R²Orch：通过自博弈与技能内化循环实现智能体编排的递归自我改进
原标题：Paper page - R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop
AI 摘要
该论文提出 R²Orch，一种基于递归自博弈的智能体编排训练范式：内循环让编排器在自选任务上自我提升，外循环优化内循环的学习方式，并通过「递归可学习性」奖励 Proposer 选择技能撤除后结果仍不确定的任务，配合技能撤除机制逐步走向零技能部署。在十二个数据集上，零技能配置的 9B 编排器在问答、数学、代码和交互任务上均优于基线，分布内平均分从 86.0 提升至 89.8，分布外从 76.1 提升至 78.8，AIME 2026 从 74.0 升至 81.3。代码与模型已开源。
正文节选
R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop Abstract R²Orch uses recursive self-play with learnable task selection and skill withdrawal to train orchestrators without hand-designed curricula, improving performance across reasoning and interactive tasks. In recent years, agentic orchestration has been widely applied to solve complex human tasks. However, existing orchestration learning still faces key challenges, including hand-des