omnimodal-llm

OmniScope is a training-free token compression framework for omnimodal LLMs that estimates audio and video relevance separately using the query as a shared anchor, achieving up to 3.53x prefill speedup and over 15% GPU memory reduction with minimal accuracy loss.