Demystifying Chains, Trees, and Graphs of Thoughts.

Area of Science:

• Artificial Intelligence
• Natural Language Processing

Background:

• Large language models (LLMs) show significant progress, with prompt engineering enhancing their capabilities.
• Structure-enhanced prompting, including Chain-of-Thought and Graph of Thoughts, guides LLM reasoning for improved task performance.

Purpose of the Study:

• To devise a general blueprint for effective and efficient LLM reasoning schemes.
• To establish the first taxonomy of structure-enhanced LLM reasoning, analyzing structures, representations, and algorithms.

Main Methods:

• In-depth analysis of the prompt execution pipeline.
• Development of a taxonomy for structure-enhanced LLM reasoning schemes, termed 'reasoning topologies'.
• Comparison of existing prompting schemes using the proposed taxonomy.

Main Results:

• A general blueprint for LLM reasoning schemes is proposed.
• The first taxonomy of structure-enhanced LLM reasoning is established, detailing reasoning topologies.
• Analysis reveals how design choices impact performance and cost in prompting schemes.

Conclusions:

• The study provides a framework for understanding and advancing structure-enhanced LLM reasoning.
• This work clarifies concepts, categorizes existing methods, and identifies future research challenges in prompt engineering.
• The proposed taxonomy facilitates the development of more effective and efficient LLM reasoning techniques.