Tree of Thoughts Prompting

Prompt engineering series on techniques eliciting reasoning capabilities of LLMs

Original Paper: Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Abstract

• The Tree of Thoughts framework enhances language models' problem-solving abilities by enabling deliberate decision-making through exploration over coherent units of text.

• It allows language models to consider multiple reasoning paths, self-evaluate choices, and make global decisions, significantly improving performance on tasks requiring planning or search.

Practical Implications

• The Tree of Thoughts framework enhances language models' problem-solving abilities on tasks requiring planning or search, such as Game of 24 and Creative Writing.

• It empowers language models to make more autonomous and intelligent decisions, improving interpretability and human alignment in model decisions.

• The framework could be beneficial for real-world decision-making applications like coding, data analysis, and robotics, opening up opportunities for further research in these areas

Methodology

• The Tree of Thoughts framework enables language models to perform deliberate decision-making by considering multiple reasoning paths and self-evaluating choices.

• The framework allows exploration over coherent units of text, serving as intermediate steps toward problem-solving.

• The approach can be considered a modern rendition of classical search methods for problem-solving, incorporating heuristics and value feedback.

• The Tree of Thoughts framework augments the associative 'System 1' of language models with a 'System 2' based on searching a tree of possible paths to the solution

Limitations

• The Tree of Thoughts framework may not be necessary for tasks where existing language models like GPT-4 already excel.

• The study only explores relatively simple tasks, leaving room for further investigation into more complex applications.

• Implementing the Tree of Thoughts method may require more resources compared to sampling methods, potentially impacting cost and efficiency.

Conclusion

• The Tree of Thoughts framework enhances language models' problem-solving abilities on tasks requiring planning or search.

• The framework empowers language models to make decisions autonomously and intelligently.

• The intersection of language models with classical AI approaches is seen as an exciting direction.

How tree of thoughts framework is different from chain-of-thought prompting?

• Tree of Thoughts allows for exploration over coherent units of text, serving as intermediate steps towards problem-solving.

• Chain-of-Thought prompting involves a chain of coherent language sequences bridging input and output, sampled sequentially.

• Tree of Thoughts enables deliberate decision-making by considering multiple reasoning paths and self-evaluating choices, allowing for global choices.

• Chain-of-Thought prompting samples thoughts coherently without explicit decomposition, leaving the decomposition of thoughts ambiguous.

How tree of thoughts framework is different from self-consistency prompting?

• Tree of Thoughts allows language models to explore multiple reasoning paths over thoughts in a tree structure.

• Self-Consistency Prompting samples chains of coherent thoughts and returns the most frequent output, exploring a richer set of thoughts.

Paper Infographic