arXiv has announced that starting October 1, 2026, each submitter can submit at most two papers per calendar month, and rejected submissions still consume the monthly quota. The policy aims to address a surge in submissions, particularly in AI where paper volume grew more than sixfold in two years, and the burden of low-quality and AI-assisted papers on volunteer moderators.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Starting October 1, 2026, arXiv imposes a cap of at most 2 submissions per submitter per calendar month. What this rate limit truly targets is not AI itself, but the structural mismatch: after AI pushed paper production speed higher, the volunteer review system cannot expand in sync.
The rule arXiv laid out is itself not complicated: each submitter may submit at most 2 papers per calendar month, effective October 1, 2026. But what truly determines the practical binding force of this rule are several easily overlooked details.
First, the restriction applies to all authors and all categories, not only to the AI field where submissions have surged, and the 2-paper quota is shared across all of arXiv, not 2 for cs.AI and another 2 for cs.LG. Second, the restriction counts submission attempts rather than the number of papers ultimately successfully posted—even if a paper is rejected, it still consumes one quota for that month.
Third, the restriction targets the submitter who actually submits the paper, not all co-authors of the paper. If a paper has 5 authors and A is responsible for submitting it, then this paper only consumes one of A's quotas and does not deduct from the quotas of the other 4. On this basis, arXiv also reminds co-authors that they need to coordinate who is responsible for submission.
arXiv also retains another restriction in place since 2024: at any time, at most 3 submissions may be in active status. With the two restrictions stacked, the operating space of high-frequency submitters is compressed simultaneously from two directions: monthly total volume and concurrent inventory.
The rule leaves one exit: if a paper is proactively deleted by the submitter before formal announce, that submission does not count toward the 2-per-month limit and does not occupy one of the 3 active submission slots. But for situations of rushing a conference deadline or having a backlog of papers on hand, arXiv currently has no plans to open a special channel, and the official advice is to plan submissions in advance.
The direct backdrop of the rate-limit order is a steep submission volume curve. In September 2016, arXiv received 9,869 submissions in a single month; by September 2024, this number reached 20,569; and in September 2026, 40,363. Over ten years it more than quadrupled, but the most recent doubling took only two years.
More noteworthy is the structural difference. arXiv specifically singled out the cs.AI category under computer science: from 2024 to 2026, submissions in this category grew more than 6-fold, while other arXiv categories overall roughly doubled over the same period. AI is the steepest segment of this round of growth.
The pressure of submission volume does not stop at the numbers themselves. Behind the 40,363 submissions, nearly 9,000 support tickets were created for arXiv staff and reviewers.
What needs to be distinguished is that the causal strength between submission volume growth and AI-assisted writing is only described qualitatively by the source, with no quantitative attribution given. What can be confirmed is synchronization in time: the spread of AI tools and the accelerated rise in submission volume occurred within the same window.
arXiv officially named several types of low-quality papers: thin papers with a very narrow scope and very thin content; salami papers that cut one complete piece of work into several small slices and submit them separately; and a growing number of papers that rely heavily on AI writing and have very high textual density.
There is a key distinction here: arXiv has not prohibited researchers from using AI. Under existing rules, as long as authors disclose it and the research itself meets arXiv's requirements for scholarly value and advancing the field, using AI to assist research is not a problem. What is being targeted is the side effect that AI greatly lowers the threshold for producing low-value papers.
In its announcement, arXiv specifically mentioned this change: before AI appeared, there was in practice a relatively easy-to-judge realistic upper limit on how many submissions one person could independently produce; now, this upper limit is increasingly difficult to judge.
In February this year, Nature focused on the same issue. Human-computer interaction researcher Raphael Wimmer used the AI tool Prism to spend only 54 seconds turning an experiment he had never actually done into a paper, and remarked that writing papers had never been so easy, and clogging the scientific publishing process had never been so easy. This case is a source paraphrase, used to illustrate the magnitude of the decline in the generation threshold.
arXiv is not an academic journal in the traditional sense; after submission, a paper does not need to first undergo complete peer review before it can be made public. But this does not mean that any randomly generated PDF can be posted directly—arXiv still needs to review whether submissions conform to the corresponding category, basic academic standards, and scholarly interest requirements.
