Math papers on arXiv are 2.24 times the expected number, while machine learning shows no abnormal acceleration

Jasper Dekoninck posted a set of paper counts on X. The data are grouped by month of first submission on arXiv and fitted with an exponential trend including a month term: exp(a+bt+c_month). First, the totals: in September 2026, arXiv had 40,363 papers in reality versus a fitted expectation of 25,382, or 1.59 times.

Mathematics is the most extreme. In September, there were 8,167 papers in reality versus an expected 3,639, or 2.24 times. In the figure, the blue line is the actual monthly paper count, the orange line is the fitted trend, and the dashed orange line is the forecast for 2026.

Computer science overall had 17,747 papers versus an expected 13,166, or 1.35 times. The increase is not in cs.AI and cs.LG. In the replies, it was noted that within CS, computational complexity, algorithms, and game theory grew the most.

cs.LG basically hugs the expected line. Dekoninck says that machine learning was already growing very fast before AI agents appeared; agents can automate a lot of theoretical work, but fields that need resources such as compute have not accelerated noticeably.

Biology is the other side. bioRxiv had 5,049 papers in September in reality versus an expected 5,800, or 0.87 times. COVID is a clear outlier, and paper counts have not significantly exceeded trend. His hypothesis is that experiments are more of a bottleneck than knowledge work.

Divergence is also visible in other fields: physics 1.44 times, statistics 1.63 times, medicine and health 1.23 times, economics 1.15 times, quantitative finance 1.13 times, electrical engineering and systems science 0.97 times, and quantitative biology 0.98 times.

If you only look at AI news, it is easy to think machine learning papers are the most explosive. This dataset suggests the opposite: math and theoretical CS are growing faster. One explanation is that agents lowered the marginal cost of knowledge work, while resources such as experiments, compute, and data have not been automated. This judgment is not conclusive, but the divergence is already written in the charts.

Data and interactive charts on his personal website: https://jasperdekoninck.xyz/research-output/ ; original post: https://x.com/j_dekoninck/status/2113962730025865452 .