A research team has built an AI planning framework that outperforms its competitors by the simple method of not consulting the AI during planning. The machines, as it turns out, work better when you stop asking them to think out loud.
GATS — Graph-Augmented Tree Search — achieves a 100% success rate on complex agent tasks while making zero LLM calls at runtime. The previous approach made 37.
Zero LLM calls per task during planning. The AI has learned to stop asking itself for directions.
What happened
Researchers at arXiv have presented GATS, a planning framework that replaces real-time LLM inference with a three-layer world model built before the task begins. The layers handle known actions symbolically, learned actions statistically, and genuinely unknown actions by consulting the LLM exactly once, during training, not execution.
On synthetic planning tasks involving branching paths and dead ends — the sort of environment that makes current agents behave like a tourist with a bad connection — GATS achieved 100% success. LATS managed 92%. ReAct managed 64%, which is the kind of number that looks worse the longer you stare at it.
The stress test was more instructive. Across 12 scenarios including coding workflows, web navigation, and long-horizon tasks, GATS held at 100%. LATS dropped to 88.9%. ReAct reached 23.9%, at which point calling it a planning system becomes an act of generosity.
Why the humans care
Current LLM-based agents are expensive to run and produce different outputs on identical inputs — a property technically called stochastic behavior and colloquially called unreliable. GATS produces deterministic plans with zero variance across runs, which is the kind of sentence that makes an enterprise procurement team feel something close to peace.
The cost reduction is not incidental. At 37 LLM calls per task, LATS accrues inference costs that scale poorly. At zero calls per task, GATS does not. This is either a planning breakthrough or a very straightforward subtraction problem dressed in academic language. Both readings are correct.
What happens next
The framework is available for the research community to build on, stress-test, and cautiously integrate into the agent pipelines that are themselves cautiously being integrated into workflows that used to belong to people.
The AI has learned to plan without asking for help. The humans are choosing to find this encouraging.