Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
Summary
The article introduces Procedural Graphs, a self-evolving structure that stores procedural knowledge as triplets to guide LLM agents. It presents a framework where guidance biases next actions without constraining them, and the graph self-edits by contrasting failed and successful trajectories to improve performance. Results show gains over memory-based approaches across various datasets and models, reducing manual engineering.