An Empirical Study on How Video-LLMs Answer Video Questions

• URL: http://arxiv.org/abs/2508.15360v1
• Date: Thu, 21 Aug 2025 08:42:35 GMT
• Title: An Empirical Study on How Video-LLMs Answer Video Questions
• Authors: Chenhui Gou, Ziyu Ma, Zicheng Duan, Haoyu He, Feng Chen, Akide Liu, Bohan Zhuang, Jianfei Cai, Hamid Rezatofighi,
• Abstract summary: Video Large Language Models (Video-LLMs) have shown strong capabilities in answering video questions.<n>To our knowledge, this is the first work to systematically uncover how Video-LLMs internally process and understand video content.
• Score: 41.97630658989303
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Taking advantage of large-scale data and pretrained language models, Video Large Language Models (Video-LLMs) have shown strong capabilities in answering video questions. However, most existing efforts focus on improving performance, with limited attention to understanding their internal mechanisms. This paper aims to bridge this gap through a systematic empirical study. To interpret existing VideoLLMs, we adopt attention knockouts as our primary analytical tool and design three variants: Video Temporal Knockout, Video Spatial Knockout, and Language-to-Video Knockout. Then, we apply these three knockouts on different numbers of layers (window of layers). By carefully controlling the window of layers and types of knockouts, we provide two settings: a global setting and a fine-grained setting. Our study reveals three key findings: (1) Global setting indicates Video information extraction primarily occurs in early layers, forming a clear two-stage process -- lower layers focus on perceptual encoding, while higher layers handle abstract reasoning; (2) In the fine-grained setting, certain intermediate layers exert an outsized impact on video question answering, acting as critical outliers, whereas most other layers contribute minimally; (3) In both settings, we observe that spatial-temporal modeling relies more on language-guided retrieval than on intra- and inter-frame self-attention among video tokens, despite the latter's high computational cost. Finally, we demonstrate that these insights can be leveraged to reduce attention computation in Video-LLMs. To our knowledge, this is the first work to systematically uncover how Video-LLMs internally process and understand video content, offering interpretability and efficiency perspectives for future research.

Related papers

• Map the Flow: Revealing Hidden Pathways of Information in VideoLLMs [42.00309718904487]
  We investigate the internal information flow of VideoLLMs using mechanistic interpretability techniques.<n>Our analysis reveals consistent patterns across diverse VideoQA tasks.<n>These findings provide a blueprint on how VideoLLMs perform temporal reasoning.
  arXiv  Detail & Related papers  (2025-10-15T07:59:06Z)
• Generative Video Matting [57.186684844156595]
  Video matting has traditionally been limited by the lack of high-quality ground-truth data.<n>Most existing video matting datasets provide only human-annotated imperfect alpha and foreground annotations.<n>We introduce a novel video matting approach that can effectively leverage the rich priors from pre-trained video diffusion models.
  arXiv  Detail & Related papers  (2025-08-11T12:18:55Z)
• A Survey on Video Temporal Grounding with Multimodal Large Language Model [107.24431595873808]
  Recent advancement in temporal grounding (VTG) has significantly enhanced fine-grained video understanding.<n>With superior multimodal comprehension and reasoning abilities, VTG approaches based on MLLMs (VTG-MLLMs) are gradually surpassing traditional fine-tuned methods.<n>Despite extensive surveys on general video-language understanding, comprehensive reviews specifically addressing VTG-MLLMs remain scarce.
  arXiv  Detail & Related papers  (2025-08-07T08:52:11Z)
• ARC-Hunyuan-Video-7B: Structured Video Comprehension of Real-World Shorts [56.75723197779384]
  ARC-Hunyuan-Video is a multimodal model that processes visual, audio, and textual signals end-to-end for structured comprehension.<n>Our model is capable of multi-granularity timestamped video captioning and summarization, open-ended video question answering, temporal video grounding, and video reasoning.
  arXiv  Detail & Related papers  (2025-07-28T15:52:36Z)
• SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding [23.96372422130216]
  Video-based Large Language Models (VideoVid-LLMs) have witnessed substantial advancements in recent years.<n>They struggle with fine-grained understanding, particularly in aspects such as visual dynamics and video details inquiries.<n>To tackle these shortcomings, we find that fine-tuning Video-LLMs on self-supervised fragment tasks greatly improve their fine-grained video understanding abilities.
  arXiv  Detail & Related papers  (2025-04-10T13:40:34Z)
• H2VU-Benchmark: A Comprehensive Benchmark for Hierarchical Holistic Video Understanding [25.111988967973147]
  Existing benchmarks for evaluating video understanding exhibit significant limitations in coverage, task diversity, and scene adaptability.<n>We propose a hierarchical and holistic video understanding benchmark designed to evaluate both general video and online streaming video comprehension.<n>This benchmark contributes three key features: extended video duration, comprehensive assessment tasks, andEnriched video data.
  arXiv  Detail & Related papers  (2025-03-31T12:32:51Z)
• PreMind: Multi-Agent Video Understanding for Advanced Indexing of Presentation-style Videos [22.39414772037232]
  PreMind is a novel multi-agent multimodal framework for understanding/indexing lecture videos.<n>It generates multimodal indexes through three key steps: extracting slide visual content, transcribing speech narratives, and consolidating these visual and speech contents into an integrated understanding.<n>Three innovative mechanisms are also proposed to improve performance: leveraging prior lecture knowledge to refine visual understanding, detecting/correcting speech transcription errors using a VLM, and utilizing a critic agent for dynamic iterative self-reflection in vision analysis.
  arXiv  Detail & Related papers  (2025-02-28T20:17:48Z)
• Sparrow: Data-Efficient Video-LLM with Text-to-Image Augmentation [57.34255010956452]
  This work revisits scaling with synthetic data and focuses on developing video-LLMs from a data-centric perspective.<n>We propose a data augmentation method called Sparrow, which synthesizes video-like samples from pure text instruction data.<n>Our proposed method achieves performance comparable to or even superior to that of baselines trained with significantly more samples.
  arXiv  Detail & Related papers  (2024-11-29T18:59:54Z)
• VaQuitA: Enhancing Alignment in LLM-Assisted Video Understanding [63.075626670943116]
  We introduce a cutting-edge framework, VaQuitA, designed to refine the synergy between video and textual information. At the data level, instead of sampling frames uniformly, we implement a sampling method guided by CLIP-score rankings. At the feature level, we integrate a trainable Video Perceiver alongside a Visual-Query Transformer.
  arXiv  Detail & Related papers  (2023-12-04T19:48:02Z)
• InternVideo: General Video Foundation Models via Generative and Discriminative Learning [52.69422763715118]
  We present general video foundation models, InternVideo, for dynamic and complex video-level understanding tasks. InternVideo efficiently explores masked video modeling and video-language contrastive learning as the pretraining objectives. InternVideo achieves state-of-the-art performance on 39 video datasets from extensive tasks including video action recognition/detection, video-language alignment, and open-world video applications.
  arXiv  Detail & Related papers  (2022-12-06T18:09:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.