ATPO: 自适应树策略优化用于多轮医疗对话

Effective information seeking in multi-turn medical dialogues is crucial for accurate diagnosis, particularly when dealing with incomplete information. The uncertainty inherent in user-agent interactions poses challenges for aligning Large Language Models (LLMs), which we model as a Hierarchical Markov Decision Process (H-MDP). Conventional Reinforcement Learning (RL) methods like Group Relative Policy Optimization (GRPO) struggle with long-horizon credit assignment, while Proximal Policy Optimization (PPO) suffers from unstable value estimation in this context. To address these issues, we propose a novel uncertainty-aware Adaptive Tree Policy Optimization (ATPO) algorithm. Our method adaptively allocates the rollout budget to enhance decision-making efficiency and accuracy in multi-turn dialogues.