self-refine

**Self-refine** is the **iterative prompting method where a model repeatedly generates output, evaluates it, and refines it toward better quality** - it formalizes draft-to-revision behavior within inference time. **What Is Self-refine?** - **Definition**: Closed-loop generation pattern of initial draft, self-feedback, and improved rewrite. - **Iteration Structure**: Can run fixed rounds or terminate when quality criteria are satisfied. - **Feedback Source**: Self-generated critique, rubric scoring, or external validator signals. - **Task Applicability**: Useful for writing, code generation, and constrained-format responses. **Why Self-refine Matters** - **Output Quality**: Multiple passes usually produce clearer and more accurate final responses. - **Error Recovery**: Early draft mistakes can be corrected before final delivery. - **Prompt Control**: Refine loop can enforce style, completeness, and policy constraints. - **Operational Flexibility**: Works without model retraining, using only inference-time logic. - **Cost Balance**: Additional passes add compute cost but can reduce human rework. **How It Is Used in Practice** - **Rubric Design**: Define explicit criteria for what counts as improved output. - **Iteration Limits**: Set max rounds and quality thresholds to control latency. - **Verification Step**: Add final consistency check before returning refined response. Self-refine is **a practical iterative-improvement framework for LLM applications** - structured revision loops can significantly enhance final-output reliability with manageable inference-time overhead.