Repository files navigation

Official implementation for paperTree of Thoughts: Deliberate Problem Solving with Large Language Models with code, prompts, model outputs.Also checkits tweet thread in 1min.

1. Set up OpenAI API key and store in environment variableOPENAI_API_KEY (seehere).

2. Installtot package in two ways:

• Option 1: Install from PyPI

pip install tree-of-thoughts-llm

• Option 2: Install from source

git clone https://github.com/princeton-nlp/tree-of-thought-llmcd tree-of-thought-llmpip install -r requirements.txtpip install -e.# install `tot` package

The following minimal script will attempt to solve the game of 24 with4 5 6 10 (might be a bit slow as it's using GPT-4):

importargparsefromtot.methods.bfsimportsolvefromtot.tasks.game24importGame24Taskargs=argparse.Namespace(backend='gpt-4',temperature=0.7,task='game24',naive_run=False,prompt_sample=None,method_generate='propose',method_evaluate='value',method_select='greedy',n_generate_sample=1,n_evaluate_sample=3,n_select_sample=5)task=Game24Task()ys,infos=solve(args,task,900)print(ys[0])

And the output would be something like (note it's not deterministic, and sometimes the output can be wrong):

10 - 4 = 6 (left: 5 6 6)5 * 6 = 30 (left: 6 30)30 - 6 = 24 (left: 24)Answer: (5 * (10 - 4)) - 6 = 24

Run experiments viash scripts/{game24, text, crosswords}/{standard_sampling, cot_sampling, bfs}.sh, except in crosswords we use a DFS algorithm for ToT, which can be run viascripts/crosswords/search_crosswords-dfs.ipynb.

The very simplerun.py implements the ToT + BFS algorithm, as well as the naive IO/CoT sampling. Some key arguments:

• --naive_run: if True, run naive IO/CoT sampling instead of ToT + BFS.
• --prompt_sample (choices=[standard,cot]): sampling prompt
• --method_generate (choices=[sample,propose]): thought generator, whether to sample independent thoughts (used in Creative Writing) or propose sequential thoughts (used in Game of 24)
• --method_evaluate (choices=[value,vote]): state evaluator, whether to use the value states independently (used in Game of 24) or vote on states together (used in Creative Writing)
• --n_generate_sample: number of times to prompt for thought generation
• --n_evaluate_sample: number of times to prompt for state evaluation
• --n_select_sample: number of states to keep from each step (i.e.b in the paper's ToT + BFS algorithm)

logs/ contains all the trajectories from the paper's experiments, except forlogs/game24/gpt-4_0.7_propose1_value3_greedy5_start900_end1000.json which was reproduced after the paper (as the original experiment was done in a notebook) and achieved a 69% score instead of the original 74% score due to randomness in GPT decoding. We hope to aggregate multiple runs in the future to account for sampling randomness and update the paper, but this shouldn't affect the main conclusions of the paper.

Setting up a new task is easy, and mainly involves two steps.

• Set up a new task class intot/tasks/ and task files intot/data/. Seetot/tasks/game24.py for an example. Add the task totot/tasks/__init__.py.
• Set up task-specific prompts intot/prompts/. Seetot/prompts/game24.py for an example. Depending on the nature of the task, choose--method_generate (choices=[sample,propose]) and--method_evaluate (choices=[value,vote]) and their corresponding prompts.

Please cite the paper and star this repo if you use ToT and find it interesting/useful, thanks! Feel free to contactshunyuyao.cs@gmail.com or open an issue if you have any questions.

@misc{yao2023tree,title={{Tree of Thoughts}: Deliberate Problem Solving with Large Language Models},author={Shunyu Yao and Dian Yu and Jeffrey Zhao and Izhak Shafran and Thomas L. Griffiths and Yuan Cao and Karthik Narasimhan},year={2023},eprint={2305.10601},archivePrefix={arXiv},primaryClass={cs.CL}}