Research · Nov 5, 2026, 00:00 UTC

Turn-level GRPO method TurnWise improves RL for streaming video understanding

TurnWise assigns separate advantages to each response or waiting decision instead of one trajectory-level advantage, and averages token losses per turn. The authors report it outperforms prior RL approaches for streaming video understanding without a learned critic or extra rollouts.

• Topicresearch
• Topicbenchmark