Researchers Eliminate LLM Inference From Agent Planning, Achieve 100% Success Rate With Graph-Augmented Search

A new framework called GATS combines tree search with layered world models to eliminate LLM calls during AI agent planning while outperforming existing methods like LATS and ReAct. In stress tests across coding, web navigation, and long-horizon tasks, GATS achieved 100% success compared to 88.9% for LATS and 23.9% for ReAct, while requiring zero LLM calls per task versus 37 for LATS.

Why it matters: As AI agents move toward production use, reducing computational costs and eliminating stochastic behavior through deterministic planning is critical for scalability and reliability.