PaperlayerView exact revision on arXiv

Source-linked AI summary

Is the arXiv Apocalypse upon us? A break in preprint growth coincident with agentic AI

Carlo R. Contaldi

arXiv:2610.34449v1physics.soc-phcs.DLstat.AP

TL;DR

The paper asks whether aggregate arXiv submissions show a distinct signature following the arrival of agentic AI. It analyzes submission counts and metadata using piecewise trends and author-concentration measures. It reports an early-2026 break separate from AI-category growth, accompanied by a shift toward single authorship.

  • Problem

    The paper asks whether aggregate arXiv submissions contain a distinct signature of the transition to agentic tools, without relying on LLM-detection methods.

    Source: §I · p. 1

  • Method

    The analysis fits a seasonally adjusted piecewise log-linear trend to submission counts and uses metadata to examine author concentration.

    Source: A. Trend and seasonal model · p. 2, D. Author concentration · p. 3

  • Results

    The early-2026 break is separate from AI-research growth, concentrated in desk-based fields, and accompanied by a shift toward single-author papers.

    Source: E. A shift to single-author papers · p. 6, §VII · p. 8

  • Takeaways & Limitations

    Publication counts in affected fields after late 2025 are no longer comparable with earlier counts, while the submission record alone cannot establish whether mathematics' additional papers represent more established results.

    Source: §VII · p. 9

    The method measures aggregate submission-volume response, not which papers used agentic tools or their quality, and an unrelated change affecting formal fields could produce a similar pattern.

    Source: C. Limitations · p. 8

Abstract

from arXiv · show

As an experiment in agentic assisted analysis, we used an agentic AI model to analyse the effect of agentic AI on arXiv submissions; data acquisition, analysis and the first draft were carried out by the model under the author's direction. Monthly arXiv submissions have grown exponentially for three decades, at about $9\%$ per year. Fitting a piecewise log-linear trend with seasonal modulation to the full record, we find two post-pandemic breaks in growth rate. The first, in mid-2022, precedes ChatGPT and is carried by the expansion of AI research within computer science (CS); outside CS, growth from 2022 to 2025 remained at $8\%$ per year. The second, between December 2025 and February 2026, follows the release of models capable of extended autonomous work and is absent from the AI categories themselves. Since December 2025 submissions without a CS listing have run $17\%$ above trend, and in September 2026 total submissions reached a record $40\,363$. The excess is largest in mathematics, where submissions doubled within a year, and follows a gradient from formal and theoretical fields to experimental ones that is not ordered by surveyed rates of AI-tool adoption. In mathematics the increase is distributed across subfields and authors, two-thirds of it arising from a larger number of distinct authors rather than from a few prolific submitters. Across fields the clearest signature is a shift to single authorship: solo-authored papers rose by $70\%$ within a year in mathematics and condensed matter, and by $44\%$ in quantitative biology even as its total output fell, but not in experimental particle physics. The analysis is deliberately agnostic of LLM-detection methods: it uses only submission counts and metadata, with treatment timing fixed by calendar dates, and is therefore free of the outcome-triggered timing bias of author-level studies.

I. INTRODUCTION

The paper asks whether aggregate arXiv submissions show a distinct signature of the shift to agentic tools, using counts and metadata rather than LLM-detection methods. Its data comprise monthly submission counts and category-level records, with cross-lists and overlapping groups handled explicitly.

I. INTRODUCTION

  • The analysis locates growth changes from the submission record before comparing them with agentic-model release dates, avoiding author-level treatment timing triggered by output.It uses submission counts and metadata, not LLM-detection classifiers or word-frequency signatures.Source: §I · p. 1

II. DATA

  • The aggregate series contains monthly new-submission counts from July 1991 to October 2026; October 2026 is excluded because it is incomplete and first subject to the per-submitter limit.That month is discussed separately.Source: §II · p. 2
  • Category-level API counts cover January 2022 to October 2026, include cross-lists, and can overlap; non-CS groups remove papers carrying any cs.* listing.Shared archive aliases affect classification, and API totals agree with published aggregate counts to within about 2%.Source: §II · p. 2
  • For mathematics author-concentration analysis, the study retrieved full metadata for 3,945 submissions in September 2025 and 8,203 in September 2026.Both samples comprise submissions without a computer-science listing.Source: §II · p. 2

III. METHOD

The method models log monthly submissions with piecewise growth and seasonal adjustments. Category groups are defined from API queries, with counts reported for September 2026.

