# Large Language Model Guided Tree-of-Thought

**Authors:** Jieyi Long  
**arXiv:** [2305.08291](https://arxiv.org/abs/2305.08291)  
**Date:** 2023-05-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.08291v1/

## Key points

- Auto-regressive large language models (LLMs) have potential in solving complex reasoning tasks
- LLMs struggle with certain types of problems that require trial and error problem-solving techniques
- Jieyi Long introduces the Tree-of-Thought (ToT) framework to address this issue
- ToT technique is inspired by how the human mind approaches complex reasoning tasks through a tree-like thought process that allows for backtracking when necessary
- The ToT framework aims to improve the problem-solving capabilities of LLMs by augmenting them with additional modules such as a prompter agent, a checker module, a memory module, and a ToT controller
- These modules engage in a multi-round conversation with the LLM to solve a given problem using the ToT framework
- The memory module records the conversation and state history of the problem-solving process, which allows the system to backtrack to previous steps of the thought process and explore other directions from there
- Long implemented a ToT-based solver for Sudoku puzzles and experimental results showed that it significantly increased success rate of Sudoku puzzle solving
- Long's paper presents an innovative approach to improving LLMs' problem-solving capabilities by incorporating human-like trial-and-error techniques into their decision making processes.

## Summary

In the field of artificial intelligence, auto-regressive large language models (LLMs) have shown great potential in solving complex reasoning tasks. However, these models still struggle with certain types of problems that require trial and error problem-solving techniques. To address this issue, Jieyi Long introduces the Tree-of-Thought (ToT) framework in their paper "Large Language Model Guided Tree-of-Thought". The ToT technique is inspired by how the human mind approaches complex reasoning tasks through a tree-like thought process that allows for backtracking when necessary. The ToT framework aims to improve the problem-solving capabilities of LLMs by augmenting them with additional modules such as a prompter agent, a checker module, a memory module, and a ToT controller. To solve a given problem using the ToT framework, these modules engage in a multi-round conversation with the LLM. The memory module records the conversation and state history of the problem-solving process, which allows the system to backtrack to previous steps of the thought process and explore other directions from there. To verify the effectiveness of this proposed technique, Long implemented a ToT-based solver for Sudoku puzzles. Experimental results showed that the ToT framework significantly increased the success rate of Sudoku puzzle solving. Overall, Long's paper presents an innovative approach to improving LLMs' problem-solving capabilities by incorporating human-like trial-and-error techniques into their decision making processes. This could have significant implications for various fields that rely on AI powered systems to solve complex problems.