Pyramid of Thought: A Novel Approach for Enhancing Chain-of-Thought Reasoning in Large Language Models

Abstract

Chain-of-Thought (CoT) prompting shows promise in unleashing the latent reasoning abilities of large language models (LLMs). While CoT prompts guide LLMs through a step-wise reasoning chain, its traditional format encounters challenges with truly complex problems. In response, we introduce the Pyramid of Thought (PoT), a pioneering prompting strategy inspired by the Pyramid Principle. Unlike CoT's linear chain, Our PoT establishes a layered structure of 'facts'—mutually exclusive or exhaustive subsets relevant to the problem. This hierarchical architecture guides LLMs through multiple reasoning steps, offering focused clarity to tackle complexity. Through extensive experiments, PoT demonstrates significant accuracy improvement of 6% compared to CoT.