This repository provides the official implementation for the ICML 2024 paper "Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models." It offers a unified framework for language agents to perform reasoning, acting, and planning, targeting researchers and developers working with large language models for complex tasks. The core benefit is a novel search algorithm that integrates these capabilities, improving performance on benchmarks like HotPotQA and WebShop.

How It Works

LATS employs a tree search mechanism to explore possible actions and reasoning steps. It unifies reasoning, acting, and planning by treating them as nodes and branches within a search tree. This approach allows agents to systematically explore a problem space, evaluate potential trajectories, and select optimal actions, effectively combining deliberation with execution. The method is advantageous for its ability to handle complex, multi-step tasks that require both understanding and interaction.

Quick Start & Requirements

• HotPotQA & Programming:
  • Install: pip install -r requirements.txt
  • Prerequisites: OpenAI API key (export OPENAI_API_KEY=<your key>)
  • Run: sh lats.sh (HotPotQA) or sh run_lats.sh (Programming)
  • Docs: Project Website, Paper
• WebShop:
  • Install: Clone WebShop repo, install from source, then pip install -r requirements.txt
  • Prerequisites: OpenAI API key, local WebShop environment running on localhost.
  • Run: sh lats.sh (after modifying lats.py with your WebShop port)

Highlighted Details

• Official implementation for ICML 2024 paper "Language Agent Tree Search".
• Demonstrates unified reasoning, acting, and planning for LLM agents.
• Includes code, prompts, and model outputs for experiments.
• Offers integrations with LangChain (LangGraph) and LlamaIndex.

Maintenance & Community

The project is associated with Andy Zhou and the paper's authors. Contact is available via email (andyz3@illinois.edu) or GitHub issues.

Licensing & Compatibility

The repository does not explicitly state a license in the README. Users should verify licensing for commercial use or integration into closed-source projects.

Limitations & Caveats

Log files for HotPotQA and WebShop experiments are too large to be included. The README mentions code adapted from reflexions/tree/main, suggesting potential dependencies or licensing considerations from that source.