ICML 2026 review policies may skew scores

The signal may be real, but the cleaner explanation is probably a mix of reviewer-composition effects, author self-selection, and LLM-polished tone rather than a pure policy-driven scoring bias.

• –Policy B still forbids LLMs from judging strengths/weaknesses or writing the review, so any score gap would likely come from tone/compliance differences rather than an official scoring shortcut.
• –Policy A papers are guaranteed Actual Policy A reviewers, while Policy B papers draw a mixed A/B pool, so raw score comparisons are confounded before any LLM effect shows up.
• –The Reddit/X sample is self-selected, and people with strong feelings are more likely to respond, so the poll is better for hypothesis-building than for proof.
• –ICML says it will analyze score distributions and compensate systematic differences, which means this is a live process-design issue, not just forum gossip.