Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

The authors present Procedural Graphs, a representation of procedural knowledge as (procedure, relation, procedure) triplets that provides step-level guidance to LLM agents while allowing them to deviate when needed. An LLM refiner compares failed and successful trajectories, edits the graph’s topology and attributes, and retains only changes that improve held-out validation performance. Experiments on several datasets and backbone models show the approach consistently outperforms memory-based baselines, with additional gains from the self-evolution mechanism.