The paper introduces Intrinsic Self-Correction (ISC) in generative language models (LMs), aiming to enable small LMs to correct their initial outputs automatically. Unlike previous methods that rely on complex prompts and verification processes, ISC is designed to be intrinsic and spontaneous, allowing LMs to self-verify and self-modify their answers. The authors propose Partial Answer Masking (PAM) to enhance the model's ability to self-verify, and conduct experiments on LMs with 6 billion to 13 billion parameters using two tasks: commonsense reasoning and factual knowledge reasoning. The results show that ISC significantly improves the accuracy of the generated outputs, demonstrating the effectiveness of the proposed method in enhancing the self-correction capabilities of even small LMs. The paper also discusses the limitations and future directions, emphasizing the importance of domain knowledge and multiple iterations for more accurate self-correction.The paper introduces Intrinsic Self-Correction (ISC) in generative language models (LMs), aiming to enable small LMs to correct their initial outputs automatically. Unlike previous methods that rely on complex prompts and verification processes, ISC is designed to be intrinsic and spontaneous, allowing LMs to self-verify and self-modify their answers. The authors propose Partial Answer Masking (PAM) to enhance the model's ability to self-verify, and conduct experiments on LMs with 6 billion to 13 billion parameters using two tasks: commonsense reasoning and factual knowledge reasoning. The results show that ISC significantly improves the accuracy of the generated outputs, demonstrating the effectiveness of the proposed method in enhancing the self-correction capabilities of even small LMs. The paper also discusses the limitations and future directions, emphasizing the importance of domain knowledge and multiple iterations for more accurate self-correction.