And a large part of this system relies on volunteer reviewers. Once submission volume rises, this is where things first cannot keep up. This is a judgment that can reasonably be inferred from the source description: paper generation speed can expand rapidly, but review capacity is limited by manpower and time and cannot expand in sync.
Thomas Dietterich, chair of the arXiv Editorial Advisory Committee and professor emeritus at Oregon State University, mentioned that currently only a small portion of authors submit a large number of low-quality papers, yet consume a disproportionate amount of review time; the result is not only increased workload for reviewers, but other normal submissions may also be delayed by days or even weeks. This statement comes from a source paraphrase and is a qualitative judgment by a relevant party.
From this, the design logic of the rate-limit order can be understood: rather than passively accumulating on the review side, it is better to limit submission frequency at the entrance. The provision that rejected papers still consume quota exists precisely because as long as a submission enters the review process, it consumes reviewer time.
arXiv explicitly stated that this new rule is currently more like a temporary speed-limiting measure, not a final solution.
The first question that needs time to be rejudged is: after increasingly advanced AI tools enter scientific research, what exactly should be a reasonable paper production speed for authors? Before AI appeared, this speed had a relatively easy-to-judge realistic upper limit; now this upper limit has become blurred, and rules are therefore difficult to fix once and for all.
The second question is tools and processes: arXiv needs to upgrade its own review tools and processes, and then decide what long-term rules should be adopted.
Therefore, the positioning of this rate-limit order is to buy buffer time for the review side, rather than to make a ruling on AI-assisted writing itself. arXiv has not prohibited researchers from using AI, and this point has not changed in the rules.
From the confirmed rules, what rate limiting can directly compress is the upper limit on monthly submission counts for an individual submitter, as well as the number of submissions simultaneously in active status. For the situation in which a small portion of authors submit a large number of low-quality papers and consume a disproportionate amount of review time, this is a hard constraint at the entrance.
But whether the new rule can truly alleviate reviewer burden, and the problem of normal submissions being delayed by days or even weeks, the source does not give a conclusion. This is an effect judgment that has not yet been verified and requires observing actual data after implementation.
Another unresolved point is the coordination cost among co-authors. The rule only constrains the submitter; a paper only consumes the quota of the one submitter, which means teams need to decide on their own who is responsible for submission. The source only notes that coordination is needed and does not give a specific mechanism.
In addition, arXiv characterizes the new rule as a temporary measure, and long-term rules have not yet been determined. This means that the current 2-paper cap, the 3-active-submission cap, and details such as deletion not counting toward quota may all be subject to subsequent adjustment.
The factual basis of this article comes from a WeChat public account report (QbitAI, published October 2, 2026), which paraphrases the content of arXiv's official announcement. Among these, the specific provisions of arXiv's rate-limiting rules, effective date, submission volume data, cs.AI growth multiple, number of support tickets, officially named types of low-quality papers, AI disclosure rules, and characterization as a temporary measure are all based on arXiv official information as paraphrased by the source; Thomas Dietterich's statement and the case of the Prism tool generating a paper in 54 seconds are source paraphrases, treated in this article as source claims and not independently verified. The causal strength between submission volume growth and AI-assisted writing is only described qualitatively by the source, and this article does not make a quantitative attribution.
arXiv's 2-per-month rate limit appears on the surface to be a submission quota, but in essence, after AI pushed paper production into the fast lane, human review capacity has become the new physical bottleneck. The rule caps submission attempts rather than publication counts; rejections also consume quota; and the quota is shared across all of arXiv, not divided by category. Whether this temporary speed limit can alleviate pressure on the review side, the source does not give a conclusion; long-term rules depend on how arXiv redefines a reasonable paper production speed in the AI era.
Primary report
Primary source