How Fast Is Machine-Learning Research Growing? arXiv Submission Counts 2007–2025, Growth-Curve Model Selection, Out-of-Sample Tests, and a Bayesian Change-Point Analysis
Year: 2026
Type: preprint
Abstract: I retrieved yearly counts of arXiv papers listed in cs.LG, cs.AI, cs.CL, cs.CV and stat.ML (cross-lists included) for 2007–2025 from the arXiv API on 18 October 2026. cs.LG grew from 84 papers (2007) to 46,022 (2025). Mean log growth was 0.42 per year in 2008–2015 and 0.47 in 2016–2020, then 0.12 in 2021–2025; the doubling time lengthened from about 1.5 years to 6.0. In sample, a logistic curve fits best (AICc −56.2, against −48.8 for Gompertz and −34.7 for exponential). Fitted to 2007–2019 only, however, the logistic collapses to the exponential and over-predicts 2025 by a factor of 4.4. Saturation was not identifiable before it began. A conjugate Bayesian single change-point model on log-growth gives weak evidence of a break in cs.LG (log Bayes factor between −0.7 and 0.9 depending on the prior; the modal break is 2020–21) and strong evidence for cs.CV (log BF 4.1). Over the same period cs.AI and cs.CL grew faster than cs.LG, so the cs.LG slowdown does not imply a slowdown of AI research on arXiv overall.
Author: Aguru, Manoj Kumar
Language: English
Subfield: Statistics, Probability and Uncertainty
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222903816
Cites:
Cited by:
Related to: