ICML 2026 reviewers score batches differently

This looks less like a mystery and more like a classic reviewer-calibration problem, where local batch effects can swamp any global expectation of fairness.

• –Different topic areas can attract reviewers with very different norms for what counts as a strong paper.
• –Reviewer harshness is often uneven across batches, especially when reviewers are assigned in clusters or subfields.
• –Raw scores are usually noisy and not directly comparable across batches without calibration.
• –If ICML uses any normalization, it is likely limited and cannot fully erase batch-to-batch variance.
• –The main signal is probably relative ranking within a reviewer group, not the absolute score number.