GATS framework achieves 100% success rate on agent planning tasks with zero LLM calls during inference

A new planning framework called GATS (Graph-Augmented Tree Search) introduces a three-layer world model to reduce reliance on LLM inference during multi-step agent planning tasks. The framework combines systematic UCB1-based tree search with exact symbolic action matching (L1), statistics learned from execution logs (L2), and LLM-based prediction for unknown actions (L3).

In synthetic planning tasks with branching paths and dead-ends, GATS achieved a 100% success rate, outperforming LATS (92%) and ReAct (64%). On a 12-scenario stress test spanning coding workflows, web navigation, and long-horizon tasks, GATS maintained a 100% success rate, while LATS dropped to 88.9% and ReAct to 23.9%.

GATS requires zero LLM calls per task during planning, compared to 37 per task for LATS, and produces deterministic plans with zero variance across runs. The authors argue this approach substantially outperforms LLM-guided exploration for agent planning by reducing computational costs and eliminating stochastic behavior.