arXiv now permits each submitter to send at most two new papers per calendar month, while retaining a ceiling of three active submissions at any one time. The rule began on 1 October and applies across every subject category. It is a blunt response to a precise operational problem: papers can now be produced faster than a largely volunteer moderation system can inspect them.[1]
September brought 40,363 submissions and almost 9,000 support tickets. The corresponding submission counts were 20,569 in September 2024 and 9,869 in September 2016. That is a 96.2% increase in two years and a 309% increase over the decade, calculated from arXiv’s published figures. The repository says cs.AI alone grew more than sixfold over two years, while moderators reported more narrow papers, sliced-up “salami” submissions and dense AI-written manuscripts that did not meet arXiv’s scholarly standards.[1]
The rule counts attempts because attempts consume review time
How the new rate limit works
| Situation | Treatment |
|---|---|
| Two submitted papers in one calendar month | The submitter must wait until the next month, whether the papers are accepted or rejected.Source 1 |
| Submissions in different categories | They share the same two-paper allowance.Source 1 |
| Papers still on hold when a new month begins | They do not consume the new month’s allowance, but they remain part of the three-active-submission cap.Source 1 |
| Paper submitted for several co-authors | Only the person who presses submit uses an allowance; the co-authors’ allowances remain available.Source 1 |
| Submission deleted before announcement | It does not count toward the monthly or active limit.Source 1 |
This design explains both the rule’s fairness logic and its weakness. Moderators spend time on a submission even when they reject it, so the allowance counts attempts. But the limit attaches to a person rather than a research group. A team with many co-authors can rotate the submitting role; an independent researcher cannot create more submitters. The policy reduces the maximum rate of an individual account, not the total production capacity of a large collaboration.[1][2]
Arrival is outpacing the volunteer checkpoint
The moderation pressure in September
| Measure | Value | Context |
|---|---|---|
| Submissions | 40,363 | About 1,345 per day over a 30-day month.Source 1 |
| Volunteer moderators | Nearly 300 | Nature reports that they collectively check about 1,000 manuscripts per day.Source 2 |
| Waiting backlog | Nearly 4,000 | The backlog delays legitimate papers as well as low-value ones.Source 2 |
| Support tickets | Almost 9,000 | A second workload beyond the manuscripts themselves.Source 1 |
arXiv moderation is a gate for relevance, format and scholarly interest, not a correctness judgment equivalent to journal peer review. That makes attention the limiting input. A paper that takes minutes to generate can still require a domain expert to recognize that its references are fabricated, its contribution has been split unnaturally or its dense prose hides no meaningful advance.[1][2]
A quota changes behaviour before it changes quality
The most constructive effect may be editorial. Two slots force a prolific submitter to decide which work is mature enough to occupy them. A paper rejected on basic standards now carries an opportunity cost. The less constructive effect is displacement: authors can route papers through colleagues, delay legitimate clusters of work or move low-quality submissions to repositories with fewer safeguards. The rule cannot tell a rigorous third paper from a synthetic one.
Longer-term remedies need to operate closer to the scarce resource. Better triage could identify duplicate or sliced submissions, verify citations and detect obvious template residue before a volunteer reads the paper. More professional moderation would reduce dependence on donated time. Stronger provenance and disclosure could help readers and moderators understand how a manuscript was produced without treating AI assistance itself as disqualifying.
The two-paper rule is therefore best read as an infrastructure alarm. Scientific text generation has become cheap enough that access to a repository can no longer be governed only by the assumption that writing a paper is slow. The expensive act is deciding whether the paper belongs in the scholarly record. arXiv has rationed the queue while the research community works out how that decision should be made at machine-era volume.
Sources & further reading
- Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy arXiv Blog · 1 Oct 2026 · Accessed 16 Oct 2026
- ArXiv versus AI slop: will capping preprint submissions work? Nature · 15 Oct 2026 · Accessed 16 Oct 2026