# Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models

**Authors:** Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, Yu-Xiong Wang  
**arXiv:** [2310.04406](https://arxiv.org/abs/2310.04406)  
**Date:** 2023-10-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.04406v1/

## Key points

- The paper introduces a framework called LATS (Language Agent Tree Search) to enhance decision-making capabilities of large language models (LLMs)
- LLMs have impressive performance in decision-making tasks but lack autonomy
- LATS leverages the strengths of LLMs in planning, acting, and reasoning
- Inspired by Monte Carlo tree search in model-based reinforcement learning
- LATS utilizes LLMs as agents, value functions, and optimizers for enhanced decision-making processes
- Incorporates an environment for external feedback to enable deliberate and adaptive problem-solving mechanism
- Experimental assessments conducted across diverse domains such as programming, HotPotQA, and WebShop
- Results demonstrate applicability of LATS for both reasoning and acting tasks
- Impressive scores achieved with GPT-4 for programming on HumanEval (94.4%) and with GPT-3.5 for web browsing on WebShop (75.9%)
- Highlights how LATS can unify reasoning, acting, and planning within language models to overcome limitations as autonomous agents
- Findings showcase effectiveness and generality of this approach across different domains

## Summary

The paper titled "Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models" introduces a framework called LATS (Language Agent Tree Search) that aims to enhance the decision-making capabilities of large language models (LLMs). LLMs have shown impressive performance in various decision-making tasks, but often lack autonomy. LATS addresses this limitation by leveraging the strengths of LLMs in planning, acting, and reasoning. Inspired by Monte Carlo tree search in model-based reinforcement learning, LATS utilizes LLMs as agents, value functions, and optimizers to repurpose their latent capabilities for enhanced decision-making processes. A key aspect of this framework is the incorporation of an environment for external feedback. This deliberate and adaptive problem-solving mechanism goes beyond existing techniques' limitations. Experimental assessments were conducted across diverse domains such as programming, HotPotQA, and WebShop to evaluate the effectiveness of LATS. The results demonstrate its applicability for both reasoning and acting tasks. For instance, when using GPT-4 for programming on HumanEval, LATS achieves an impressive score of 94.4%. Similarly, with GPT-3.5 for web browsing on WebShop, it attains an average score of 75.9%. Overall, the study highlights how LATS can unify reasoning, acting, and planning within language models to overcome their limitations as autonomous agents. The findings showcase the effectiveness and generality of this approach across different domains.