Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
The paper proposes a graph-based control layer that lets LLM agents refine their own procedures after failures.
Procedural Graphs encode what an agent should do, in what order, and under what conditions as linked procedure triplets. At each step, the system locates the agent in the graph and uses nearby nodes to guide the next action without fully prescribing it. A refiner compares failed and successful trajectories, edits the graph, and keeps changes only when validation performance holds or improves. The authors report gains across datasets, task types, and LLMs, including cases where the system starts from a minimal or flawed graph. HF Daily Papers' note
Procedural Graphs encode what an agent should do, in what order, and under what conditions as linked procedure triplets. At each step, the system locates the agent in the graph and uses nearby nodes to guide the next action without fully prescribing it. A refiner compares failed and successful trajectories, edits the graph, and keeps changes only when validation performance holds or improves. The authors report gains across datasets, task types, and LLMs, including cases where the system starts from a minimal or flawed graph. HF Daily Papers' note
score 5