This paper introduces Intrinsic Self-Correction (ISC), a method to enable small language models (LMs) to self-correct their responses without relying on external verification. The approach involves a self-triggered process where the model evaluates its own answers and modifies them if errors are detected. The key components of ISC include constructing self-correction data and employing Partial Answer Masking (PAM) to facilitate self-verification. The self-correction pipeline is designed to be integrated into the model's training process, allowing it to self-correct during response generation. The study evaluates ISC on two tasks: commonsense reasoning and factual knowledge reasoning, using LMs with parameter sizes ranging from 6 billion to 13 billion. The results show that ISC significantly improves the accuracy of generated answers, with some models achieving up to a 5.6% improvement. The method is effective for both large and small LMs, demonstrating that even small models can benefit from intrinsic self-correction. ISC is implemented through a data construction pipeline that generates self-correction data, which is then used for fine-tuning. The PAM technique is introduced to enable the model to self-verify its answers by focusing on the verification process rather than the entire answer. This approach allows the model to learn from its own mistakes and improve its responses over time. The experiments show that ISC enhances the model's ability to self-correct, leading to more accurate and reliable outputs. The method is particularly effective in scenarios where the model's initial response is incorrect, as it allows the model to recognize and correct these errors autonomously. The results indicate that ISC is a promising approach for improving the performance of LMs, especially in tasks where accuracy is crucial. The study also highlights the importance of self-verification in the self-correction process, as it enables the model to assess its own responses and make necessary adjustments. Overall, the findings suggest that intrinsic self-correction can significantly enhance the quality of generated answers, making it a valuable technique for improving the capabilities of language models.This paper introduces Intrinsic Self-Correction (ISC), a method to enable small language models (LMs) to self-correct their responses without relying on external verification. The approach involves a self-triggered process where the model evaluates its own answers and modifies them if errors are detected. The key components of ISC include constructing self-correction data and employing Partial Answer Masking (PAM) to facilitate self-verification. The self-correction pipeline is designed to be integrated into the model's training process, allowing it to self-correct during response generation. The study evaluates ISC on two tasks: commonsense reasoning and factual knowledge reasoning, using LMs with parameter sizes ranging from 6 billion to 13 billion. The results show that ISC significantly improves the accuracy of generated answers, with some models achieving up to a 5.6% improvement. The method is effective for both large and small LMs, demonstrating that even small models can benefit from intrinsic self-correction. ISC is implemented through a data construction pipeline that generates self-correction data, which is then used for fine-tuning. The PAM technique is introduced to enable the model to self-verify its answers by focusing on the verification process rather than the entire answer. This approach allows the model to learn from its own mistakes and improve its responses over time. The experiments show that ISC enhances the model's ability to self-correct, leading to more accurate and reliable outputs. The method is particularly effective in scenarios where the model's initial response is incorrect, as it allows the model to recognize and correct these errors autonomously. The results indicate that ISC is a promising approach for improving the performance of LMs, especially in tasks where accuracy is crucial. The study also highlights the importance of self-verification in the self-correction process, as it enables the model to assess its own responses and make necessary adjustments. Overall, the findings suggest that intrinsic self-correction can significantly enhance the quality of generated answers, making it a valuable technique for improving the capabilities of language models.