GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning

Short summary

GATS (Graph-Augmented Tree Search) is a planning framework that eliminates LLM calls during inference by combining UCB1-based tree search with a three-layer world model (symbolic matching, learned statistics, and LLM prediction for unknowns). It achieves 100% success on 12 challenging scenarios versus 88.9% for LATS and 23.9% for ReAct, with zero LLM calls per task during planning. Results show systematic search with learned world models can outperform LLM-guided exploration for agent planning.

• •GATS eliminates LLM calls during planning using UCB1 tree search + layered world model
• •100% success on 12 stress-test scenarios vs 88.9% LATS and 23.9% ReAct
• •Deterministic plans with zero variance across runs, zero LLM calls per task

Generated with AI, which can make mistakes.