Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
21 days ago
- LLM agents struggle with long-horizon planning, often losing track of objectives and invoking tools out of order.
- The Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets to guide agent actions.
- A guidance model uses the agent's active node and surrounding subgraph to provide situational, non-binding action suggestions.
- The graph self-evolves by comparing failed and successful trajectories, editing its topology to improve performance.
- Self-evolution starts from a minimal skeleton and can match or surpass hand-designed graphs, even repairing flawed expert priors.
- The method consistently outperforms memory-based baselines across datasets, task types, and LLMs without manual engineering.