arXiv Caps Submissions at Two a Month as AI-Generated Papers Flood the Pipeline

arXiv, the preprint server that functions as the de facto first stop for nearly every new machine learning paper, imposed hard submission limits on October 1: two new papers per author per month, with no more than three submissions active in the review pipeline at once. It is the most restrictive change in the archive's three-decade history, and arXiv's own announcement ties it directly to an AI-driven surge in volume that moderators say they can no longer keep pace with.

The numbers behind the decision are stark. arXiv logged 40,363 submissions in September 2026, against 20,569 in September 2024 and just 9,869 a decade earlier in September 2016 — submissions nearly doubled in two years after doubling over the previous eight. The cs.AI category alone has grown more than sixfold in the past two years, and September generated roughly 9,000 support tickets, an unmanageable load for a moderation system still run largely by volunteers. Thomas Dietterich, chair of arXiv's computer science advisory committee, framed the cap as targeting a small slice of prolific submitters producing 'thin' or salami-sliced papers who 'consume a disproportionate fraction of the moderators' time' — work that generative AI tools have made dramatically cheaper to produce at scale.

The rate limit builds on a narrower policy arXiv rolled out in May 2026: a one-year submission ban for papers showing 'incontrovertible evidence' of unchecked AI generation, such as fabricated citations or leftover chatbot meta-commentary (one flagged paper reportedly included the line 'would you like me to make any changes?'). That policy was itself a response to a documented spike in the broader literature — a Lancet-published analysis found fabricated references in biomedical papers climbing from roughly 1 in 2,828 in 2023 to 1 in 458 by 2025, and to 1 in 277 in the first seven weeks of 2026, a roughly twelvefold rise in under three years.

Not everyone treats the cap as an unambiguous fix. AI researcher Nathan Lambert, whose widely circulated RLHF reference work would not have cleared the new monthly limit had it been in effect, argued publicly that arXiv should lean the opposite direction — accepting nearly everything and letting the field sort quality downstream rather than gatekeeping at submission time.

This feels like the early stages of a slow death of Arxiv — Nathan Lambert, on the new submission caps

Lambert also predicted that peer-review gating, currently applied only to review and position papers, will expand to ordinary technical submissions within two to three years if volume keeps climbing at the current rate. The mechanics of the new policy give that prediction some teeth: rejected submissions still count against an author's monthly quota, and the three-active-submission ceiling means labs that split a large project into coordinated papers, or post follow-up corrections, can find themselves bottlenecked for weeks whenever the moderation queue backs up — a friction that falls hardest on exactly the high-output labs producing the frontier-model and benchmark papers this category covers.