跳到正文
arXiv Artificial Intelligence· Mingda Zhang, Hanwen Zhang, Qiang Huang, Zijia Wang, Pengfei Guo, Yuchen Zhang, Jionghao Zhu, Xiaoying Tang·· 4 小时前AI 评分34

EvoSteer:基于参考锚定信用分配的在线自演化图编排

EvoSteer: Online Self-Evolving Graph Orchestration via Reference-Anchored Credit Assignment

AI 导读

EvoSteer 提出一种在线自演化图编排范式,编排器在执行过程中构建运行团队,并依据执行特征与学习到的价值估计修复看似合理但失败的步骤。配套的 Anchored Trajectory Balance(AnchorTB)以冻结参考平衡子轨迹,为每个编排动作分配系数;Validated Skill Admission 则在共享名义测试预算下通过序贯检验决定技能晋升。

来源:arXiv Artificial Intelligence · arxiv.org