Do Audio-Visual Large Language Models Really See and Hear?
This reveals a fundamental modality bias in AVLLMs, which is an incremental insight for researchers in multimodal AI, addressing how these models integrate audio and vision.
The study conducted the first mechanistic interpretability analysis of Audio-Visual Large Language Models (AVLLMs), finding that while they encode rich audio semantics internally, these capabilities are suppressed in final text generation when audio conflicts with vision, due to visual bias in fusion layers and limited audio alignment during training.
Audio-Visual Large Language Models (AVLLMs) are emerging as unified interfaces to multimodal perception. We present the first mechanistic interpretability study of AVLLMs, analyzing how audio and visual features evolve and fuse through different layers of an AVLLM to produce the final text outputs. We find that although AVLLMs encode rich audio semantics at intermediate layers, these capabilities largely fail to surface in the final text generation when audio conflicts with vision. Probing analyses show that useful latent audio information is present, but deeper fusion layers disproportionately privilege visual representations that tend to suppress audio cues. We further trace this imbalance to training: the AVLLM's audio behavior strongly matches its vision-language base model, indicating limited additional alignment to audio supervision. Our findings reveal a fundamental modality bias in AVLLMs and provide new mechanistic insights into how multimodal LLMs integrate audio and vision.