One Ranking, Any Budget: Matryoshka Evidence-to-Context Frame Selection for Long-Video Understanding
Matryoshka Evidence-to-Context (MEC) is a training-free method for selecting frames from long videos for Large Multimodal Models (LMMs). It solves the problem of choosing subsets for different frame budgets by creating one priority sequence. This sequence concentrates relevant evidence and gradually adds broader context, allowing truncation to any required budget without reselection.