65 upvotes
Turn-Level Credit Boosts Streaming Video RL Decisions
What is Turns Matter: Advancing Streaming Video RL via Turn-Level Credit?
The paper introduces a method that gives reinforcement-learning agents credit for each decision turn while watching a video stream, instead of only at the end of a whole episode. This finer-grained feedback lets models better learn when to answer and when to keep watching.
Who it is for
Researchers building real-time video understanding systems can use this to improve response timing without redesigning reward functions.