Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Large language models (LLMs) are becoming increasingly powerful, capable of performing a wide range of tasks. However, they still struggle with tasks that require exploration, strategic planning, or where initial decisions play a critical role. To address this challenge, researchers have introduced a new framework called Tree of Thoughts (ToT).

How ToT Works

ToT allows LLMs to perform deliberate decision-making by considering multiple reasoning paths and evaluating choices before selecting the next course of action. It also enables them to look ahead or backtrack when necessary to make informed decisions.

Here's a breakdown of how ToT works:

• Tree Structure: ToT represents the problem-solving process as a tree. Each node in the tree represents a partial solution, and the branches represent different paths that can be taken to reach the final solution.
• Thought Decomposition: The problem-solving process is broken down into smaller steps called "thoughts." A thought can be a sentence, a paragraph, or even a single word, depending on the complexity of the problem.
• Thought Generation: The LLM generates different potential thoughts for each step in the problem-solving process. These thoughts can be generated using a variety of techniques, such as sampling from a pre-trained language model or using a rule-based system.
• Thought Evaluation: The LLM evaluates each potential thought to determine how likely it is to lead to a solution. This evaluation can be based on a variety of factors, such as the coherence of the thought, its consistency with the problem constraints, and its similarity to known solutions.
• Search Algorithm: The LLM uses a search algorithm to explore the tree of thoughts and select the most promising path. Different search algorithms can be used depending on the nature of the problem.

Benefits of ToT

• Generality: ToT can be applied to a wide range of problems that require reasoning and planning.
• Modularity: The different components of ToT (thought decomposition, generation, evaluation, and search) can be easily adapted to different problems.
• Adaptability: ToT can be adapted to different LLM capabilities and resource constraints.
• Convenience: ToT does not require any additional training of the LLM.

Experiments and Results

Researchers have evaluated ToT on three novel tasks that challenge existing LLM inference methods:

• Game of 24: The goal of this game is to use four numbers to create the number 24 using basic arithmetic operations.
• Creative Writing: The goal of this task is to generate a creative story based on a prompt.
• Mini Crosswords: The goal of this task is to solve a mini crossword puzzle.

ToT achieved superior results on all three tasks compared to existing LLM inference methods. These results demonstrate the potential of ToT for improving the problem-solving capabilities of LLMs.

Conclusion

Tree of Thoughts is a promising new framework that can help LLMs solve problems that require reasoning and planning. ToT is general, modular, and adaptable, making it a valuable tool for a wide range of applications. As LLMs continue to develop, ToT is likely to play an increasingly important role in helping them solve complex problems.