A. Trend and seasonal model

  • The log-count trend allows growth-rate changes at knots, whose locations are denoted by τj and whose changes are denoted by cj.Time is measured in years and Θ is the Heaviside function.Source: A. Trend and seasonal model · p. 2
  • Category groups are defined by arXiv API queries, and all listed “No CS listing” groups exclude papers carrying any cs.* listing.The “No CS listing” aggregate is the total minus the CS group and is exact.Source: Table I · p. 2
  • Eleven monthly offsets capture recurring annual or fractional-year patterns, while a weekday term accounts for month-to-month differences in weekdays.The weekday term lowers the final aggregate model's BIC by 65; the fitted weekday elasticity is 0.87 for the aggregate.Source: A. Trend and seasonal model · p. 2

B. Break detection

Break detection searches knot configurations using BIC, with an adjustment for residual autocorrelation. The field comparison measures post-2025 excess and growth changes against a common pre-period trend.

B. Break detection

  • Candidate knot configurations are compared by BIC, and knot-location uncertainty is the range with ΔBIC below 2 from the optimum.Because final-model residuals have lag-one autocorrelation ρ = 0.21, the analysis also recomputes ΔBIC using the effective number of independent months.Source: B. Break detection · pp. 2–3
  • Figure 1 juxtaposes monthly submissions against a no-knot model fitted from January 2010 to November 2022 and shows trailing-year growth against historical annual-growth ranges.Dashed lines mark ChatGPT's public release and late-2025 agentic-model releases.Source: Figure 1 · p. 3

C. Field comparison

  • For each category group, excess is the mean residual from a January 2022–November 2025 no-knot trend over December 2025–September 2026.The study also compares year-on-year growth, searches for a group-specific knot, and uses the non-CS versus CS excess difference as a difference-in-differences estimate.Source: C. Field comparison · p. 3

D. Author concentration

The author-concentration analysis decomposes mathematics submission growth using author metadata, while monthly samples track solo authorship across four non-CS groups.

D. Author concentration

  • Mathematics paper growth is decomposed into changes in distinct authors, mean papers per author, and mean authors per paper.Author names are normalized by removing diacritics, punctuation, and case; 2026 statistics are also computed on size-matched subsamples.Source: D. Author concentration · p. 3

E. Author-count series

  • Monthly author-list samples cover mathematics, condensed matter, quantitative biology, and experimental high-energy physics from January 2022 to September 2026.The first two groups have 1,000 sampled papers per month; all papers are sampled for the two smaller groups.Source: E. Author-count series · p. 3
  • Figure 2 presents mean excess over each group's January 2022–November 2025 trend and year-on-year growth for mathematics without a CS listing, other non-CS groups, and computer science.The excess averages December 2025–September 2026; groups without a CS listing are blue and CS groups vermilion.Source: Figure 2 · p. 4

IV. RESULTS

The aggregate series departs increasingly from its long-run exponential trend, and a two-knot fit identifies a robust late-2025/early-2026 break distinct from the less certain mid-2022 break. The later break follows releases of models designed for extended autonomous work, though the association is not established as causal.

A. Aggregate series

  • 28–34% above extrapolation in January–September 2026, compared with 9–12% in 2025; September 2026 was 10σ above the fitted trend.The baseline fit grew at 9.0% per year over January 2010–November 2022.Source: A. Aggregate series · p. 4
  • 30.0% year-on-year growth in January–September 2026 was 4σ from the 1997–2022 mean, whereas 2024 and 2025 growth was individually unremarkable.The respective growth rates were 11.2% in 2023, 18.2% in 2024, and 16.6% in 2025; historical mean and standard deviation were 10.0% and 5.0%.Source: A. Aggregate series · p. 4

