Hejie Cui

Multi-turn RL agents can learn far more effectively by explicitly monitoring and controlling uncertainty at both the token and turn levels, leading to more stable training and higher performance.