Self Refine: Definition and Examples

Self Refine is an iterative improvement approach in which a large language model (LLM) simultaneously plays the roles of generator, critic, and reviser. Instead of producing a single response, the model first produces an initial draft, then analyzes it to identify weaknesses, and finally rewrites it incorporating its own feedback. This cycle can be repeated several times until a satisfactory quality level is reached.

This technique is directly inspired by the human writing process: one rarely writes a perfect text in a first draft. We reread, identify weak passages, rephrase, restructure. Self Refine replicates this cognitive mechanism by formalizing it in a structured prompt. The key lies in the model's ability to step back from its own output and apply explicit evaluation criteria.

Concretely, a Self Refine prompt decomposes into three distinct phases: initial generation, feedback (where the model critiques its own output according to defined criteria), and revision (where it produces an improved version integrating that feedback). Each phase can be handled in a separate prompt or chained within the same exchange. The major advantage is that no additional model or training data is needed.

Self Refine was formalized in a 2023 research paper by Madaan et al. and demonstrated significant improvements on diverse tasks such as code generation, mathematical reasoning, writing, and dialogue. This technique has become a pillar of advanced prompt engineering and naturally combines with other approaches like Chain of Thought or multi-agent systems.