“Self-Refine: Iterative Refinement with Self-Feedback,” submitted to arXiv on March 30, 2023 by Aman Madaan, Niket Tandon, Prakhar Gupta, and a large team including Uri Alon, Yiming Yang, and Peter Clark, showed that a language model can improve its own work by playing both author and critic in a loop.
The method has three steps that repeat. The model generates an initial output. The same model then produces written feedback on that output, pointing out weaknesses. The model uses that feedback to refine the output. This continues for several rounds. Crucially, it uses a single frozen model with no supervised training data, no additional fine-tuning, and no reinforcement learning - everything is done through prompting at inference time.
Across seven diverse tasks, from dialogue response to code optimization, Self-Refine improved task performance by roughly 20 percentage points on average over a single-pass baseline, using models including GPT-3.5, ChatGPT, and GPT-4. The result helped establish that self-critique is a usable lever, though later work noted models are better at refining when given an external signal than when relying purely on their own judgment.
Why business readers should care: Self-Refine shows that a model reviewing and revising its first draft can substantially raise quality at the cost of extra inference calls. It is a practical, training-free way to trade compute for better output in writing, coding, and reasoning tasks.