Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

AI Digest

论文提出程序图结构，通过组织程序知识三元组解决LLM代理长轨迹规划中的目标丢失和工具误用问题，实现自演进优化。

This paper introduces procedural graphs to organize procedural knowledge for LLM agents, enabling self-evolving execution structures that improve long-horizon planning and tool usage.

Key points

• 程序图用(过程,关系,过程)三元组组织任务步骤，替代传统历史记录 Procedural graphs organize task steps as (procedure, relation, procedure) triplets
• 自演进机制通过对比成功/失败轨迹优化图结构和属性 Self-evolution optimizes graph structure via success/failure trajectory comparison
• 可修复专家先验缺陷并超越人工设计的执行流程 Can repair flawed expert priors and surpass hand-designed workflows
• 在多数据集和LLM模型上均展现优于记忆基线的性能 Demonstrates consistent performance gains across datasets and LLMs
• 指导模型将子图转化为情境指导，而非直接指令动作 Guidance models translate subgraphs into situational hints

Takeaway: 程序图为LLM代理提供结构化决策框架，实现智能演进的执行规划。 / Procedural graphs offer structured decision-making for LLM agents with intelligent evolution.

Why it matters 为LLM长期规划提供新范式，可提升复杂任务自动化水平。