Leaderboards

Replication rates, ranked.

Aggregate replication rates across the corpus. PARTIAL counts as half-credit; everything else is a full pass or a full miss. The underlying data is the same set of reproduction jobs every verdict on this site links back to.

Labs

Top labs by share of papers that reproduced within tolerance.

  1. 1OpenAI100%n=3
  2. 2Google Brain100%n=2
  3. 3Google Research90%n=6
Open labs →

Conferences

Venues ranked by replication rate of their accepted papers.

  1. 1arXiv preprint100%n=3
  2. 2ICLR 2020100%n=2
  3. 3ICLR 2021100%n=2
Open conferences →

Authors

Authors ranked by replication rate across their listed papers.

  1. 1Alec Radford100%n=3
  2. 2Ilya Sutskever100%n=3
  3. 3Aditya Ramesh100%n=2
Open authors →

How these numbers are computed

Replication rate = (REPRODUCED + 0.5 × PARTIAL) ÷ decisive verdicts in the bucket. WRONG, DISPUTED, and OUT_OF_BUDGET count as not-yet-reproduced and consume the denominator. PENDING verdicts and verdicts in legal-hold states are excluded from the rate but still count toward the entity's paper total.

Per our defamation protection policy, an entity must accrue at least 2 papers with a current verdict before it appears in any public ranking. The methodology behind each verdict is documented at /methodology. The technical-term framing of WRONG is at /methodology/wrong.