Example

ICLR 2025: Feedback Assignment Versus Selected Revision Preferences

A paper is a written research report; a review critiques it. Thakkar and colleagues’ author-posted draft (arXiv v1, 2025-04-13) reports AI advice on ICLR 2025 reviews in October–November 2024. Papers were assigned by chance to advice for no, half or all reviews; half-group reviews were also selected by chance. Posting depended on need and automated checks; revisions were voluntary. Selected reviews gained about 14 words more on average than unselected reviews; roughly 80-word gains described voluntary updates, not a chance-assigned comparison. Two researchers preferred revisions in 89 of 100 original–revised pairs restricted to reviews using over 60% of 3–4 suggestions. That selected preference concerns clarity and usable suggestions, not experiment-wide scientific accuracy.

0

1

Updated 2026-11-09

Contributors are:

Who are from:

Tags

Scientific Oversight of AI-Assisted Research

1Cademy