Procedural Graphs: agents maintain their own verifiable procedure graphs

简体中文

Public-observation parsing diagnostic. Implements a procedure–relation–procedure graph and a conservative edit gate, not LLM generation or real-environment execution.

Paper information

Field	Details
Paper	arXiv 2609.09153
Organization	Google (first listed affiliation of the first author)
First public release	2026-09-08 (arXiv v1)
Authors' code	No public author repository was found on 2026-09-12.
Adapter / method	procedural-graphs
Local reproduction code	src/auto_research/agent_research/latest_20260912.py

Original paper

Motivation and method

Represent successful experience as a procedure–relation–procedure graph rather than free-text fragments. Turn new experiences into candidate local edits, then admit them to long-term structure only after held-out verification. This reduces contamination from bad experiences.

flowchart LR
  O[Public trajectory] --> T[Procedure triples]
  T --> C[Candidate local edit]
  G[Existing procedure graph] --> C
  C --> V[Held-out verification gate]
  V -->|Pass| U[Update procedure graph]
  V -->|Reject| G

Key figure from the paper

Figure 2: Core architecture, main modules, and their connections. From the original paper; copyright remains with the authors.

Local reproduction and limits

See three-seed evomem-mini results. The local gate uses a deterministic proxy: an existing route must retain its first and last nodes. The policy sees no hidden answer or plan, and calls no real tools. This verifies graph updates and rejection paths only.