SELF-REFINE is an innovative approach that enables large language models (LLMs) to iteratively refine their outputs through self-feedback. Inspired by human problem-solving, SELF-REFINE involves generating an initial output, using the same LLM to provide feedback on it, and then refining the output based on this feedback. This process is repeated iteratively until a stopping condition is met. The method does not require additional training data or reinforcement learning and relies solely on the LLM for generation, feedback, and refinement. The evaluation of SELF-REFINE across seven diverse tasks, including dialog response generation, code optimization, and math reasoning, demonstrates significant improvements over conventional one-step generation methods. On average, SELF-REFINE achieves a 20% absolute improvement in task performance. The approach is effective even with state-of-the-art LLMs like GPT-4, showcasing its potential to enhance the quality of outputs without extensive supervision. SELF-REFINE's effectiveness is attributed to the specific and actionable feedback it generates, which helps refine outputs more effectively than generic feedback. The method also shows resilience to sub-optimal feedback, making it robust in various scenarios. However, the approach has limitations, such as the need for sufficient few-shot modeling capabilities in the base models and the lack of open-sourced models for evaluation. Despite these limitations, SELF-REFINE offers a promising direction for improving LLM performance through iterative refinement.SELF-REFINE is an innovative approach that enables large language models (LLMs) to iteratively refine their outputs through self-feedback. Inspired by human problem-solving, SELF-REFINE involves generating an initial output, using the same LLM to provide feedback on it, and then refining the output based on this feedback. This process is repeated iteratively until a stopping condition is met. The method does not require additional training data or reinforcement learning and relies solely on the LLM for generation, feedback, and refinement. The evaluation of SELF-REFINE across seven diverse tasks, including dialog response generation, code optimization, and math reasoning, demonstrates significant improvements over conventional one-step generation methods. On average, SELF-REFINE achieves a 20% absolute improvement in task performance. The approach is effective even with state-of-the-art LLMs like GPT-4, showcasing its potential to enhance the quality of outputs without extensive supervision. SELF-REFINE's effectiveness is attributed to the specific and actionable feedback it generates, which helps refine outputs more effectively than generic feedback. The method also shows resilience to sub-optimal feedback, making it robust in various scenarios. However, the approach has limitations, such as the need for sufficient few-shot modeling capabilities in the base models and the lack of open-sourced models for evaluation. Despite these limitations, SELF-REFINE offers a promising direction for improving LLM performance through iterative refinement.