Global distribution of AI research
646,589 arXiv papers posted since January 2016
Papers per year
Ranking
How we count: Method
646,589 arXiv papers posted since January 2016
How we count: Method
How we count: Method
How we count: Method
How we count: Method
This site maps where AI research is done, using every AI paper posted to arXiv since 2016.
We start with every paper first posted to arXiv [1] since 1 January 2016 in cs.AI, cs.LG, cs.CL, cs.CV, cs.NE, cs.RO, cs.MA or stat.ML, as its primary category or a cross-list. We annotated the first page of every paper with Claude Haiku 5.5, which lists each author's affiliations as printed. Any affiliation that doesn't appear in the page text is dropped.
We match the affiliations to the Research Organization Registry [2] and place each one in the country the paper prints. So a lab counts where it is, not where its company has its headquarters.
Each paper counts once, split equally among its authors and then between each author's first two affiliations [3]. A paper by two MIT authors and one Tsinghua author gives 0.67 to the US and 0.33 to China. About 5% of the credit can't be placed in a country. Most of it comes from authors whose first page names no affiliation, companies that print no place, and independent researchers.
The same model gives each paper one of 20 subfields from its title and abstract. Main conference papers are those in the main track (workshops are excluded) of 18 conferences: NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, ACL, EMNLP, NAACL, COLM, AAAI, IJCAI, KDD, SIGIR, WWW, ICRA, IROS and RSS. They are matched by title and authors to the official proceedings, from DBLP [4] and the conferences' own lists. Citations come from Semantic Scholar [5]. A highly cited paper is among the 10% most cited papers of its year and subfield [6].
Researchers are told apart by Semantic Scholar's author IDs [5], corrected with DBLP's author profiles [4]. Their moves between institutions are traced from the affiliations on their papers [7]. Population and GDP come from the World Bank [8], R&D spending from the OECD [9] and UNESCO [8], and Taiwan's GDP from the IMF [10].
New papers are added daily, the night after arXiv announces them (last update: 31 October 2026). Citations, researchers and moves are refreshed monthly (last update: 27 October 2026). Conference papers are added as proceedings come out (last update: 26 October 2026).
The dataset is free to download from Hugging Face as Parquet and CSV. It holds about 320 MB:
Emails, page text and the tables about individual researchers are not included. arXiv's metadata [1], ROR [2] and DBLP [4] are CC0, and ROR's locations come from GeoNames [11] under CC BY 4.0. World Bank data [8] is CC BY 4.0, and NSF's university R&D data [12] is free to use.
Issues or questions? Contact me on X: @ArtKanke.