From Senses to Decisions: The Information Flow of Auditory and Visual Perception in Multimodal LLMs

In three linesInterpretability study on audio-visual LLMs (AVLLMs): traces information flow between audio and visual tokens in Qwen2.5-Omni and Video-SALMONN2 Plus (3B/7B scales). Authors demonstrate audio-visual tokens can be discarded post-information-transfer without prediction degradation, improving inference efficiency.

Summary generated by Claude — human-verified