arXiv counts rejected papers against a new two-per-month submitter cap

arXiv is rationing moderator time. Its moderators are volunteers, and low-quality submissions are holding up good papers for days or weeks, socket.dev reported ^[10]. Thomas Dietterich, an Oregon State University distinguished professor emeritus who chairs arXiv's Editorial Advisory Council, said "a relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators' time" ^[11].

The older three-at-a-time rule ^[5] limits how many papers one submitter can have in the queue at once. The monthly cap limits how many go in. For a single account that comes to 24 papers a year ^[23].

The count is tied to the person who submits. Co-authors do not use up their own allowance when somebody else submits the paper ^[6]. A group of five authors who each submit their own papers gets ten slots a month. A group that sends everything through its principal investigator gets two ^[24].

The monthly count also hits a pattern arXiv named directly. arXiv wrote: "Our moderators are observing an increase in thin papers of narrow scope, as well as 'salami' papers, where a single work is broken up and submitted as a set of smaller papers" ^[2]. A work cut into four papers now takes one submitter two months to post ^[25].

September's record compares with 20,569 submissions a year earlier and 9,869 a decade before ^[7]. Volume nearly doubled in twelve months ^[21]. It is about four times the level of ten years ago ^[22]. Within computer science, the cs.AI category grew more than sixfold in two years ^[9]. Some of the extra papers are at least cheap to reject. Wijers told Science that moderators could turn some down at a glance, citing "a four-page paper that has 50 subsections in it, some of which do not contain any text" ^[15].

Moderators expected fake papers after ChatGPT launched but saw no change through 2024. Wijers said the surge began in early 2025, first in computer science and then in other fields ^[20]. Last October, computer science moderators stopped considering review articles and position papers unless a journal or conference had accepted them after peer review ^[17]. In December the mathematics section stopped granting automatic endorsement on an institutional email alone, and arXiv extended that rule to every section in January ^[18]. A new submitter now needs that email together with either authorship of an accepted paper in the relevant domain or an endorsement given personally by an established arXiv author in the same field ^[19]. As Wijers put it to Science, "almost all the AI generated rubbish comes from first-time submitters" ^[16].

The endorsement rules screen people on the way in. The monthly cap applies to submitters who are already in, the small group Dietterich described ^[11]. I think a per-account count is the right tool for a volunteer queue under this load. It takes no judgement to enforce. A determined author can still get around it by having co-authors submit ^[6]. AI tools remain allowed when their use is disclosed and the work meets arXiv's standards ^[13]. arXiv's statement said "AI tools are making it easy for authors to flood arXiv and other repositories with these low-value papers" ^[1]. The reporting covers only arXiv's response.