| name | graph-augmented-tree-search-agent-planning |
| description | GATS - eliminate LLM calls during agent planning by combining UCB1 tree search with a layered world model (exact symbolic match / learned statistics / LLM-for-unknown). Deterministic, zero-variance plans, 100% success on stress tests vs LATS/ReAct. Use when building LLM-agent planners that are too slow/costly/stochastic from in-loop LLM inference. |
GATS — Graph-Augmented Tree Search (from arXiv:2607.08894)
Planning framework that gets LLM-class planning success without calling the LLM during
inference. Combines systematic UCB1 tree search with a three-layer world model.
When to use
- LLM-agent planning (LATS, ReAct) is too slow / expensive / non-deterministic in production.
- You need deterministic, reproducible plans (zero variance across runs).
- You have (or can build) an environment model or execution logs to learn from.
Core architecture
Replace in-loop LLM next-action prediction with:
- UCB1 tree search over the action space (systematic exploration, not sampling).
- A layered world model that predicts outcomes WITHOUT an LLM call:
- L1 — Exact symbolic action matching: known actions → deterministic lookup (fast path).
- L2 — Statistics learned from execution logs: empirical success/transition distributions.
- L3 — LLM-based prediction: used ONLY for genuinely unknown actions (cold start).
- LLM is invoked at most during L3 (rare once L2 is populated); after warm-up, inference uses zero
LLM calls.
Implementation steps
- Define the action space and a simulator / replayable execution log.
- Build L1 as a deterministic rule/lookup table for actions with known effects.
- Populate L2: mine execution logs → (state, action) → outcome frequency table; smooth with
Laplace / Bayesian prior.
- L3: route to an LLM only when L1+L2 have no confident prediction; cache the result back into L2.
- Run UCB1 over the tree:
UCT = Q(a) + c·√(ln N_parent / N_a), expand using the world model, prune
dead-ends (branching paths that L2/L3 mark as failing).
- Return the highest-value path as the plan.
Results (paper)
- 100% success vs LATS 92% / ReAct 64% on synthetic; 100% vs LATS 88.9% / ReAct 23.9% on 12-scenario
stress test (coding, web nav, long-horizon).
- Zero LLM calls per task during planning (vs 37 for LATS); zero variance across runs.
Pitfalls
- World-model quality dominates; bad L2 statistics → tree search optimizes the wrong thing. Validate
L2 coverage before trusting plans.
- Cold-start (empty L2) degrades to L3 cost — seed L2 from any available logs first.
- UCB1 exploration constant
c needs tuning; too high → wasted expansions, too low → premature
convergence to a suboptimal deterministic plan.
- Environment non-stationarity: re-mine L2 periodically or plans rot.
Verification
- Compare success rate, LLM-calls/task, and run-to-run variance vs LATS/ReAct on the same tasks.
- Report plan determinism (std dev across N runs = 0 target).