The paper introduces a novel training technique called Reflective Augmentation (RefAug) to enhance the mathematical reasoning capabilities of language models (LMs). Traditional data augmentation methods focus on expanding the training set by creating additional instances, which is effective for standard single-round question-answering tasks but does not foster deeper understanding. RefAug addresses this by incorporating reflection into each training instance, encouraging the model to consider alternative perspectives and engage in reflective reasoning. This method is designed to improve performance in both standard and complex reflective reasoning scenarios. Key contributions of RefAug include: 1. **Enhanced Problem-Solving Skills**: RefAug boosts the model's accuracy in standard single-round math reasoning tasks by +7.2%. 2. **Improved Reflective Reasoning**: It significantly enhances the model's performance in reflective math reasoning tasks, such as follow-up questions and error correction, with gains of +12.3%, +22.3%, +10.6%, and +9.2% respectively. 3. **Complementary to Existing Techniques**: RefAug complements traditional data augmentation methods, leading to even greater improvements when combined. The paper also includes extensive experiments on various math reasoning tasks, demonstrating the effectiveness of RefAug. Additionally, it explores the impact of different components of the reflective section and the stability of the annotation process. The results show that RefAug not only enhances basic problem-solving skills but also cultivates reflective reasoning abilities, making it a valuable complement to existing augmentation techniques.The paper introduces a novel training technique called Reflective Augmentation (RefAug) to enhance the mathematical reasoning capabilities of language models (LMs). Traditional data augmentation methods focus on expanding the training set by creating additional instances, which is effective for standard single-round question-answering tasks but does not foster deeper understanding. RefAug addresses this by incorporating reflection into each training instance, encouraging the model to consider alternative perspectives and engage in reflective reasoning. This method is designed to improve performance in both standard and complex reflective reasoning scenarios. Key contributions of RefAug include: 1. **Enhanced Problem-Solving Skills**: RefAug boosts the model's accuracy in standard single-round math reasoning tasks by +7.2%. 2. **Improved Reflective Reasoning**: It significantly enhances the model's performance in reflective math reasoning tasks, such as follow-up questions and error correction, with gains of +12.3%, +22.3%, +10.6%, and +9.2% respectively. 3. **Complementary to Existing Techniques**: RefAug complements traditional data augmentation methods, leading to even greater improvements when combined. The paper also includes extensive experiments on various math reasoning tasks, demonstrating the effectiveness of RefAug. Additionally, it explores the impact of different components of the reflective section and the stability of the annotation process. The results show that RefAug not only enhances basic problem-solving skills but also cultivates reflective reasoning abilities, making it a valuable complement to existing augmentation techniques.