Main idea:

• Vanilla implementation of intrinsic self-correction.
  • A single model responses, give feedbacks, and refines.
  • No fine-tuning is involved. Everything is done on the natural language level.

Concern:

• According to the Google DeepMind paper which was published after this, self-refine worked because, in the evaluation part, the initial prompt was incomplete, and the secondary feedback gave extra information to guide the model in a better way.
• After adjusting the prompt, they found out that the single response outperformed the self-refine.