Learning to Coach for Experiential Learning
The paper trains a separate coach model to turn an AI’s past attempts into guidance the same frozen actor can use.
L2C rewards the coach when its advice leads the actor to a correct guided response. The authors test rewards aimed at improving the same problem and rewards meant to transfer across other instances. In math reasoning and interactive text games, the trained coach beats self-refinement and an untrained coach. More coaching iterations improved accuracy more efficiently than simply giving the actor a larger decoding budget. ArXiv · AI/CL/LG's note
L2C rewards the coach when its advice leads the actor to a correct guided response. The authors test rewards aimed at improving the same problem and rewards meant to transfer across other instances. In math reasoning and interactive text games, the trained coach beats self-refinement and an untrained coach. More coaching iterations improved accuracy more efficiently than simply giving the actor a larger decoding budget. ArXiv · AI/CL/LG's note
score 5