Tree-of-Thought: Improving GPT-4 Reasoning by 900%

Here is a summary of the key points in the PDF:

Large language models are increasingly being used for general problem solving, but they are still limited by their token-level, left-to-right generation process.

The authors propose a new Tree of Thoughts (ToT) framework to address this. ToT frames problem solving as search through a tree of "thoughts", where each thought is an intermediate step towards the solution.

This allows the language model to:

- Generate and explore multiple potential thoughts at each step

- Evaluate the progress of different thoughts using self-evaluation prompts

- Perform lookahead and backtracking to make global decisions

The authors propose 3 tasks to test ToT: Game of 24, Creative Writing and Crosswords.

ToT significantly outperforms standard input-output and chain-of-thought prompting baselines on all 3 tasks. This shows the benefits of ToT's ability to explore, evaluate and search through different reasoning paths.

The key benefits of ToT are:

- Generality: It generalizes existing prompting methods

- Modularity: The components can be varied independently

- Adaptability: It can accommodate different problem properties and resource constraints

- Convenience: It only requires a pre-trained language model

The ToT framework shows how deliberate search through a tree of "thoughts" can help large language models solve problems that require planning and search, beyond their standard left-to-right generation process.

Code to be (isn't yet) released at: https://github.com/ysymyth/tree-of-thought-llm

// via: https://kagi.com/summarizer/index.html?url=https%3A%2F%2Farx...