Official Repo of Language Agent Tree Search (LATS)

Official implementation for paper Language Agent Tree Search Unifies Reasoning Acting and Planing in Language Models with code, prompts, model outputs.

More can be found at our project website or paper

Check out our demo, CodeLATS at our demo

Reasoning + Acting (HotPotQA)

Setup

To get started:

Clone this repo and move to the HotPotQA directory:

git clone https://github.com/andyz245/LanguageAgentTreeSearch && cd LanguageAgentTreeSearch/hotpot

Install the module dependencies into your environment:

pip install -r requirements.txt

Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Set the scripts and run paper experiments

sh lats.sh

--n_generate_sample: number of times to prompt during expansion/sampling
--n_evaluate_sample: number of times to prompt for state evaluation
--iterations: maximum number of trajectories to sample

Reasoning (Programming)

Setup

To get started:

Clone this repo and move to the HotPotQA directory:

git clone https://github.com/andyz245/LanguageAgentTreeSearch && cd LanguageAgentTreeSearch/programming

Install the module dependencies into your environment:

pip install -r requirements.txt

Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Set the scripts and run paper experiments

sh run_lats.sh

Code adapted from https://github.com/noahshinn024/reflexion/tree/main

Decision-making (WebShop)

Setup

To get started:

Clone this repo and move to the WebShop directory:

git clone https://github.com/andyz245/LanguageAgentTreeSearch && cd LanguageAgentTreeSearch/webshop

Install WebShop from source and run environment instance locally. Follow the instructions here (https://github.com/princeton-nlp/WebShop)
Install the module dependencies into your environment:

pip install -r requirements.txt

Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Change localhost in lats.py to your local port running WebShop
Set the scripts and run paper experiments

sh lats.sh

--n_generate_sample: number of times to prompt during expansion/sampling
--n_evaluate_sample: number of times to prompt for state evaluation
--iterations: maximum number of trajectories to sample

Trajectories

programming/root/ contains all the trajectories from the paper's experiments on programming. Please use get_acc.py with the log path to get the actual accuracy. HotPotQA and WebShop logs were too large to upload, feel free to email if interested.

Citations

Please cite the paper and star this repo if you use LATS and find it interesting. Feel free to contact andyz3@illinois.edu or open an issue if you have any questions.

@misc{zhou2023language,
      title={Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models}, 
      author={Andy Zhou and Kai Yan and Michal Shlapentokh-Rothman and Haohan Wang and Yu-Xiong Wang},
      year={2023},
      eprint={2310.04406},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}