Verified project record

princeton-nlp/tree-of-thought-llm

tree-of-thought-llm is a Python framework that enables large language models to explore multiple reasoning paths and evaluate intermediate states using tree search algorithms. It frames problem-solving as a search over a tree of thoughts rather than a single sequential prompt, targeting complex, multi-step reasoning tasks.

Project overview

By framing problem-solving as a deliberate search over a tree of thoughts rather than a single sequential prompt, tree-of-thought-llm addresses the lack of exploration in standard LLM prompting, offering a structured approach to improve multi-step reasoning outcomes.

Project type: AI Agent
Use cases: Search & Research
Deployment: Refer to project documentation
License: MIT

Best for

Researchers and developers who need a framework that enables LLMs to perform deliberate exploration and tree search for complex, multi-step reasoning tasks.

Key capabilities

Implements the ToT search using Breadth-First Search to deliberate over reasoning steps.
Implements the ToT search using Depth-First Search, used for crosswords experiments.
Runs standard Chain-of-Thought or input-output sampling for baseline comparison.
Allows users to easily add new tasks by creating task classes, data files, and prompts.

Limitations and risks

Outputs are not deterministic due to randomness in LLM decoding, which can affect success rates.
Requires an external LLM API such as GPT-4, which may incur paid service costs and introduce latency from multiple LLM calls.

Getting started

Install the package from PyPI using pip install tree-of-thoughts-llm, set the OPENAI_API_KEY environment variable, and use a Python script to call the solve function.

Evidence and sources

GitHub project description: [NeurIPS 2023] Tree of Thoughts: Deliberate Problem Solving with Large Language Models
README: Install from PyPI
README: Official implementation for paper Tree of Thoughts: Deliberate Problem Solving with Large Language Models
README: The very simple ``run.py`` implements the ToT + BFS algorithm, as well as the naive IO/CoT sampling.
README: The following minimal script will attempt to solve the game of 24 with 4 5 6 10 (might be a bit slow as it's using GPT-4):

Reviewed at: 2026-07-15T17:15:26Z