduce MeCo, a novel timestamp-free framework that enables video LLMs to fully harness their intrinsic semantic capabilities for temporal localization tasks. Rather than outputting boundary timestamps, MeCo partitions videos into holistic event and transition segments based on the proposed structural token generation and grounding pipeline, derived from video LLMs' temporal structure understanding capability. We further propose a query-focused captioning task that compels the LLM to extract fine-grained, event-specific details, bridging the gap between localization and higher-level semantics and enhancing localization performance. Extensive experiments on diverse temporal localization tasks show that MeCo consistently outperforms boundary-centric methods, underscoring the benefits of a semantic-driven approach for temporal localization with video LLMs.