HuggingFace Daily Papers· Heming Zou, Qi Wang, Yixiu Mao, Yun Qu, Yuyi Guo, Zeyuan Liu, Chenglin Li, Weijie Liu, Saiyong Yang, Xiangyang Ji·· 10 天前AI 评分40
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
AI 导读
TreeOPSD 框架通过树感知后验自蒸馏解决长时序强化学习的信用分配稀疏性和策略漂移问题，在 Deep Research 等基准上显著提升 Qwen3-30B-A3B-Thinking-2507 的平均准确率 4.97 点。该框架无需价值网络或外部教师模型，通过递归回溯终端奖励生成隐式层级信用，并映射为有界教师分布以约束策略更新。
来源：HuggingFace Daily Papers · arxiv.org