LATS: Language Agent Tree Search¶
UCT-style search with complete rollouts, environment reward and reflection.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | University of Illinois Urbana-Champaign. |
| First public release | 2023-10-06 |
| Author code | lapisrocks/LanguageAgentTreeSearch. |
| Local adapter or method key | lats |
| Local reproduction code | src/auto_research/agent_research/ |
Paper method and evidence¶
Use an LM as agent, value function and optimizer within MCTS. External execution supplies rewards; reflections on failed traces guide later exploration and backtracking.
flowchart LR
A[UCT selection] --> B[Roll out complete plans]
B[Roll out complete plans] --> C[Score and reflect]
C[Score and reflect] --> D[Backpropagate value or backtrack]Original paper figure¶
Figure shows search and reflection. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
The latest paper reports GPT-4 HumanEval pass@1 92.7% and GPT-3.5 WebShop score 75.9. Also tests coding, interactive QA, navigation and math; no production A/B.
Local reproduction¶
Four full plans per PlanBench episode; exact-plan rewards and reflections/backtracking on failures, then select by cumulative value. Joint success 1.0000, mean cost 4.0000, rollouts/reflections/backtracks 480/360/360. Metrics. Leaves are complete plans, not LM value calls. No HumanEval executor or WebShop; local search is much smaller.