What the program chairs wrote
On 31 March 2026 the ICLR program chairs published a retrospective. ICLR 2026 received 19,525 valid, format-compliant submissions. 779 were desk-rejected for procedural or content violations. In the section on hallucinated references they wrote that many submissions included one or more references to documents that did not exist, or that had seriously wrong bibliographic details.
They described an automated system that extracted references and checked them against bibliographic databases and web search. They also wrote that the system had a significant false-positive rate, for example when authors had translated a non-English title. Area chairs reviewed the flags, then the program chairs checked them. Each paper with flagged references was reviewed by at least three people. Papers with confirmed hallucinated references were desk-rejected, with an appeal channel. The chairs wrote that this process partly explains a higher desk-reject rate than in past years. They did not publish a count of how many papers that was.
The earlier policy, and a vendor’s count
On 19 November 2025 the same conference said that extensive or careless use of language models often produces hallucinated references, that such hallucinations would be treated as a code-of-ethics violation, and that papers with those problems were being desk-rejected. Detection tools were described as a way to triage, with action only after an area chair or senior area chair found concrete evidence.
Separately, GPTZero reported that it scanned 300 submissions, that its tool flagged 90, and that after human review it counted 50 with at least one hallucinated citation. That number is GPTZero’s, from a sample, not a figure in the ICLR retrospective. Do not add it to the 779 desk rejects. The chairs’ 779 covers procedural and content violations as a whole, not hallucinated references alone.
What this means for a checker
The system the chairs described is the same shape as RefCheck: extract a reference, look it up in bibliographic databases, and do not treat a database miss as final when the title may have been translated or the record is thin. RefCheck’s “unconfirmed” verdict exists for that reason. “Likely fabricated” is for a formatted citation that the queried databases do not support.
RefCheck was not the system ICLR described, and this page does not claim it would reproduce their decisions. False positives are the point of their human review. If you are submitting, the practical page is check references before you submit.