Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
Yuxing Lu, Yicheng Chen, Shanchan Wu, Sercan Ö. Arık — Google / Georgia Tech / Peking University
Key Insight
A localized, editable execution graph that biases the next action and can self-evolve — the strongest architecture fit of the Sept 9 research run (98/100 fit).
What's Actually Supported
Consistent gains over memory baselines across multiple datasets, tasks, and LLMs, per the paper; self-evolution further improves results and can repair flawed expert priors.
Caveat / Limitations
No public code was found today. Guidance generation adds a model call; transfer and selective reuse of guidance are flagged as future work by the authors themselves.
