THE
CONFERENCE
SIGNAL.
54,721 researchers. The 1,726 who never miss a cycle. And the ones already founding the labs you compete with.
Built from public publication data: every accepted NeurIPS and ICLR paper, 2023–2025.
THE MARKET YOU CAN'T SEE
Hiring an AI researcher right now means competing for people you cannot see. They are not on job boards. They are heads-down on the next paper, and by the time a name reaches you they have already signed or left to start a lab.
The talent still leaves a trail. Every serious researcher signs their work. The papers they publish at the venues where the field judges itself show who does the work once and vanishes, who does it on repeat, who is accelerating, and who has gone quiet to raise a round.
SO WHAT?
The data trail exists whether you use it or not. The labs winning the research talent war built maps like this before they opened a search. If you are hiring from published research networks without a recurrence filter, you are working a list of 54,721 when the actual target is 1,726.
THE SIGNAL
Rank every researcher by how many of the six cycles they appear in and the population collapses into a steep curve. A single paper proves a person can do the work once. The fixtures prove they do it on repeat.
The only signal that survives the noise of a conference that drew 21,575 submissions in 2025.
SO WHAT?
A single paper is noise. Three or more cycles is signal. The difference between the 1,726 fixtures and the 36,792 one-and-done tail is not just output: it is reliability, depth, and the probability they are already on someone else's shortlist. Filter to recurrence first.
THE FOUNDER FACTORY
Read the top of the recurrence list and you are reading a cap table. These labs were founded by recurring names in this dataset, with well past five billion dollars raised between them in the last two years.
The money is pooling in four directions: world models, embodied learning, diffusion models, and AI security.
If you are hiring into world models, robotics, diffusion, or AI security: you are bidding against a freshly funded company started by the person at the top of this list.
SO WHAT?
The most cited names in this dataset are not candidates: they are competitors. Levine's lab and Ermon's lab are hiring from the same proceedings you are. Every dollar they raise is a dollar going toward pulling the rising fixtures before you can reach them. Know the founder map before you open a search.
THE RISING STARS
The founders are the present. The next ones are accelerating fastest, not yet locked into a lab or a cap table.
Not on anyone's radar three years ago. Now publishing like fixtures. Screen these first: the all-six fixtures are not reachable.
SO WHAT?
The 392 breakouts are the most actionable segment in this dataset. Not yet locked into tenured positions. Not yet founding. Publishing fast enough that their trajectory is visible. These are the researchers you can still reach, but the window is measured in months, not hiring cycles.
SUPPLY VERSUS DESTINATION
China overtook the United States as the largest contributor of papers in 2025, with Tsinghua the most prolific university. But contribution and elite concentration are two different things. China is the supply. The West is the destination.
of the Chinese-origin recurring authors are now at non-Chinese labs: Berkeley, CMU, Stanford, NUS, NTU, Google.
The recruiting opening is the gap between where this talent is trained and where it is employed.
SO WHAT?
Half the elite was trained in China and most now works in the West. That gap: 56% of Chinese-origin recurring authors at non-China institutions is the least-contested hiring surface in AI research. If you can hire globally or sponsor, you are competing for a pool that almost nobody is working.
WHERE THEY TRAINED, WHERE THEY SIT
The recurring talent comes through a narrow set of graduate programmes, then concentrates at an equally narrow set of employers. The geography is a barbell: established American powerhouses on one end, Chinese and Singaporean institutions rising fast on the other, and very little in between.
MIT, Stanford, CMU, Berkeley dominate AI doctoral training, but Tsinghua, USTC, SJTU, Peking, and Zhejiang now host recurring authors. The drain is starting to reverse.
SO WHAT?
The drain from Chinese institutions is accelerating. Tsinghua, USTC, and SJTU are producing recurring authors at NeurIPS and ICLR, and those authors are landing at Berkeley, CMU, Stanford, and Singapore. The search starts upstream, at the programmes that feed the labs.
WHAT A FOUNDER SHOULD BE WATCHING
THE FOUNDER EXITS
When a fixture leaves academia, a lab and a hiring spree follow within months. Each departure drains a sub-field and opens a backfill.
THE FOUR FUNDED THEMES
World models, robotics, diffusion, AI security. The talent and the capital are converging there, and so should your search.
THE RISING 392
The breakout authors are the next founders and the most hireable researchers on the list, reachable now, and not for long.
THE CHINA GAP
Half the elite is Chinese-origin and most is employed in the West. If you can hire globally or sponsor, it is the least-contested pool on earth.
THREE MOVES TO MAKE NOW
HIRE THE RECURRENCE, NOT THE SINGLE PAPER.
Three-plus cycles separates the 17,929 who keep producing from the 36,792 who appeared once.
Filter your conference sourcing to repeat authors, and ignore the one-and-done tail.
SKIP THE ANCHORS, WORK THE RISING MIDDLE.
The all-six fixtures are tenured or already founding. The hireable signal is the 392 breakouts and the recurring not-yet-tenured.
Screen for trajectory, and start with the steepest curves.
REACH THEM BETWEEN PAPERS.
A rising fixture is being courted off these same proceedings by Meta, OpenAI, and DeepMind, or is about to found.
The window on a rising name is months, not years.
WE KNOW
THE 392.
DO YOU?
Most of the 54,721 are already placed, tenured, or founding. The approachable cohort is the rising middle: breakout authors not yet locked in, not yet on a cap table. We know who they are in your specific research area, what they'd move for, and who else is already approaching them.
Tell us the role: the research thread, the level, the location. We'll show you who from this dataset fits, who's approachable, and what the competitive landscape looks like before you start.