AI Reviews Are Biased Towards the Appendix
About half or more of the weaknesses GPT and Claude flagged came from appendices, compared with 8.4% for human reviewers.
The study examined 2,000 papers from ACL, EMNLP, COLM, and ICLR. The pattern held across API, coding-agent, and web access methods. Models singled out fine-grained technical details deep in appendices that were not central to the papers’ contributions. Human appendix criticisms focused mainly on presentation and content tied to the core idea.
HF Daily Papers' note
The study examined 2,000 papers from ACL, EMNLP, COLM, and ICLR. The pattern held across API, coding-agent, and web access methods. Models singled out fine-grained technical details deep in appendices that were not central to the papers’ contributions. Human appendix criticisms focused mainly on presentation and content tied to the core idea.
HF Daily Papers' note
score 4