B. Breaks in the growth rate

  • A two-knot search improves BIC by a further 41 (25 with neff) over the one-knot fit, placing the first knot in mid-2022.A one-knot fit improved BIC by 83 (52 with neff) over pandemic knots alone; its preferred location was the turn of 2022–2023, with post-knot growth about 16% per year.Source: B. Breaks in the growth rate · p. 4
  • December 2025–February 2026 is the second knot’s interval, and excluding September 2026 leaves it there; the first break cannot be separated from pandemic recovery.The authors therefore regard the second break as robust, but not the first.Source: B. Breaks in the growth rate · p. 4
  • The second break follows releases of models designed for extended autonomous work; earlier capable models were not associated with detectable aggregate breaks.If causal, the response would have a threshold character, appearing with tools able to execute multi-step research and writing tasks rather than conversational assistance.Source: B. Breaks in the growth rate · p. 4

C. Field dependence

The two aggregate breaks have different field origins: the earlier acceleration reflects computer-science growth, while the later excess is concentrated outside CS. The response varies across fields, with mathematics leading and theoretical/formal work generally rising more than experimental or computational fields.

C. Field dependence

  • 8.1% annual growth without a CS listing versus 24% with one from January 2022 to November 2025 attributes the earlier aggregate acceleration to CS and AI research.The corresponding rates were 47% for cs.AI and 44% for cs.CL; papers without a CS listing remained consistent with the archive’s historical rate.Source: C. Field dependence · p. 4, Table II · p. 5
  • 17.1% above trend for papers without a CS listing contrasts with a 5.5% CS excess; cs.AI and cs.CL show no excess or 2026 break.The difference-in-differences excess outside CS relative to CS is +11.6 percentage points, and the authors characterize the 2026 surge as increased production in other fields, not continuation of the AI-research boom.Source: C. Field dependence · p. 5, Table II · p. 5
  • Mathematics had the largest excess at 27% (z = 21), contributing 13,726 of 28,041 additional non-CS submissions between January–September 2025 and January–September 2026.Its monthly year-on-year growth rose from near 15% through February 2026 to 108% in September; statistics, quantum physics, economics, and several other fields also had excesses, while quantitative biology had a deficit.Source: C. Field dependence · p. 5
  • The field ordering does not track surveyed agentic-tool adoption: mathematics had the largest response despite mathematicians reporting the lowest adoption in the available survey.The authors describe the ordering instead as following the extent to which research involves formal reasoning and writing rather than experiment or observation.Source: C. Field dependence · p. 5
  • 21.6% versus 10.6% year-on-year growth in 2026 for hep-th and hep-ex supports a theory–experiment gradient, but their trend-residual comparison is inconclusive.The hep-ex residual was +15% and hep-th +14%, but hep-ex’s fitted trend was nearly flat and noisy (1.7% per year), with low significance (z = 3.4 against 8.3).Source: C. Field dependence · pp. 5–6
  • September 2025–September 2026 mathematics submissions rose by factors of 2.9 in differential geometry and 2.0 in algebraic geometry, versus 1.6 in numerical analysis and optimisation and control.Other formal subfields increased by factors of 2.3–2.5, showing the increase spans mathematics subfields while computational subfields rose less.Source: C. Field dependence · p. 6

D. Concentration among authors

The mathematics increase was distributed across authors and subfields, with most logarithmic growth attributable to a larger author pool. Across several fields, solo authorship rose sharply after late 2025, though experimental particle physics showed no such shift.

D. Concentration among authors

  • 67% of mathematics’ logarithmic growth came from more distinct authors, while high-output authors’ papers remained a minority of the excess.The increase was spread across the fourteen largest subfields, whose paper-count ratios ranged from 1.6 to 2.9.Source: D. Concentration among authors · p. 6

E. A shift to single-author papers

  • Mathematics’ solo-authored share rose from 28.8% to 37.7%, with a September 2025 break about six months before its submission break.Condensed matter’s solo share rose from 7.6% to 11.8%, with a December 2025 break coinciding with its submission break.Source: E. A shift to single-author papers · p. 6
  • Solo-authored papers rose 70% in mathematics and 71% in condensed matter, while multi-author papers rose 33% and 13%, respectively.Solo-authored papers accounted for 46% of mathematics’ increase and 29% of condensed matter’s increase.Source: E. A shift to single-author papers · p. 6
  • Quantitative biology’s solo-authored papers rose 44% even as total output fell 3.4%, while multi-author papers declined 9%.The solo-fraction level shift occurred in November–December 2025; explanations for the multi-author decline were not tested.Source: E. A shift to single-author papers · p. 6
  • Experimental particle physics showed no solo-authorship break; solo-authored papers grew 13% over two years, versus 30% for multi-author papers.Its solo-authored fraction remained consistent with its earlier mean.Source: E. A shift to single-author papers · p. 7

F. The per-submitter limit

The per-submitter limit began on 1 October 2026, but early October counts showed no visible reduction in submission volume. The paper estimates that its potential effect is bounded by authorship metadata, which only identifies papers that might have been displaced.

