Query-Adaptive Video Context Compression and Frame-Level Visual Dependence for Long-Video Speculative Drafting
Select video draft tokens dynamically by question relevance and score visual dependence per frame so long-video drafters focus on informative moments instead of uniform strides.