A research team including Google has proposed "Procedural Graph," a method to explicitly manage and evolve the action sequences of LLM (Large Language Model) agents.

Conventional LLM agents select their next action based on accumulated history, which often leads to challenges such as losing sight of the objective over long-term tasks or calling tools in an inappropriate order. In contrast, Procedural Graph employs a graph structure consisting of triplets of "procedure, relation, procedure," explicitly indicating what the agent should do next.

A key feature of this framework is its self-evolving capability. An LLM-based refiner compares failed trajectories with successful ones to edit the structure and attributes of the graph. The system is designed so that only edits that maintain or improve performance on validation data are applied. This allows the construction of graphs that outperform expert-designed structures without the need for manual engineering.

In validations using multiple datasets and models, the research team confirmed that Procedural Graph consistently demonstrates higher performance compared to traditional memory-based baselines.

Source: