The authors introduced Procedural Graph, a graph of procedures that stores knowledge of what to do, in what order, and under what conditions. The authors describe LLM agents as programs with a language model that plan actions over long horizons and use external tools.

At each step, the system locates the agent’s active node, and a separate model turns the surrounding subgraph into situational guidance for the next action. The guidance biases the solver’s next action but does not dictate it.

An LLM refiner, a model for refining the graph, compares failed and successful trajectories and edits the graph’s topology and attributes. The authors write that across multiple datasets, task types, and LLMs, this approach delivered consistent gains over memory-based baselines.

Claim check:

  • The authors introduced Procedural Graph, a graph of procedures for storing knowledge of what to do, in what order, and under what conditions. (confirmed by the publication itself: evidence; «We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions.»)
  • The authors describe LLM agents as programs with a language model that plan actions over long horizons and use external tools. (confirmed by the publication itself: evidence; «Large language models are increasingly deployed as agents that plan over long horizons and act through external tools.»)
  • At each step, the system locates the agent’s active node, and a separate model turns the surrounding subgraph into situational guidance for the next action. (confirmed by the publication itself: evidence; «At each decision step, the framework localizes the agent’s active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver’s next action without dictating it.»)
  • The guidance biases the solver’s next action but does not dictate it. (confirmed by the publication itself: evidence; «At each decision step, the framework localizes the agent’s active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver’s next action without dictating it.»)
  • An LLM refiner compares failed and successful trajectories and edits the graph’s topology and attributes. (confirmed by the publication itself: evidence; «The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph’s topology and attributes»)
  • The authors write that across multiple datasets, task types, and LLMs, this approach delivered consistent gains over memory-based baselines. (confirmed by the publication itself: evidence; «Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines»)

Primary sources:

score 63.3 · kind research · revision 1 · stories st-ee5zbi