Who Wins the Conflict? Mechanistic Interpretability of Text Bias in Audio LLMs
For researchers and developers of multimodal AI systems, this work provides mechanistic insights and a practical intervention to mitigate text bias in audio LLMs, though the method is training-free and incremental.
This work presents the first mechanistic analysis of text dominance in Audio LLMs, revealing that text actively suppresses intact audio representations. A training-free intervention called back-patching, which routes late-layer audio activations to earlier layers, consistently reduces this bias.
While Audio Large Language Models (Audio LLMs) excel at multimodal understanding, they suffer from text dominance, a bias where models blindly favor text over acoustic evidence, causing hallucinations. However, the internal mechanisms underlying how these models behave when audio and textual inputs contradict each other remain unexplored. In this work, we present the first mechanistic analysis of this phenomenon by tracing the propagation of internal representations across layers. Our investigation reveals three key findings: (i) text dominance is systematically and empirically across models; (ii) while text and audio rely on functionally distinct pathways, they ultimately converge into a shared semantic space in late layers; and (iii) the text pathway does not erase audio information, but rather actively suppresses intact audio representations. Building on these insights, we leverage back-patching, a training-free intervention that routes late-layer audio activations back into earlier layers. This amplifies the audio representations, enabling them to overcome textual suppression. Our evaluation shows that back-patching consistently reduces text dominance, paving the way for mechanistic multimodal alignment under conflict.