ICLR's 2 September caps stand while Science and Nature split on the 2027 count

Retraction Watch’s weekend reads, posted 17 October 2026, links a 7 October Science article on conference submission caps. The caps are older than that roundup. On 2 September the ICLR 2027 program chairs published Submission policies for ICLR 2027. The author guidelines are the rule text.

Two outlets, two totals

Science, Jeffrey Brainard, 7 October at 1:40 p.m. ET, says ICLR received about 500 submitted papers ten years ago, and that the 2027 deadline last month brought more than 40,000. The guidelines set that full-paper deadline at 25 September 2026, Anywhere on Earth, after abstracts on 18 September.

Nature, Dalmeet Singh Chawla, 13 October, says submissions to ICLR 2027 look set to exceed 62,000, up from 490 a decade ago. It quotes program chair Jacob Andreas, of MIT: “Historically, the vast majority of new researchers have focused their effort on a single paper or project, and we want to incentivize this practice, rather than encouraging people to submit very large numbers of, usually AI-generated, papers.”

The 2 September post and the guidelines print no 2027 total. The 31 March retrospective says ICLR 2026 received 19,525 valid, format-compliant submissions. Science writes 19,500 where that page writes 19,525.

Science writes that 20% of the 19,500 submissions had no qualifying reciprocal reviewer. The September blog says “about 20%” of ICLR 2026 submissions, “computed by comparing the author lists to the final PC list,” and it does not print 19,500. A quarter of that set was desk-rejected. Those that reached full review were accepted at about half the rate of submissions with an author on the program committee. The chairs say they expect these figures to be much higher for 2027. That is their expectation.

The limits on the guidelines page

No author may appear on more than 20 papers. Each author may appear on at most one paper in which no author is an eligible reciprocal reviewer. In the page’s words: if you and all of your co-authors have never had a paper accepted at a major machine learning conference or journal, you may submit at most one paper this year. Submissions still over either quota at the full-paper deadline are randomly desk-rejected until every author is under quota.

The blog’s “new authors” are people who have never published at a major AI, ML, computer-vision, robotics, or NLP conference. The guidelines’ list also includes journals, among them JMLR and TMLR, and it excludes workshop papers, position papers, and similar side publications. An accepted NeurIPS 2026 paper alone does not qualify a reviewer for this cycle. Eligibility uses papers accepted by the abstract deadline.

The chairs write that 20 “does not represent a reasonable number of submissions,” and: “We personally find it difficult to imagine how one author could contribute meaningfully to 20 papers.” Fewer than 0.2% of authors submitted more than 20 papers last year, they write, and no author had more than 20 accepted. They say a tighter cap on short notice could have put junior submitters in large groups at risk, because those authors do not always control which papers a principal investigator expects to join. They also say the rules invite gift authorship, that they are monitoring for it, and that reports can go to [email protected].

Authors on three or more papers must review at least six, unless they hold an organizing-chair role. Every submission needs at least one qualified author registered for at least three reviews. If nobody qualifies, that requirement is waived and the one-paper cap applies. OpenReview profiles are the record, and a wrong profile is grounds for desk rejection.

ARR’s October lottery

ACL Rolling Review’s sustainable reviewing policy applies from October 2026. It says 38% of the roughly 17,000 May 2026 submissions came from authors with no available publication record. A paper is guaranteed review only with a qualified service contributor. The rest enter a lottery, and papers the lottery does not pick are desk-rejected. One person can cover at most two submissions. Authors are capped at 20 submissions, and at five as first or joint-first author.

Science describes a qualified ARR reviewer as someone with two papers at AI conferences on a designated list. ARR’s qualifications page is more specific: at least two major-venue papers for someone with a PhD or a senior academic post, and at least three for a doctoral student or for an industry researcher with a master’s degree.

One NeurIPS detector figure

On 2 June the NeurIPS blog said Pangram 3.3.2 gave 273 of 969 position-track submissions a 100% score, and that this score should not be read as “100% of the text is AI-generated.” With shorter windows the share at 90% to 100% fell from 42.7% to 12.7%, and the chairs said 178 submissions on that track would be desk-rejected. That is a June result for one track.

For ICLR 2027 the operative numbers are the guidelines: 20 papers, one paper with no qualified reciprocal reviewer, and a random desk-reject of the excess. Science’s figure is more than 40,000. Nature’s, six days later, looks set to pass 62,000. ICLR’s own pages print neither.