What is the objective of reasoning with reinforcement learning?

Damek Davis, Benjamin Recht

We show that several popular algorithms for reinforcement learning in large language models with binary rewards can be viewed as stochastic gradient ascent on a monotone transform…