Back to glossary

Term

Self-Refine

Self-Refine is a prompting technique in which a model critiques its own output and improves it across several iterations — without human feedback.

Self-Refine — explained in more detail

Self-Refine (Madaan et al., 2023) has a single language model take on three roles: first it generates an initial answer, then it critiques that answer with concrete improvement suggestions, then it rewrites the answer based on its own critique. The loop runs until the model itself signals “no further improvement needed” or an iteration limit kicks in. Everything happens without external feedback and without fine-tuning.

Example / Practical use

Strong on tasks with clear evaluation criteria: code generation, mathematical proofs, code optimization, longer texts with style requirements. Across several benchmarks, the original paper reported average gains of roughly 20% over plain one-shot answers. Practical note: more than 3–5 iterations rarely add further value.

How it differs from similar concepts

Unlike Self-Consistency, Self-Refine works sequentially and improves a single answer instead of sampling several in parallel. Unlike RLHF, there is no training — the “self-correction” happens entirely at inference time via prompts.

Entdecke mehr

Themenuebersicht