This paper explores the intrinsic self-correction capabilities of Large Language Models (LLMs), focusing on the role of "confidence" in this process. The authors identify that LLMs can understand and assess their confidence levels, which is crucial for effective self-correction. They propose an "If-or-Else" (IoE) prompting framework to guide LLMs in evaluating their confidence and refining responses accordingly. Extensive experiments on four LLMs across six benchmark datasets demonstrate that the IoE-based prompt consistently improves the accuracy of self-corrected responses compared to traditional methods. The study highlights the importance of considering LLMs' confidence in self-correction and provides a practical framework to enhance this capability. The research also discusses limitations and potential risks, emphasizing the need for further exploration and caution in deploying such techniques.This paper explores the intrinsic self-correction capabilities of Large Language Models (LLMs), focusing on the role of "confidence" in this process. The authors identify that LLMs can understand and assess their confidence levels, which is crucial for effective self-correction. They propose an "If-or-Else" (IoE) prompting framework to guide LLMs in evaluating their confidence and refining responses accordingly. Extensive experiments on four LLMs across six benchmark datasets demonstrate that the IoE-based prompt consistently improves the accuracy of self-corrected responses compared to traditional methods. The study highlights the importance of considering LLMs' confidence in self-correction and provides a practical framework to enhance this capability. The research also discusses limitations and potential risks, emphasizing the need for further exploration and caution in deploying such techniques.