Research library

Critique and revision Research note

The first draft is not always the best draft

Feedback can change an output without changing model weights.

The finding

Self-Refine uses the same model to generate an answer, give feedback, and revise it. Across seven tasks, the authors report improvements over one-step generation using human assessments and automatic metrics, without additional training. [1]

What it means for Pingpong

The relevance to Pingpong is the value of allocating a separate opportunity to review. The product uses later models for that opportunity and supplies a fresh frame for evaluating earlier work. Its final answer is the result of the selected sequence, rather than a page of unrelated first drafts.

The limit

This study supports iterative revision in its tested settings. It does not establish that replacing the self-reviewer with another model is always better, or that each additional pass improves correctness.

Source

Self-Refine: Iterative Refinement with Self-Feedback

Aman Madaan et al. (2023)

Research paper, arXiv archive
DOI: 10.48550/arXiv.2303.17651

Paper PDF

This is Pingpong's interpretation of external research, not a result from a Pingpong experiment or an endorsement by the authors. Editorial standard