level: technical
large language model agents often struggle with multi-step planning due to high inference costs and unpredictable behavior. existing methods like lats and react depend on many llm calls during planning. gats, or graph-augmented tree search, replaces those calls with a three-layer world model. the model combines exact symbolic matching, learned statistics from past runs, and llm predictions only for unknown actions. this design keeps planning fast and stable.
gats uses ucb1-based tree search to explore possible action sequences. the first layer catches actions seen before with perfect accuracy. the second layer uses execution logs to estimate outcomes for familiar but variable actions. the third layer calls an llm only when an action is completely new. this layered approach avoids repeated llm queries, cutting cost and noise. on synthetic tasks with branching paths and dead ends, gats reached a 100% success rate, beating lats at 92% and react at 64%.
the framework was also tested on 12 harder scenarios, including coding workflows and web navigation. gats maintained high success while using far fewer llm calls. by shifting most prediction work to cheaper, deterministic layers, it makes agent planning practical for real-world use. the method shows that combining classic search with learned world models can outperform pure llm-driven strategies.
why it matters: it reduces llm inference costs and improves reliability in ai agents that plan and act in complex environments.