arXiv 实施每月两篇投稿限额：应对 AI 论文激增致志愿者审核过载与平台生态失衡

arXiv now limits submissions to two papers per person per month because AI tools have driven monthly submissions past 40,000, doubling in two years and growing sixfold in the AI category alone, overwhelming its volunteer moderators.

Starting October 1, 2026, the preprint platform Arxiv will allow only two submissions per person per calendar month. The existing cap of three active submissions at any given time, in place since 2024, stays.

According to Arxiv, the limit applies across all categories and only to the person who submits the paper. Co-authors aren't affected. Rejected papers still count toward the limit because they eat up moderation time. But if authors delete a paper before publication, they get that slot back.

AI-fueled growth is overwhelming moderators

Arxiv points to the sharp rise in submissions, partly driven by AI tools, as the reason for the new rule. In September 2016, the platform received 9,869 submissions. By September 2024, that number hit 20,569. By September 2026, it reached 40,363. That surge generated nearly 9,000 support tickets for staff and moderators. In the cs.AI category alone, submissions grew more than sixfold in two years.

Arxiv recorded more submissions in September 2026 than in any other single month in its history. | Image: Arxiv

Volunteer moderators are seeing more thin, narrowly focused papers, "salami" papers sliced into tiny pieces, and dense texts written by AI. Thomas G. Dietterich, chair of the Arxiv Editorial Advisory Council, calls the volunteer moderators the backbone of quality control.

"However, a relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators’ time," he writes. "This is unfair to authors who continue to submit quality papers — their papers can be delayed for days or weeks as a result." On Bluesky, he adds that some authors submit dozens of papers.

Submissions to the cs.AI category have grown more than sixfold since early 2024. | Image: Arxiv

Dietterich also hopes the limit will curb "micro-result" papers that describe, say, one extra ablation study or a minor tweak to an architecture. Arxiv calls the rule a temporary fix until its own moderation tools can scale. Dietterich says he hopes the cap can be raised or dropped later.

Three rounds of tightening in under a year

Arxiv had already tightened rules for computer science papers in November 2025. Since then, review articles and position papers must go through peer review. In May 2026, Dietterich announced a one-year ban for authors whose papers clearly contain unverified LLM outputs like hallucinated sources.

AI conferences are dealing with the same problem. ICLR 2027 had already received about 50,000 abstracts before its deadline. ICLR 2026 got around 19,500 valid submissions. Among roughly 5,000 accepted papers at NeurIPS 2025, GPTZero found at least 100 fabricated citations, even though each paper had at least three reviewers.

Major funding arrives as Arxiv goes independent

After 25 years, arXiv is splitting from Cornell University and has been transitioning to an independent nonprofit since July 1, 2026. In late September, the Simons Foundation International, XTX Markets, and the Siegel Family Endowment pledged a total of $17.2 million, spread over three to five years. arXiv plans to put part of the money toward technical upgrades, with a focus on handling AI-generated content.

The major AI labs, meanwhile, are pushing hard to position their models as research tools. In August, Google Deepmind expanded its Co-Scientist into a system designed to plan experiments, control lab equipment, and generate manuscripts. Anthropic is building its own biology lab where Claude is meant to guide robots through experiments. In September, OpenAI announced an "automated research intern" that handles defined research tasks under human oversight.

A study involving Princeton and the UK AI Safety Institute took a different angle. It had Claude Opus 4.8 work for six days on research questions from two unpublished NeurIPS papers. The original authors, acting as reviewers, rejected both sets of results.