ReFoCUS: Reinforcement-guided Frame Optimization for Contextual Understanding

ReFoCUS (Reinforcement-guided Frame Optimization for Contextual Understanding) introduces a novel framework that employs online policy-gradient reinforcement learning to optimize frame selection for video-based large multi-modal models (LMMs). By utilizing reward signals from reference models, ReFoCUS learns to select frames that enhance temporally grounded responses without requiring explicit supervision, improving reasoning accuracy on various video QA benchmarks. This advancement is significant for practitioners as it allows for more effective video understanding in AI applications by aligning frame selection strategies with internal model utility, thereby enhancing contextual relevance in responses.