Researchers from Google, the Georgia Institute of Technology, and Peking University introduced Procedural Graphs, a framework for organizing process knowledge into editable graph structures. The study, submitted to arXiv on September 8, demonstrates that this approach helps AI agents maintain focus during multi-step tasks.
The system functions by representing processes as triplets of procedure, relation, and procedure. It utilizes a localization mechanism to track active steps and a guidance model to provide situational advice to the LLM. This replaces vague instructions with specific, step-level guidance based on the surrounding subgraph.
A self-evolution mechanism allows the graphs to improve by learning from past successes and failures. An LLM refiner validates modifications, ensuring only performance-enhancing changes are accepted. This automated process produced structures that matched or exceeded the quality of hand-designed graphs.
The framework was tested across seven benchmarks using four LLMs, including Gemini 3.1 Pro and Claude Sonnet 4.6. It achieved a first or joint-first ranking in 21 out of 24 settings. On the BFCL v3 benchmark, the method provided a 9 percentage point improvement in function calling.
In enterprise simulations, the technology significantly increased long-horizon survival rates, rising from 6% to 34%. Specifically, Gemini 3.1 Pro showed a 28 percentage point improvement on the EnterpriseArena benchmark, which evaluates the ability of agents to manage extended business processes.
