Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Summary

This paper introduces the Tree of Thoughts (ToT) framework, a novel approach to enhance the problem-solving capabilities of Large Language Models (LLMs). Unlike the standard approach of chain-of-thought prompting, which explores a linear sequence of reasoning, ToT allows the LLM to explore multiple reasoning paths simultaneously, creating a tree-like structure of thoughts. The framework involves three key steps: (1) thought generation, where the LLM proposes multiple distinct thought steps at each decision point; (2) thought evaluation, where each thought is evaluated based on its consistency, progress towards the solution, and other relevant criteria; and (3) thought selection, where the most promising thoughts are selected for further exploration. The paper demonstrates ToT's effectiveness on a variety of challenging reasoning tasks, including arithmetic word problems, creative writing, and game playing (e.g., solving the game of 24). Experimental results show that ToT consistently outperforms both standard LLM prompting techniques and chain-of-thought approaches. The core contribution is the ability to enable more strategic and deliberate problem-solving by LLMs by allowing them to maintain and explore a diverse set of potential solution paths, fostering exploration, self-correction and improved final results.

Key Takeaways

1. The Tree of Thoughts (ToT) framework enables LLMs to explore multiple reasoning paths simultaneously, leading to improved problem-solving.
2. ToT utilizes three key steps: thought generation (creating multiple thought steps), thought evaluation (assessing the quality of each thought), and thought selection (choosing the best thoughts for further exploration).
3. Experiments show that ToT outperforms standard and chain-of-thought LLM prompting on various reasoning tasks.
4. ToT facilitates more strategic and deliberate problem-solving, fostering exploration and self-correction in LLMs.