Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
The paper proposes a graph layer that tells an agent what procedural step it is in, then lets the LLM choose the next action.
Procedural Graphs encode “what to do” as linked procedure triplets, giving agents step-level guidance without hard-coding every move. The graph can update itself by comparing failed and successful runs, keeping edits that improve validation performance and retaining rejected ones as warnings. The authors say the method can start from a minimal skeleton, match or beat hand-built graphs, and repair flawed expert designs. Across tested datasets, task types, and LLMs, it outperformed memory-based baselines. ArXiv · AI/CL/LG's note
Procedural Graphs encode “what to do” as linked procedure triplets, giving agents step-level guidance without hard-coding every move. The graph can update itself by comparing failed and successful runs, keeping edits that improve validation performance and retaining rejected ones as warnings. The authors say the method can start from a minimal skeleton, match or beat hand-built graphs, and repair flawed expert designs. Across tested datasets, task types, and LLMs, it outperformed memory-based baselines. ArXiv · AI/CL/LG's note
score 5