Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

gist

Procedural Graphs keep what-to-do knowledge outside model weights as an editable directed graph of procedure nodes and attributed transitions. At each step the framework localizes the active node and a guidance model turns the local neighborhood into situational advice that biases the solver without hard-coding the next action. Offline, an LLM refiner contrasts failed and successful trajectories, proposes topology and attribute edits, and commits only candidates that hold or improve held-out validation while remembering rejections. From a minimal skeleton the loop matches or beats hand-designed graphs and can repair a bad expert prior.

ideas

• Mirror knowledge graphs for procedures. Store (procedure, relation, procedure) triplets so agents answer what-to-do with an inspectable graph instead of reconstructing order from a flat history.
• Guide from the local neighborhood. Locate the active node, retrieve its k-hop transitions with condition/guidance/pitfalls attributes, and generate step-level situational advice rather than dumping the full graph.
• Bias without dictating. Append generative guidance to the solver prompt so the ReAct loop stays free to reason while admissible transitions stay explicit.
• Evolve under a validation gate. Contrast failed versus successful trajectories, edit topology and attributes, keep only candidates that do not hurt held-out score, and retain rejected edits as negative constraints.
• Construction modes matter. Evolution from scratch can match or surpass expert graphs; iterative repair recovers when a hand-crafted prior initially hurts performance.

quotes