AI breakthroughs draw over half of topic switchers, authors say

Beichen Lu and four co-authors said in an arXiv preprint submitted Oct. 19 that AI breakthroughs drew over half of topic-shifting researchers into artificial intelligence across 36 years of computer-science data.

It matters because the authors link post-Transformer concentration to lower peer-review confidence: reviewer confidence fell by 1 point on a 5-point scale, according to the Oct. 21 arXiv listing.

The preprint, which is not yet peer reviewed, examines 7.2 million papers, 73,089 peer reviews and 3.7 million researchers. The authors said the evidence points to a constraint on AI progress that is not captured by researcher counts, publication volume or investment alone.

“Breakthroughs depend on diverse attention, which collective concentration can deplete,” the authors wrote in the abstract. They said computer science expanded while research focus tightened after the Transformer, a neural-network architecture central to recent AI systems.

The authors describe a repeated cycle from AlexNet to large language models. In their account, each breakthrough increased the field’s “carrying capacity for ideas,” but larger breakthroughs were followed by faster depletion of the idea diversity available to subsequent work.

The finding cuts against a simple scale-based view of AI research, in which more people and more funding should reliably speed discovery. The preprint instead argues that attention itself can become a scarce input when researchers converge on the same dominant area.

The authors said the pattern was not limited to computer science.

The paper also points to peer review as a possible reinforcement mechanism. The authors said simulations showed that reciprocal reviewing, a policy in which authors review work in exchange for submitting or participating, can give more influence to prolific authors and narrow what the field accepts.

Those claims remain early. The arXiv record provides an abstract and bibliographic details, but it does not establish external peer review, independent replication or whether the reported patterns would hold under alternative measures of research attention.

The authors call for institutions to protect diverse attention in artificial intelligence rather than concentrate incentives around the leading paradigm. The arXiv record lists the paper under Artificial Intelligence, Computers and Society, and General Economics.

Source: arXiv preprint, arXiv cs.AI, Oct. 21, 2026