GATS achieves 100% success with zero LLM calls during planning

GATS (Graph-Augmented Tree Search) is a novel planning framework for LLM agents that eliminates the computational and stochastic overhead of traditional LLM-guided search. Developed by Maureese Williams and Dymitr Nowicki, GATS combines systematic UCB1-based tree search with a three-layer world model. Layer 1 uses exact symbolic action matching for known operations, Layer 2 applies statistics learned from execution logs to predict outcomes, and Layer 3 falls back to an LLM for unknown actions. This design ensures that during the planning phase, the LLM is never called—zero inference calls per task—making the process deterministic and dramatically cheaper.

In benchmark tests, GATS achieved a 100% success rate on synthetic planning tasks with branching paths and dead-ends, versus 92% for LATS and 64% for ReAct. On a stress test of 12 challenging scenarios—including coding workflows, web navigation, and long-horizon tasks—GATS again hit 100%, while LATS dropped to 88.9% and ReAct to 23.9%. Crucially, GATS required 0 LLM calls per task during planning, compared to 37 for LATS, and produced plans with zero variance across runs. This work suggests that systematic search with learned world models can outperform LLM-guided exploration for multi-step agent planning, opening the door to more reliable and cost-effective AI agents.