Research library

Critique and revision Research note

Asking again can make a correct answer wrong

Revision needs a reason beyond the existence of another pass.

The finding

Huang and colleagues study intrinsic self-correction: revision without external feedback. On the reasoning tasks they evaluate, models struggle to correct themselves and sometimes become less accurate after another attempt. [1]

What it means for Pingpong

Pingpong brings another model into the review, but that model is not an answer key. The reason to retain the original question and earlier responses is to let a reviewer identify a specific gap, not to assume the latest wording must be superior. A good chain must be allowed to preserve a correct answer.

The limit

The title describes the models and conditions studied, not an eternal limit on AI. For Pingpong, the important risk is regression: tests must count correct-to-incorrect revisions as well as successful corrections.

Source

Large Language Models Cannot Self-Correct Reasoning Yet

Jie Huang et al. (2023)

ICLR 2024
DOI: 10.48550/arXiv.2310.01798

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