谷歌研究人员与乔治亚理工学院、北京大学合作提出了一种名为"过程图"的新型执行框架，旨在为大语言模型代理提供结构化的步骤级指导。1该框架以三元组形式（程序、关系、程序）存储程序知识，相比传统隐式的历史积累方式，能够提供更加显式和可控的执行指导。1
过程图的核心创新在于其自进化机制。1该机制由大语言模型优化器驱动，通过对比失败和成功的执行轨迹来自动调整图的拓扑结构和属性，仅当编辑能够保留或改进验证性能时才予以提交。1这一设计使框架能够从最小骨架开始自动构建，最终可以匹配甚至超越手工设计的图，并能有效修正存在缺陷的专家先验。1
在实验评估中，该框架在多个数据集、不同任务类型和多种骨干大语言模型上均表现出优异成绩，一致地超越了记忆型基线方案，而自进化过程进一步提升了整体性能。1
Researchers from Google, Georgia Institute of Technology, and Peking University have introduced procedural graphs, a novel framework designed to guide language model agents through step-by-step execution structures 1. The approach stores procedural knowledge as (procedure, relation, procedure) triplets, offering explicit execution guidance compared to implicit accumulation through historical trajectories 1.
The framework incorporates a self-evolution mechanism wherein an LLM optimizer compares failed and successful execution paths to refine the graph's topology and attributes, only committing edits that preserve or improve validation performance 1. Rather than requiring manual design, procedural graphs automatically construct themselves from a minimal skeleton, demonstrating the ability to match or even surpass hand-crafted graphs while correcting defective expert priors 1. Google researchers Yuxing Lu, Yicheng Chen, Shanchan Wu, and Sercan Arik contributed to the work alongside collaborators from the two academic institutions 1.
Across multiple datasets, task types, and underlying language models, procedural graphs consistently outperform memory-based baselines, with self-evolution delivering further performance improvements 1.
评论
还没有评论，欢迎留下第一条。