Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
21 days ago
- Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets, providing step-level guidance to agents.
- Guidance is used as a bias that influences the solver's next action without dictating it, allowing deviation when needed.
- The graph evolves via an LLM refiner that contrasts failed and successful trajectories, committing edits that preserve or improve held-out validation performance.
- It can repair flawed expert priors and builds effective graphs from minimal skeletons.
- Across multiple datasets, task types, and LLMs, Procedural Graph outperforms memory-based baselines, with self-evolution adding further gains.