TurnWise: Turn-Level Credit Assignment for Streaming Video RL
November 3, 2026
TurnWise introduces a turn-level GRPO framework for streaming video understanding to fix reward misalignment. By using silence-completed prefix scoring for credit assignment, it prevents redundant responses and encourages useful follow-ups in interactive video settings.
HOW THIS AFFECTS YOU
●
builderThis can reduce redundancy in streaming video models you are deploying.
●
researcherYou can improve video agent training by aligning rewards with specific interaction turns rather than entire trajectories.