Seems ICML is rejecting MANY unanimous positively rated papers [D]
A researcher's 4444-rated paper rejected; reviewers inflate scores to avoid debate
A Reddit post by user AffectionateLife5693 reveals a disturbing trend at ICML 2025: many papers with unanimously positive ratings are being rejected. The author's own submission (ID 4444, later 4443 after rebuttal) was rejected as expected. They explain that the conference's rebuttal phase, intended to encourage reviewers to reconsider scores, has backfired. Area Chairs (ACs) now seek homogeneous ratings among reviewers, creating pressure to increase scores just to end prolonged discussions. Many disengaged reviewers comply, inflating scores. Conversely, reviewers who initially gave high scores often refuse to update them even after their concerns are fully addressed.
One example cited: a reviewer wrote 'Thank you for the rebuttal. The paper is valuable. The rebuttal addressed all my concerns' but kept the score at 4 (out of 10). This inflation leads to uniformly high scores across papers, yet the conference has limited capacity. The author predicts a repeat of last NeurIPS's mass rejection of high-scored papers. They advocate returning to simpler peer review: independent evaluations by reviewers, ACs assessing quality and consistency, and resolving borderline cases through discussion. The current complex mechanism, they argue, worsens an already flawed system.
- A paper with a pre-rebuttal score of 4444 and unanimous positive reviews was rejected at ICML 2025
- Reviewers inflate scores to avoid back-and-forth, with one reviewer keeping a score of 4 even after saying all concerns were addressed
- ACs push for homogeneous ratings, leading to inflated scores and likely mass rejections of deserving papers, mirroring last NeurIPS
Why It Matters
Review score inflation undermines peer review integrity, risking fair acceptance of high-quality ML research.