This repository demonstrates a novel "Tree-of-Thought" (ToT) prompting technique designed to enhance the reasoning capabilities of Large Language Models (LLMs) like ChatGPT. It targets users seeking to improve LLM accuracy on complex reasoning tasks, offering a single-prompt solution that outperforms traditional Chain-of-Thought (CoT) methods by enabling self-correction and knowledge accumulation.

How It Works

The ToT prompting approach simulates a group of experts independently reasoning through a problem step-by-step. Each "expert" shares its thought process, allowing for self-correction and consensus-building. This method encourages the LLM to explore multiple reasoning paths and identify errors, leading to more accurate outcomes compared to single-path CoT prompting.

Quick Start & Requirements

• Install: No specific installation required; the technique is applied via prompt engineering.
• Prerequisites: Access to an LLM like ChatGPT (GPT-3.5 or GPT-4).
• Resources: Minimal; requires only API access or a chat interface to the LLM.
• Docs: tree-of-thought-prompting GitHub

Highlighted Details

• Improves ChatGPT 3.5's reasoning to match ChatGPT 4 on specific complex problems.
• Emulates organizational decision-making structures for LLM reasoning.
• Explores concepts like collaboration, X-as-a-Service, and hierarchy for LLM interactions.
• Provides a single-prompt alternative to multi-call ToT frameworks.

Maintenance & Community

• The project welcomes contributions for prompt refinement and new examples.
• Maintained by Dave Hulbert.

Licensing & Compatibility

• Licensed under MIT.
• Compatible with commercial use and closed-source applications.

Limitations & Caveats

The effectiveness is demonstrated on a limited set of complex reasoning problems, and the prompt may require further refinement for optimal performance across diverse tasks. The repository is primarily a proof-of-concept for a prompting technique rather than a software library.