Self-Hint Prompting Improves Zero-shot Reasoning in Large Language Models via Reflective Cycle
Abstract
Chain-of-Thought (CoT) has brought a fresh perspective to improve the reasoning ability of large language models (LLMs). To relieve the burden of manual design in CoT, Zero-shot CoT has pioneered a direct interaction with LLMs. Based on it, researchers attempt to optimize reasoning paths through various prompting approaches like reflection, selection, and planning. However, few studies have focused on the possibility of combining all these strategies through a cognitive theory. Inspired by experiential learning, this paper proposes a new zero-shot prompting method based on Kolb's reflective cycle, named Self-Hint prompting. Specifically, Self-Hint prompting introduces an automated iterative interaction approach to simulate the conscious reflection process, which uses intermediate observations as hints to guide LLMs. We have conducted comprehensive experiments on various math reasoning benchmarks. The empirical results on GPT models demonstrate the effectiveness of our method. Proposed Self-Hint prompting consistently outperforms other zero-shot baselines.