The idea in one paragraph.
One model, no training, three roles: generate a draft, critique it, and revise using the critique - iterated until it stops improving.
THE MACHINERY
Generate-Feedback-Refine loop
The same model plays author and critic in alternation.
Self-generated feedback
Critique is produced by the model itself, with no external signal.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Requires no supervised training data, additional training, or reinforcement learning; one LLM is generator, refiner, and feedback provider.
Inspect a source excerpt
Self-Refine does not require any supervised training data, additional training, or reinforcement learning, and instead uses a single LLM as…Read the passage in context on arXiv ↗
Across 7 diverse tasks, outputs are preferred by humans and automatic metrics over conventional one-step generation, improving by ~20% absolute on average.
Even GPT-4 can be further improved at test time by this approach alone.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- No external signal: the critic shares the generator's blind spots, so errors it cannot see it cannot fix.
- Iteration can degrade output when the critique is wrong.
Why it belongs in this atlas.
The purest statement of test-time compute as a harness lever, and the clearest illustration of its limit. Contrast with Reflexion, which reaches outside for a real failure signal.