Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
What it is
This paper introduces the Procedural Graph, a framework that organizes procedural knowledge into (procedure, relation, procedure) triplets to guide LLM agents. It uses a guidance model to translate the active subgraph into step-level advice and a self-evolving loop to refine the graph topology based on trajectory outcomes.
The problem it solves
LLM agents often lose track of objectives, invoke tools out of order, and repeat unproductive actions as their interaction history accumulates, due to the lack of structured procedural knowledge.
What you could use it for
- Improving long-horizon planning consistency in LLM agents
- Preventing tool invocation order errors in complex workflows
- Automating the refinement of agent execution strategies without manual engineering
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.