Confidence Matters: Revisiting Intrinsic Self-Correction Capabilities of Large Language Models This paper investigates the intrinsic self-correction capabilities of large language models (LLMs), focusing on the role of "confidence" in the self-correction process. The study identifies that LLMs possess the ability to assess their own confidence in responses, which is crucial for effective self-correction. The authors propose an "If-or-Else" (IoE) prompting framework to guide LLMs in evaluating their confidence, enabling intrinsic self-corrections. The IoE-based Prompt is designed to allow LLMs to retain answers with high confidence while revising answers with low confidence. The framework is evaluated on multiple benchmarks, demonstrating consistent improvements in accuracy compared to existing methods. The study highlights that LLMs can assess their own confidence in responses, which is essential for self-correction. The IoE-based Prompt is shown to be effective in enhancing self-correction capabilities, particularly in mathematical reasoning and other tasks. The framework is compared with the Critical Prompt, which directly instructs LLMs to find and correct errors. The IoE-based Prompt is found to be more effective, as it reduces the risk of over-criticism and improves the reliability of self-correction. The paper also explores the impact of confidence on self-correction, showing that LLMs can improve their self-correction capabilities by incorporating confidence assessments. The IoE-based Prompt is shown to be effective in both deterministic and open tasks, with the ability to assess confidence levels and guide self-correction. The framework is evaluated on multiple benchmarks, including GSM8K, SVAMP, HotpotQA, Sports, LLC, and Domestic Robot, demonstrating consistent improvements in accuracy. The study also investigates the effectiveness of the IoE-based Prompt in multi-modal reasoning tasks, showing that it outperforms existing methods. The framework is shown to be effective in both language-based and multi-modal reasoning tasks, with the ability to assess confidence levels and guide self-correction. The paper also explores the integration of the IoE-based Prompt with existing prompting techniques, such as Chain-of-Thought (CoT) and Rephrase-and-Respond (RaR), showing that it can be effectively combined with these methods. The study concludes that the IoE-based Prompt is an effective framework for enhancing the self-correction capabilities of LLMs, with the ability to assess confidence levels and guide self-correction. The framework is shown to be effective in both deterministic and open tasks, with the ability to assess confidence levels and guide self-correction. The paper also highlights the importance of confidence in the self-correction process, showing that LLMs can improve their self-correction capabilities by incorporating confidence assessments. The study also discusses the limitations and potential risks of the framework, including the need for further research and the potential for misuse.Confidence Matters: Revisiting Intrinsic Self-Correction Capabilities of Large Language Models This paper investigates the intrinsic self-correction capabilities of large language models (LLMs), focusing on the role of "confidence" in the self-correction process. The study identifies that LLMs possess the ability to assess their own confidence in responses, which is crucial for effective self-correction. The authors propose an "If-or-Else" (IoE) prompting framework to guide LLMs in evaluating their confidence, enabling intrinsic self-corrections. The IoE-based Prompt is designed to allow LLMs to retain answers with high confidence while revising answers with low confidence. The framework is evaluated on multiple benchmarks, demonstrating consistent improvements in accuracy compared to existing methods. The study highlights that LLMs can assess their own confidence in responses, which is essential for self-correction. The IoE-based Prompt is shown to be effective in enhancing self-correction capabilities, particularly in mathematical reasoning and other tasks. The framework is compared with the Critical Prompt, which directly instructs LLMs to find and correct errors. The IoE-based Prompt is found to be more effective, as it reduces the risk of over-criticism and improves the reliability of self-correction. The paper also explores the impact of confidence on self-correction, showing that LLMs can improve their self-correction capabilities by incorporating confidence assessments. The IoE-based Prompt is shown to be effective in both deterministic and open tasks, with the ability to assess confidence levels and guide self-correction. The framework is evaluated on multiple benchmarks, including GSM8K, SVAMP, HotpotQA, Sports, LLC, and Domestic Robot, demonstrating consistent improvements in accuracy. The study also investigates the effectiveness of the IoE-based Prompt in multi-modal reasoning tasks, showing that it outperforms existing methods. The framework is shown to be effective in both language-based and multi-modal reasoning tasks, with the ability to assess confidence levels and guide self-correction. The paper also explores the integration of the IoE-based Prompt with existing prompting techniques, such as Chain-of-Thought (CoT) and Rephrase-and-Respond (RaR), showing that it can be effectively combined with these methods. The study concludes that the IoE-based Prompt is an effective framework for enhancing the self-correction capabilities of LLMs, with the ability to assess confidence levels and guide self-correction. The framework is shown to be effective in both deterministic and open tasks, with the ability to assess confidence levels and guide self-correction. The paper also highlights the importance of confidence in the self-correction process, showing that LLMs can improve their self-correction capabilities by incorporating confidence assessments. The study also discusses the limitations and potential risks of the framework, including the need for further research and the potential for misuse.