F. The per-submitter limit

  • 31,672 October submissions were recorded by 23 October, already 14% above October 2025’s full-month total of 27,692, with no visible volume reduction in the first three weeks.The limit allows at most two new submissions per submitter per calendar month and three active submissions; rejected submissions count toward the monthly cap.Source: F. The per-submitter limit · p. 7

V. DISCUSSION

The discussion distinguishes the earlier AI-category-driven acceleration from a later rise in submissions outside AI categories. The later pattern follows releases of models built for extended autonomous work and varies across fields, but the analysis cannot identify AI-produced papers or establish individual-adopter effects.

A. Two phenomena

  • From 2023 to 2025, AI research expansion drove faster arXiv growth; from late 2025, papers in other fields rose sharply while AI categories did not.The discussion distinguishes the later phenomenon as the one associated with the submission limit.Source: A. Two phenomena · p. 7

B. What the pattern implies

  • The late-2025 pattern follows extended-work model releases by two to four months and is largest in fields centered on formal reasoning and text.Its field ordering does not match surveyed adoption rates; the authors argue this pattern is consistent with tools carrying research problems through to completion, while noting that timing and cross-field patterns do not establish causation.Source: B. What the pattern implies · p. 7

C. Limitations

  • The aggregate analysis cannot identify AI-produced papers or assess their quality, and an unrelated early-2026 change affecting formal fields could produce a similar signature.Calendar timing estimates population-level response rather than effects on individual adopters.Source: C. Limitations · p. 8

D. A data point

  • The model-assisted analysis took about 3.75 hours from first question to complete draft, including data retrieval, statistical analyses, figures, and text.The author estimates a comparable conventional analysis would have taken at least a working week.Source: D. A data point · p. 8

VI. OPEN QUESTIONS

The authors identify unresolved consequences of rising submissions for researcher training, peer review, and scholarly credit, then outline implications for moderation, evaluation, and judging mathematical output.

VI. OPEN QUESTIONS

  • Submission metadata omit career stage, so the authors cannot tell whether established researchers are replacing work that would previously have involved students and postdoctoral researchers.If so, effects on training may emerge slowly in publication counts and be difficult to reverse.Source: §VI · p. 8
  • If mathematics’ doubling extends to most fields, peer review that relies on researchers’ unpaid effort is unlikely to keep pace with submission volume.An estimated 21% of reviews submitted to a major machine-learning conference in late 2025 were generated entirely by AI.Source: §VI · p. 8
  • As human contributions shift from execution toward direction, the boundary between authored and generated work—and the basis of individual intellectual credit—will become harder to draw.The authors note that US law excludes AI-generated material without sufficient human creative control from copyright.Source: §VI · p. 8

VII. CONCLUSIONS

  • A per-account submission limit will not address the broad increase from authors each submitting one or two mathematics papers, so moderation capacity must scale with volume.Source: §VII · p. 9
  • Publication counts in affected fields after late 2025 are no longer comparable with earlier counts, requiring assessments that rely on them to account for the change.Source: §VII · p. 9
  • The submission record cannot establish whether mathematics’ doubled output represents a corresponding increase in established results; the authors leave that question to the mathematical community.Source: §VII · p. 9

Appendix A: Break locations by group

Table IV reports one additional break per category group in the January 2022–September 2026 series, along with the search window and fitted weekday elasticity.

Appendix A: Break locations by group

  • Each category group has one additional break location reported for its January 2022–September 2026 series.Source: Appendix A · p. 9
  • The procedure searches for that break from January 2023 through June 2026.Source: Appendix A · p. 9
  • Table IV also reports the fitted weekday elasticity w from Eq. (1) for each group.Source: Appendix A · p. 9
Loading 2610.34449v1…