CoFiE: Coarse-to-Fine Evidence Selection for Efficient Streaming Video Understanding
CoFiE introduces a Coarse-to-Fine Evidence Selection framework for efficient streaming video understanding using open-source Vision Language Models (VLLMs). It decouples evidence selection into two stages: coarse filtering based on visual novelty and fine refinement using query-specific attention during LLM prefill. This method achieves state-of-the-art accuracy with significant improvements in efficiency and latency.