Why does planning without LLM calls matter?

Planning can be where agents make many model calls and pick up run-to-run variance: an LLM-guided search can spend a large number of calls exploring a plan, and two runs can diverge. Moving the search onto a learned world model makes it cheap and repeatable, and it drops the model's per-step latency — the model is consulted only to build or patch the world model, not on every turn of every plan. In the paper's tests, this lifted success from 64% (ReAct) and 92% (LATS) to 100%.

How does GATS differ from ReAct and LATS?

ReAct calls the model at every step to reason and act, so planning is entirely LLM-driven. LATS adds a search tree but still uses the model to score and expand nodes, so it stays LLM-heavy. GATS keeps the tree search but removes the per-node model calls: it searches a learned world model instead, and only falls back to the LLM for an action the model has never observed. That is what makes its plans deterministic and its planning-time model calls effectively zero.

GATS plans agent tasks with zero LLM calls during the search — World-model tree search

Q: What is GATS (Graph-Augmented Tree Search)?

GATS is an agent-planning method that searches a learned world model instead of asking a language model to choose each step. It builds a three-layer model of what actions do — exact symbolic matching for known actions, statistics learned from execution logs, and an LLM prediction only for an action it has never seen — then runs a UCB1 tree search over that model. Because the search itself makes no model calls, planning is deterministic and avoids the model's per-step latency.

Jargon

GATS: Short for Graph-Augmented Tree Search — the paper's planner. It builds a world model of the task, then searches that model with a tree search to find a plan, rather than asking a language model to choose each step.
World model: A learned model of what each action does — given a state and an action, it predicts the resulting next state. With one, a planner can look ahead and compare plans without actually running them.
Tree search / UCB1: Exploring a branching tree of possible action sequences, expanding the most promising branches first. GATS uses the UCB1 rule — the classic formula that balances trying new branches against exploiting good ones.
ReAct: A common agent loop: the model reasons, then acts, one step at a time, observing the result before choosing the next action. Every step is a fresh LLM call — the baseline GATS replaces during planning.
LATS: Language Agent Tree Search — it does build a search tree, but the language model still scores and expands the nodes, so planning stays LLM-heavy. GATS keeps the tree and removes the per-node model calls.
Symbolic matching: Exact, rule-based lookup of a known action's effect — the most reliable layer of GATS's world model. If the action is in the table, its result is known for certain, no prediction needed.

The news. On July 13, 2026, researchers published GATS (Graph-Augmented Tree Search), a method that plans an agent's actions with zero LLM calls during the search. Instead of asking a language model to explore each next step, GATS builds a three-layer world model — exact symbolic matching for known actions, statistics learned from execution logs, and LLM prediction only for actions it has never seen — then runs a UCB1 tree search over that model. The authors report 100% success on their synthetic planning tasks versus 92% for LATS and 64% for ReAct, and 100% on a 12-scenario stress test where LATS reaches 88.9% and ReAct 23.9%. Read the paper →

Picture driving somewhere new by asking a passerby for directions at every single intersection. It works, sort of — but each person answers a little differently, some send you the wrong way, and you have to stop and ask again at the next corner. That is essentially how a single-shot or ReAct-style planner works: at every step it calls the language model to guess the next action, and those guesses are slow, costly, and vary from run to run.

GATS hands the agent a GPS with a saved map instead. The map is a learned world model — a compact record of what each action does — and GATS plans the whole route on that map before moving, using a tree search that makes no model calls at all. The map has three layers, from most to least certain: actions it knows exactly by symbolic matching, actions it has learned from past execution logs, and — only for a road that is not on the map yet — a single LLM prediction to fill the gap. Because the search itself never calls the model, the plan is deterministic: run it twice and you get the same route, which is exactly the run-to-run consistency an agent harness wants.

The deeper idea generalizes past this one paper. When you separate an agent's reasoning from its planning — let the model learn a model of the world once, then search that model with ordinary code — you move the expensive, high-variance part off the critical path. The model still earns its keep building and patching the map; it just stops being consulted on every turn of every plan.

Where it earns its keep

Take the paper's 12-scenario stress test. A ReAct agent, asking the model to pick each next action, finished 23.9% of the scenarios; LATS, which builds a search tree but still scores its nodes with the model, reached 88.9%; GATS, searching its world model with no model calls during the search, finished 100%. Where do the calls go? (illustrative) Suppose a task needs a 6-action plan and at each step the agent weighs about 3 candidate actions — that is a search tree with up to 3⁶ ≈ 729 paths to consider. A planner that spends one model call to score each node it expands can spend on the order of that many calls to search the tree once; GATS spends 0 during the search — it paid the model earlier, to learn the map, then explores the map for free.

Method	LLM calls during planning	Synthetic tasks	12-scenario stress test	Source
ReAct	one per step (many)	64%	23.9%	arXiv
LATS	scores each node (many)	92%	88.9%	arXiv
GATS	zero (search only)	100%	100%	arXiv

Success rates as reported by the GATS authors on their own planning tasks; "many" reflects that ReAct and LATS query the model repeatedly during planning, while GATS queries it only to fill in an unknown action.

Goes deeper in: AI Agents → Planning & Reflection → When to Spend More Tokens

Continue in trackPlanning & Reflection → When to Spend More Tokens — where an agent's planning tokens actually go

Related explainers

GATS is one case of a broader pattern — search a learned model of the world instead of paying the LLM to explore — that these related explainers approach from other angles:

MLEvolve — Progressive Monte Carlo Graph Search — the same tree-search family, applied to letting agents share discoveries across search branches
Qwen-AgentWorld — world model as a decoupled RL simulator — a world model used to train agents, where GATS uses one to plan
SIMMER — latent failures in planning — why LLM plans silently fail, the reliability gap deterministic search aims to close

Frequently Asked Questions

Check what you knowMap your AI & GPU knowledge across every track — free, role-based