Probing Cross-modal Information Hubs in Audio-Visual LLMs

TL;DR

This paper investigates how audio and visual information are internally represented in AVLLMs, revealing the role of sink tokens and proposing a simple method to improve their cross-modal reasoning.

Contribution

It uncovers the internal encoding patterns of cross-modal information in AVLLMs and introduces a training-free mitigation technique based on these insights.

Findings

AVLLMs encode integrated audio-visual info mainly in sink tokens.

A subset of sink tokens, cross-modal sink tokens, store cross-modal info.

The proposed method improves AVLLMs' reliance on integrated cross-modal information.

Abstract

Audio-visual large language models (AVLLMs) have recently emerged as a powerful architecture capable of jointly reasoning over audio, visual, and textual modalities. In AVLLMs, the bidirectional interaction between audio and video modalities introduces intricate processing dynamics, necessitating a deeper understanding of their internal mechanisms. However, unlike extensively studied text-only or large vision language models, the internal workings of AVLLMs remain largely unexplored. In this paper, we focus on cross-modal information flow between audio and visual modalities in AVLLMs, investigating where information derived from one modality is encoded within the token representations of the other modality. Through an analysis of multiple recent AVLLMs, we uncover two common findings. First, AVLLMs primarily encode integrated audio-visual information in sink tokens. Second, sink tokens do not uniformly hold cross-modal information. Instead, a distinct subset of sink tokens, which we term cross-modal sink tokens, specializes in storing such information. Based on these findings, we further propose a simple training-free hallucination mitigation method by encouraging reliance on integrated cross-modal information within cross-modal sink tokens. Our code is available at https://github.com/kaistmm/crossmodal-hub.