How are papers scored?
Mean of each paper’s reviewer ratings. Tap a bar to set a cutoff.
Jump to the top…
Where does my paper stand?
Tap each reviewer’s score to see where your mean lands.
Reviewer confidence for the next score
Reviewer score (adds a review)
What changed since the preliminary snapshot
Active submissions
Papers with discussion notes
Review scores changed
How the percentiles move
Share of papers at or above a mean of 2.50 and 2.67
… and at or above 3.00 and 3.33
Mean score needed for each top percentile
Who withdraws?
Share of papers active at the preliminary snapshot that have since been withdrawn or desk-rejected, by their preliminary mean rating.
When do reviewers raise a score?
Chance that a review’s score went up since the preliminary snapshot, by its own preliminary score and the average of the paper’s other reviews.
rarermore common
Individual review ratings
How the sub-scores line up with the rating
How long are the reviews?
Median words in the strengths and weaknesses sections, by the review’s rating.
Do reviewers agree?
Gap between the highest and lowest rating on each paper (papers with 2+ reviews).
Compare areas
The shape of each area
How paper mean ratings are spread in each area, ordered from highest to lowest mean.
Keywords
About this data
- Source. A public snapshot of OpenReview scores published by Paper Copilot (), not a live pull from OpenReview. Anonymous OpenReview API access is behind a bot challenge, so numbers were not cross-checked against the source.
- Scale. Ratings are 1–4 as recorded in that dataset. The page assumes higher is better (as in earlier ICLR scales); the meaning of each level was not verified.
- Snapshots.
- Who is counted.
- Other trackers. The ICLR 2027 Scores page reported 40,616 active submissions (40,458 rated) on 2026-11-06 13:54 UTC, before the first Paper Copilot ratings snapshot (40,273 active on 2026-11-07 02:15 UTC). The 343 papers in between had been withdrawn; they scored low (mean 1.53 vs 2.07). Active counts keep falling as authors withdraw, so totals depend on when a snapshot was taken.
- Paper score. The mean of a paper’s reviewer ratings. With 3 reviews it can only be a multiple of ⅓, so ties are common and “top 10%” can’t be split exactly: the chips pick the highest cutoff that still covers at least that share, and the readout shows the real share. Histogram bars cover a third of a point each (the left edge is the cutoff).
- History. The “What changed” charts use every Paper Copilot snapshot of the file (one every ~8 hours), each verified against its published hash. Snapshot charts (sub-scores, review length, agreement, keywords) use the latest one.
- Confidence-weighted score. With “Conf. weighted” selected at the top, a paper’s score is Σ(confidence × rating) / Σ(confidence), using each reviewer’s confidence (1–4) as recorded in the data, so confident reviewers count more. It applies to the headline numbers, the score distribution, “Where does my paper stand?”, the area comparison and the area shapes; the history charts (changes, percentiles over time, withdrawals) and keywords always use the plain mean. Weighted scores rarely tie, so “top X%” is sharper than with plain means, where many papers share 2.33 or 2.67.
- Chance of acceptance. ICLR 2027 decisions are not out, so the estimate uses ICLR 2026. Paper Copilot’s history has its scores at several points of the 2026 review period; for each, every active paper’s rank (share of papers scoring higher, ties split) is matched with its final decision. The page picks the 2026 snapshot taken the same number of days after reviews came out as the current one (reviews assumed public on 2025-11-12 and 2026-11-06) and reads the share accepted among 2026 papers at your rank, excluding papers withdrawn later (the fit is monotone in rank). It uses your rank among all areas, under the selected scoring. ICLR 2027 has about twice as many submissions and its acceptance rate is unknown, so three cases are shown, as shares of all submissions: the 2026 rate (27%), 20%, and the rate at which 2027 would accept the same number of papers as 2026 (5,351 in the main track, about 12.5%). In each case your absolute rank among the 2027 papers in review is placed on the 2026 pool: 2026-equivalent top share = your top share × (2027 papers in review ÷ 2026 papers in review) × (2026 accepted ÷ 2027 accepted).
- Percentiles. Linear interpolation across papers, computed exactly from the per-paper means. “How the percentiles move” recomputes them at every snapshot; its “still active now” view fixes the set of papers to those active in the latest snapshot.
- Area differences are small (about 0.2 points against a within-area spread of about 0.45) and are descriptive, not a statement about paper quality.
Source: Paper Copilot snapshot of OpenReview scores () · active submissions