An Empirical Study on How Video-LLMs Answer Video Questions

TL;DR

This paper systematically investigates how Video-LLMs internally process video content for question answering, revealing the roles of different layers and attention mechanisms, and proposes ways to improve efficiency based on these insights.

Contribution

It introduces a novel empirical analysis using attention knockouts to interpret Video-LLMs, uncovering layer-specific functions and guiding efficiency improvements.

Findings

Video information extraction mainly occurs in early layers.

Intermediate layers act as critical outliers in fine-grained analysis.

Spatial-temporal modeling relies more on language-guided retrieval than intra-video attention.

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.