This paper addresses the challenge of self-correction in Large Language Models (LLMs) by focusing on mistake detection, particularly in mathematical reasoning. The authors introduce a novel prompting strategy called Pedagogical Chain-of-Thought (PedCoT), which is designed to guide LLMs in identifying reasoning mistakes. PedCoT is inspired by the Bloom Cognitive Model (BCM) and consists of pedagogical principles for prompt design, a two-stage interaction process, and grounded prompts. The method is evaluated on two public datasets of varying difficulty levels, demonstrating significant improvements over strong baselines. The results highlight the importance of domain knowledge in enhancing LLMs' reasoning abilities and provide a foundation for automatic math answer grading. The main contributions include the development of PedCoT, which bridges the gap between educational theory and prompt design for LLMs, and the demonstration of its effectiveness in finding mathematical reasoning mistakes.This paper addresses the challenge of self-correction in Large Language Models (LLMs) by focusing on mistake detection, particularly in mathematical reasoning. The authors introduce a novel prompting strategy called Pedagogical Chain-of-Thought (PedCoT), which is designed to guide LLMs in identifying reasoning mistakes. PedCoT is inspired by the Bloom Cognitive Model (BCM) and consists of pedagogical principles for prompt design, a two-stage interaction process, and grounded prompts. The method is evaluated on two public datasets of varying difficulty levels, demonstrating significant improvements over strong baselines. The results highlight the importance of domain knowledge in enhancing LLMs' reasoning abilities and provide a foundation for automatic math answer grading. The main contributions include the development of PedCoT, which bridges the gap between educational theory and prompt design for LLMs, and the demonstration of its effectiveness in finding mathematical reasoning mistakes.