VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting
The episode discusses 'VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting,' a paper by Shoubin Yu, Yue Zhang, and Mohit Bansal. The paper proposes an input-adaptive reinforcement learning framework that uses training-time visual prompting and self-distillation to improve a model's ability to find and track evidence in videos. The authors show strong performance on video understanding tasks.