Two Papers a Month, Rejections Included: arXiv Imposes Its Harshest Rate Limit as AI-Driven Submissions Double in Two Years

arXiv, the preprint server that has been the de facto front door for computer science and AI research for three decades, has imposed the strictest submission limit in its history: from October 1, 2026, each submitter may file no more than two new papers per calendar month. The change was announced by community engagement manager Kat Boboris and took effect immediately.

The mechanics are unforgiving. The cap applies to whoever presses "submit" — co-authors are unaffected, so a four-person team can still file eight papers a month if each takes a turn. Rejected submissions count against the quota; only papers deleted before announcement do not. The limit applies across all categories and stacks on top of the existing ceiling of three active submissions at any given time, a rule arXiv has had since 2024.

The numbers behind the decision are stark. September 2026 brought a record 40,363 submissions, up from 20,569 in September 2024 and 9,869 in September 2016 — a 309 percent rise over a decade, and a doubling in just two years. Submissions in the computer science AI category grew more than sixfold over the same two-year span. That single month generated nearly 9,000 support tickets for staff and moderators, a corps of roughly 300 volunteers.

What moderators say they are seeing is not more science but more residue: thin papers of narrow scope, single research contributions fragmented into several "salami" submissions, survey articles that amount to annotated bibliographies, and densely AI-written text. arXiv's rules already permit AI as a research tool when it is disclosed and the work meets the archive's scholarly standards. "Many of the submissions that arXiv is now receiving do not satisfy these requirements," Boboris wrote. Rejection rates have climbed from around 4 percent to between 10 and 12 percent.

Editorial advisory council chair Thomas Dietterich, a professor emeritus at Oregon State University, put the blame squarely on a minority: "A relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators' time," delaying legitimate work "by days or weeks." Mathematician Alex Kontorovich of Rutgers was blunter still: "I'm not going to read your 42 papers this month. Pick two, your very best two."

Not everyone accepts the cure. Scott Kominers, a Harvard Business School professor and longtime arXiv contributor, argues the cap punishes everyone for the behavior of a few: "arXiv would do better to economise on moderation — not on scholarship." With no exceptions for conference deadlines, no carryover of unused slots and rejections consuming quota, a solo researcher racing a conference deadline now plans submissions the way one budgets a scarce currency.

arXiv frames the limit explicitly as a stopgap. The organization became an independent nonprofit in July 2026, and on September 23 announced a $17.2 million donation — part of it earmarked for tools to manage AI-generated content. Since May 2026 it has issued one-year bans for papers containing hallucinated references or leftover LLM prompts. The deeper fix — automated provenance checking that verifies data trails, code and version histories — may be the only approach that scales with submission volume, because human moderation bandwidth, unlike model output, does not.