A knowledge graph for what-to-do: procedure triplets that self-evolve by contrasting failed trajectories with successful ones

arXiv 2609.09153·medium signalXBlueskyLinkedInCopy link

A Procedural Graph stores (procedure, relation, procedure) triplets the way a knowledge graph stores entity triplets, answering what-to-do rather than what-is. At each decision step the framework localizes the agent's active node and a guidance model turns the surrounding subgraph into step-level situational guidance that biases the next action without dictating it, addressing the failure where long trajectories lose the objective, call tools out of order and repeat dead ends. The graph edits itself: an LLM refiner contrasts failed against successful trajectories and commits topology edits only when held-out validation performance holds or improves, retaining rejected edits to avoid repeating them, and starting from a minimal skeleton it matched or beat hand-designed graphs and repaired a flawed expert prior.