AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression
AVOC addresses challenges in understanding long audio-video content by reframing token compression as a top-K retrieval problem. It introduces a learnable module that selects the most relevant and diverse tokens from hour-long sequences, achieving state-of-the-art performance on benchmarks like OmniVideoBench.