Do Audio-Visual Large Language Models Really See and Hear?

TL;DR

Analyzing the internal mechanisms of Audio-Visual Large Language Models (AVLLMs) reveals a bias where visual information suppresses audio information, stemming from training data imbalance.

Problem

It was unclear whether AVLLMs, which integrate audio and visual information to generate text, actually utilize both modalities in a balanced manner, especially when audio conflicts with vision.

Approach

As the first mechanistic interpretability study, we analyzed how audio and visual features evolve and fuse through different layers of an AVLLM. Probing techniques were used to evaluate audio information in intermediate layers and compare it with the final output.

Results & Contribution

Useful audio information exists in intermediate layers, but deeper fusion layers exhibit a bias where visual representations suppress audio. Additionally, AVLLM audio behavior closely matches its vision-language base model, suggesting limited additional alignment to audio supervision. This study provides new mechanistic insights into modality bias in multimodal LLMs.