Multi-turn Reinforcement Learning

A reinforcement learning approach in which an agent's multi-turn interaction process serves as the training signal. The quality of the final task outcome is used as the reward, training models to make optimal tool-calling decisions and policy selections across multi-step decision chains.