Reproduce “GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maureese Williams, Dymitr Nowicki
A framework named GATS integrates layered world models with an Upper Confidence Bound (UCB1)-based tree search to enable efficient and deterministic multi-step planning for LLM agents. It achieved a 100% success rate and optimal plan lengths on synthetic tasks and a comprehensive stress test, performing all planning steps without requiring LLM inference, thus eliminating planning variance.
Key insights
- Decoupling the world model from the LLM during inference-time planning is effective for reducing computational costs and achieving deterministic agent behavior.
- A layered world model prioritizing symbolic rules and learned statistics over LLM inference can efficiently and robustly predict state transitions.
- Combining systematic search, such as UCB1, with a persistent graph memory that merges states and reuses statistics leads to superior and more reliable planning compared to LLM-guided exploration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.