Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
Summary
The paper proposes Procedural Graphs to structure procedural knowledge for LLM agents as (procedure, relation, procedure) triplets. At decision steps, an active node is localized and a guidance model provides step-level guidance to bias the next action without dictating it. The graph self-evolves by comparing failed and successful trajectories, editing topology and attributes to improve held-out performance, with reported gains over memory-based baselines across datasets and models.