Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

Researchers from Google, Georgia Tech, and Peking University have introduced the Procedural Graph, a self-evolving knowledge structure that stores procedural steps as (procedure, relation, procedure) triplets to guide LLM agents through complex tasks. Unlike traditional approaches that rely on implicit action history, the Procedural Graph provides explicit step-level guidance that biases (but doesn't dictate) agent decisions, while automatically refining itself by contrasting failed and successful trajectories. The framework consistently outperforms memory-based baselines across multiple datasets and tasks, even starting from minimal skeletons or correcting flawed expert priors.