arXiv imposes global submission limit: 2 papers per month as AI papers surge 6x

In its official blog post "Fair Moderation, Equitable Access, and AI," published on October 1, 2026, arXiv announced that each actual submitter may submit at most 2 papers per calendar month. The new rule took effect that same day and covers all subject categories including computer science, mathematics, and physics; switching submission categories cannot reset the quota. At the same time, no more than 3 papers may be in active submission status—the latter restriction has existed since 2024, and what is new this time is the unified monthly quota.

The key to understanding this set of rules is that it constrains the "act of submission," not "authorship." The quota counts only the person who actually submits the paper, not other co-authors. Even if a researcher is involved in ten papers at once, as long as he is not the one submitting them, his own monthly quota will not be used up. Conversely, the limit targets the number of papers submitted rather than the number ultimately published: even if a submission is rejected, it still consumes that month's quota, because moderators have already spent time on it.

The official announcement also clarified several boundaries. A manuscript still on hold across months does not count against the next month's two-paper quota, but it still counts toward the cap of at most three active submissions; if an author discovers a major problem before the paper is officially announced and deletes the submission, no quota is consumed. For authors who have accumulated multiple results or need to make papers public before a specific academic conference date, arXiv asks them to plan their submission timing in advance.

The direct backdrop to the submission cap is the accelerating expansion of submission volume. In September 2026, arXiv received 40,363 paper submissions, a record high in its 35-year history, and generated nearly 9,000 support tickets waiting for staff and moderators to process. Looking further back along the timeline: September 2016 saw 9,869 submissions, the same period in 2024 saw 20,569, and September 2026 reached 40,363—the first doubling took eight years, while the most recent one took only two.

The growth is not uniform across disciplines. Official statistics charts show that monthly submissions in the cs.AI category grew from a few hundred in early 2024 to more than 3,000 in September 2026, an increase of more than sixfold over two years, while submissions in other categories roughly doubled over the same period. It should be noted that a sixfold increase in AI-category papers does not equal a sixfold increase in low-quality AI-generated papers; it also includes the expansion of normal research activity.

What absorbs this torrent is a moderation mechanism based mainly on volunteer labor. arXiv's 2025 annual report shows that its moderation network has about 260 people, with only 7 dedicated operations staff plus the scientific director; each publication day requires processing 700 to 1,400 new papers, as well as up to a thousand revision appeals. The platform also needs dedicated teams to handle submission support, technical issues, and dispute appeals.

arXiv officials said moderators have observed a growing number of thin papers with narrow research scope and limited contributions, as well as salami papers that split one complete study into multiple smaller results, along with a clear increase in AI-written papers—some of which are lengthy and structurally complex yet lack the substantive academic contribution required by the platform.

The platform does not oppose researchers using AI. Official policy allows AI-assisted research and paper writing, provided it is disclosed as required and the authors are responsible for the content of the paper, and the research itself must meet academic standards. What truly troubles the platform are submissions that use AI to lower writing costs without adequately testing the content and academic value.

Thomas Dietterich, chair of arXiv's editorial advisory committee and a former president of AAAI, noted in the announcement that only a relatively small number of authors submit large quantities of low-quality papers, yet they consume a disproportionate amount of moderation time, causing the publication of other high-quality papers to be delayed by days or even weeks. This judgment shifts the justification for the submission cap from "controlling total volume" to "correcting an imbalance in resource allocation."

Robbert Wijers, a researcher involved in the platform's moderation work, estimates that arXiv has long had an average of about 4% of submissions rejected, but since low-quality AI submissions grew markedly in early 2025, that proportion has risen to about 10% to 12%. It should be made clear that this is the interviewee's estimate, not the latest official statistic disclosed in this announcement, so it can only be treated as a source claim.

Another frequently cited figure comes from a study published in Science in 2025, which estimates that researchers who adopted large language models subsequently published about 36% more papers on arXiv. Likewise, this is a research estimate rather than official platform statistics; it shows that the mode of paper production is changing, but it cannot be directly equated with the growth rate of low-quality submissions.

This submission cap is an extension of a series of governance actions by arXiv. In October 2025, the platform tightened the acceptance requirements for review and position papers in computer science categories, in principle requiring such papers to provide proof of formal peer review acceptance; in May 2026, it further clarified accountability for unchecked LLM-generated errors, including a one-year submission ban.

Now, governance measures have extended further to the submission frequency of all submitters across all disciplines. From tightening specific paper types, to holding people accountable for generated errors, to a unified monthly quota, this thread points in the same direction: under conditions where AI lowers the cost of writing, the platform is trying to use rules to redraw "what deserves to occupy moderation attention."

The actual impact of the new rule depends on what it replaces in the research system. In the AI field, arXiv is not a dispensable paper repository: traditional conferences and journals have fixed deadlines and longer review cycles, and many teams are accustomed to making results public early through arXiv so that peers can learn about new methods as soon as possible and carry out reproductions and follow-up research. Limiting submission frequency does not equal limiting research activity itself, nor does it directly change eligibility to submit to academic conferences, but it may affect when some results are made public.

Because the quota targets only the actual submitter rather than all listed authors, one foreseeable result is that major industry players such as OpenAI and DeepMind and large university research groups can work around the limit by having someone else upload, while those truly affected may be independent scholars and solo teams lacking co-authors to rotate through. This judgment is an inference based on the text of the rule, not an official explanation, and its actual extent remains uncertain.

Josh Bloom, a professor of astronomy at the University of California, Berkeley, warned in an article published on October 1 that short-term growth should not be attributed entirely to AI: September 2026 submissions rose 29% from August, but the same period in 2025 also saw a 22% month-over-month increase, and the academic calendar itself brings seasonal fluctuations. What truly deserves attention is the long-term trend—submissions in the first nine months of this year were already about 30% higher than the same period last year.