The Generative AI Gold Rush in Theoretical and Computational Research
Xiaoshn Nee, Haobo Zhong, Xiaomin Ni
cs.DL
2026-09-04
Math listed 47,127 arXiv entries in Jan–Aug 2026, 33.5% above 2025 and 11.9% above a synthetic control of 42,113. Authors with ≥5 submissions rose from 2.45% to 3.80%.
Generative models draft proofs, write code, search literature, and turn intermediate results into preprints. Output rises, but "more papers" mixes four objects: platform traffic, a field's divergence from other fields, who is submitting repeatedly, and knowledge value. Lump them together and one cannot separate post-ChatGPT platform acceleration from a mathematics-specific extra.
This paper treats the 2026 arXiv math-list surge as an identification problem. The main estimand is not "the AI share." It is how far mathematics in January–August 2026 sits above a weighted combination of other archives fitted through 2025. AI-related stories are triangulated separately and are not baked into the synthetic control.
Twenty top-level archives, January 2018 through August 2026, a balanced panel of 2,080 archive-months. List counts measure field exposure, including cross-lists; unique site-wide submissions come from the official series. The core design is a ridge-regularized convex synthetic control on seasonally adjusted log counts, trained through December 2025, with January–August 2026 held out. λ=0.01, preperiod RMSE 0.034, effective donor count 12.09. math-ph is kept out of the main donor pool because of cross-listing. Diagnostics include spatial placebos, prior-year pseudo-holdouts, donor and preperiod swaps, and leave-one-donor-out. A specification is "qualified" only if its preperiod RMSE is at most twice the primary fit.
An author panel pulls first submissions with any math. category in matched January–August windows from 2023 to 2026 via the Atom API, stores truncated SHA256 digests of normalized display names, and flags high output at five or more submissions. Mechanism evidence cites existing corpus estimates of AI-shaped text and capability results on formal proof and olympiad systems, and does not enter the control.
The math list rose from 35,307 entries in January–August 2025 to 47,127 in 2026, +33.5%. Broad physics lists rose 16.5%; unique global submissions rose 26.8% (181,595 to 230,322). Extrapolating math's own 2018–2022 trend gives a 36.8% excess. The synthetic control is 42,113, a 11.9% gap. Qualified donor and preperiod designs span 9.6% to 14.9%. Leave-one-donor-out sits between 11.3% and 13.4%. Fitted annual gaps are −0.2%, 0.7%, 1.6%, −1.9% in 2022–2025, then 11.9% in 2026. Among 15 archives that meet the fit screen, math has the largest holdout-to-preperiod RMSPE ratio; the finite-sample probability is 1/15=0.067. Prior-year pseudo-holdouts were 1.1%, 1.9%, and −6.6%.
Twenty-nine of 30 primary math. categories grew. Combinatorics (math.CO) added 1,593 submissions, 17.7% of the increase, and grew 57.4%, the fastest among categories that already had 1,000+ submissions in 2025. The ten largest contributors hold 74.3% of the rise. Growth is broad, not a single-topic spike.
Unique math-tagged submissions rose 33.6% (31,613 to 42,236), almost the same as the list. Active author units rose only 17.2% (50,302 to 58,930), so submissions grew about twice as fast as people. Solo-authored papers rose 54.7% (8,450 to 13,070), share 26.7% to 30.9%. Units with at least five submissions went from 1,234 to 2,241, 2.45% to 3.80% of active units; the ten-plus tail from 107 to 289, 0.21% to 0.49%. The maximum rose from 20 to 44. Gini moved from 0.272 to 0.313. High-output units with no matched activity in the prior two windows rose from 90 to 267. The tail thickened through both intensified output among established names and entry from low prior visibility.
For anyone watching arXiv, 2026 math really is a new volume regime, and it outran physics and the site as a whole. After the synthetic control removes shared platform acceleration, a math-specific gap of about one tenth remains, and it shows up in 2026, not in 2023 when ChatGPT became public. That timing fits delayed diffusion and a capability threshold better than a one-shot shock.
The governance claim is that candidate production is getting cheaper faster than verification. arXiv tightened endorsement in late 2025 and January 2026, which is direct evidence that moderation hit a wall. For AI4Math, the paper does not count theorems written by models. It counts traffic and author structure: a thicker repeat-output tail, more solo papers, almost every category up. The next measurement step is to join disclosures, formal checks, corrections, and citations onto the same manuscript IDs.
The comparative design identifies an anomaly, not an AI share. Reading 11.9% as "AI caused 11.9%" is a misread.
The authors are explicit. The control absorbs platform and technology changes shared with donors; donors may themselves be AI-pushed, so the math-specific gap may be a lower bound or a net, depending on contamination. A 1/15 placebo rank is 0.067, short of a conventional 5% cutoff. The panel stops in August 2026; the last four months could change the size.
List counts include cross-lists; an audit of 100 March 2026 math-list records found five with a non-math primary archive. Author units are display names, not people: variants split, homonyms merge. Triangulation uses other groups' corpus AI-trace estimates and capability evals, not manuscript-level tool disclosure and verification on this panel. Three independent researchers, with generative models used for literature, code, figures, and language under human direction. Treat "gold rush" as a metaphor. Treating it as a causal label goes beyond the design.