arxiv.org web signal

NeurIPS paper: US and Chinese labs barely cite each other

TL;DR

  • American papers are 60% of the 2012–2021 NeurIPS dataset but just 34% of the papers Chinese institutions cite.
  • Chinese papers are 34% of the dataset but only 9% of citations from American institutions.
  • Self-citation rates: USA 41%, China 21%, Europe 14%, with American labs preferring European over Chinese work.

American papers make up 60% of a decade's worth of NeurIPS output but only 34% of the papers Chinese institutions cite. The reverse is sharper: Chinese papers are 34% of the dataset and 9% of American citations.

That is the finding of a short paper by Bingchen Zhao, Yuling Gu, Jessica Zosa Forde and Naomi Saphra, built from NeurIPS proceedings 2012 to 2021 and presented at the venue's 2022 workshop on Cultures of AI. The abstract is blunt: "American and Chinese institutions cite papers from each other's regions substantially less than they cite endogamously."

Self-citation tells the same story from inside each bloc. Chinese institutions cite other Chinese institutions 21% of the time; American institutions cite other American institutions 41%; European institutions 14%. American labs reach for European work more often than Chinese work.

The paper points at concrete things that get lost. PCANet, a Chinese computer-vision method, has around 1,200 citations "largely from Chinese or East Asian institutions." On research direction, the authors describe multi-object tracking as "an active area of research in China" while "In North America, criticisms regarding the misuse of biometric data have led researchers to avoid related tasks."

The dataset ends in 2021. The authors treat language, politics and topical divergence as co-present causes rather than ranking them.

Shared on Bluesky by 1 AI expert