Preprint server arXiv has capped submissions at two per month after receiving 40,363 papers in September 2026, almost double the 20,569 of September 2024. The limit responds to thin, salami-sliced and dense AI-written papers that generated almost 9,000 support tickets in one month. The move matters for business because the channel companies use to track AI research is becoming harder to navigate.
Why arXiv limited uploads to two per month
Kat Boboris announced the rate limit last week in a post discussed on Hacker News, citing a vertiginous rise in filings. Monthly submissions to the cs. AI category rose more than sixfold from January 2024 to September 2026. For comparison, September 2016 brought only 9,869 submissions across the server. Moderators report more narrow-scope papers and single works split into smaller units, plus low-value papers that AI tools make easy to flood into repositories.
A May measure introduced a one-year ban for unchecked AI content, while an October 2025 rule required survey and position papers to have peer backing before publication on arXiv. Both types can be produced with less effort than a full study. The server retains nonprofit status, so it has less incentive to force site visits or logins, unlike commercial platforms. Its RSS feeds remain useful for monitoring, even as Reddit plans to eliminate RSS feeds next month.
The pressure concentrates in AI-related research, which has grown from obscurity into a political and economic signifier and eclipsed other categories. AI now operates at scale both in web scraping and in enabling vast submission volumes, creating a logistical and quality crisis. The signal-to-noise ratio keeps deteriorating as plausible-looking sections hide broken links between claims and evidence. A research strand has therefore emerged to measure and counter the decline.
What SciSlopBench means for research teams
A collaboration between Seoul National University and the University of Minnesota proposes to define scientific slop as failures connecting claims, evidence, citations and structure. Its benchmark SciSlopBench contains 390 AI-generated papers, each paired with a human paper on a similar problem and contribution type. It identified the AI paper in 85.9% of pairs, outperforming conventional AI-text detectors. For companies, this shifts screening from token-level checks to whole-paper reasoning audits.
The companion framework SciSlopHarness detects and repairs such failures while preserving underlying evidence, rejecting unsupported additions and reordering claims to match available data. It reduced the remaining AI-human gap by 63% over the strongest revision baseline without human reference targets. Standard revisions leave residual slop, while direct slop-aware prompting triggers reward hacking. Small teams gain a live demo for submitting papers for slop analysis, while large R and D organizations can embed the six checks into review pipelines.
The six measures cover cross-section references, macro redundancy, argument graph, citation isolation, figure exposition and evidence gap. The dataset pairs 143 papers from FARS with top-venue human papers matched by topic, plus 247 pairs from the 2025 Agents4Science competition. Performance uses PairAcc and AUROC, with detection reported at a fixed 5% false-positive rate for human papers. The authors own paper scored a moderate 40 out of 100 in their demo, which shows the tool flags weaknesses rather than certifying quality.
