From Paper Authorship to a Reachable ML Researcher Shortlist

This is the runbook for turning academic and preprint authorship into a ranked, deduplicated shortlist of named ML researchers, each resolved to a current employer and a way to reach them. It is for technical sourcers, in-house recruiters, and talent leaders hiring research scientists when LinkedIn is empty and the only trace of the person is a paper. It gives the exact retrieval, enrichment, disambiguation, author-order, and contact-resolution steps in order, using only public publication records.

Other guides cover patent-to-inventor sourcing and reading ML engineers from model hubs. This one is different: it starts from a published paper and ends at a reachable person. That path has a specific hard part, and it is not the one most people expect.

Why the employer, not the name, is the hard part

The bottleneck in publication-based sourcing is resolving a researcher's current employer, not finding their name. The name is on the paper. The employer is usually not, or is stale, or is a free-text string nobody verified.

arXiv makes no attempt to collect affiliation in a canonical form. Affiliation is an optional, free-form text field supplied at submission and never normalized or verified, so most papers carry no affiliation data at all, and the ones that do can be years out of date. Meanwhile an estimated 70% of academic researchers have no presence in standard B2B contact databases. So the two tools a sourcer usually reaches for both fail here: the paper does not tell you where the person works now, and the databases do not have them.

That is why the enrichment layer is load-bearing rather than optional. Everything after retrieval exists to answer one question the paper cannot: where does this person work today, and how do I reach them.

Size the pool before you retrieve

Before you write a query, know how big the field is and where the people are. The pool is large in the US, roughly ten times smaller in the UK, and heavily concentrated in a handful of employers in both markets.

In Refolk's index of professional profiles, the US returns 3,230 research scientists and ML researchers tagged Machine Learning plus Deep Learning; the UK returns 323 of the same, a roughly ten-to-one ratio. Google DeepMind is the top employer in the US pool; Meta leads the UK pool, appearing in 10 of 25 sampled profiles. The practical consequence: a US shortlist is a genuine market search, while a UK shortlist is really a poaching map of four or five employers, so plan outreach accordingly.

Dataset A - ML researcher pool by market (from Refolk's index).

Counts, employers, and regions come from Refolk's index for title Research Scientist or Machine Learning Researcher with skills Machine Learning and Deep Learning. The ratio is derived.

The venue map matters as much as the market map. Venue selectivity is stable even as scale explodes, so acceptance alone is a weak filter and oral or spotlight status is the durable one.

Dataset B - Top-venue scale and selectivity, 2024 (from web).

NeurIPS and ICML figures from the Conference-Acceptance-Rate repository; ICLR from the published ICLR analysis. All three sit in a 25 to 31 percent band despite counts doubling over recent years, which is why you rank on orals and spotlights rather than acceptance.

Choose categories deliberately

Pick your arXiv categories with intent, because the category growth is uneven and misleading. cs.AI grew 6.1x from 2020 to 2025 while cs.LG grew far more slowly, so leaning on cs.AI alone over-weights recent entrants and misses the deeper bench.

The gap is a labelling artefact, not real expansion of the field. cs.AI grew from 7,417 submissions in 2020 to 45,133 in 2025, roughly 43 percent compound annual growth, and October 2024 alone set an arXiv record of 24,226 new submissions. cs.LG, the older machine-learning category, grew at only about 12 percent CAGR from a 2020 base of 25,889. The post-ChatGPT influx of agent, reasoning, and multimodal work filed under the cs.AI umbrella. If you want senior, longer-tenured researchers, cs.LG and cs.CL carry the bench; if you want the newest entrants, cs.AI catches them. Use both.

Dataset C - arXiv category growth 2020 to 2025 (from web).

Figures from the arXiv volume analysis; the multiple is derived from those figures.

The procedure, in order

Here is the full pipeline from brief to shortlist. Each step names who owns it and what done looks like. Expect a full pass to take one to two days for a focused brief.

Reading author order without misreading it

Author order in ML and CS follows a first-last convention: the first author is the lead contributor and usual hire target, the last author is the senior PI who advised the work, and the corresponding author is whoever holds the email. In one analysed corpus, 84% of 2016 papers had the first author as the corresponding author, so the lead contributor is often reachable directly. But some venues borrow alphabetical ordering from mathematics, which breaks seniority reading entirely and must be discarded when interpreting a career trajectory. Always check whether a paper or venue orders alphabetically before you assign contributor rank.

Disambiguation: the fallback chain

Deduplicate researchers with a three-level fallback chain: prefer the OpenAlex author ID, fall back to ORCID, and use normalized name only as a last resort. Then audit every common name by hand.

The order matters. OpenAlex assigns a unique author ID that studies have adopted as the definitive identifier for consistency, and it uses ORCID, publication records, and citation histories as features to disambiguate. So the OpenAlex ID goes first. ORCID goes second because it is only populated for roughly 55.7% of author-paper pairs, and because ORCID stub records can carry the wrong ID. Normalized name goes last because automatic disambiguation introduces errors for common names, name variants, and cases lacking ORCID, an effect especially pronounced for Chinese names.

That ORCID coverage sets a hard ceiling. About half of author-paper pairs have no ORCID, so your pipeline must fall back to name-plus-coauthor matching for those, and that half concentrates the collision risk. When two profiles share a name, confirm the match through Google Scholar co-authors and topic continuity: a real match shares collaborators and a coherent research thread across papers. An outlier profile with an implausible H-index or a five-way name split is a sign the disambiguation lied.

Because self-reported skill tags lag the field, this publication-based approach reaches people that tag-based tools cannot. In Refolk's index, the explicit Large Language Models skill tag returned only 2 US matches and 0 at senior level, even though the field is full of senior LLM researchers. The reason is simple: people do not update their tags for a subfield that is two years old. Reading what someone actually published sidesteps that entirely.

When the pipeline above is more retrieval plumbing than you want to build, Refolk runs the same publication-to-person logic behind a plain-English query and hands back named researchers with a resolved affiliation. It does not remove your obligation to verify employer and contact, but it collapses steps two through five into one search.

How this goes wrong

Publication-based sourcing fails in predictable ways, and every failure produces a plausible-looking result that is quietly wrong. Treat each row below as a check you run before the name enters the shortlist.

Two of these deserve extra weight. The first is affiliation treated as current, which is the single most common error because the field looks authoritative and is not. The second is the lawful-basis skip, which is not a data-quality problem but a legal one.

Resolve the contact path lawfully

Resolve contact from public sources in priority order, and wrap the whole thing in a documented lawful basis before you send anything. The most reliable source of a researcher's current email is their most recent publication; ORCID is the fallback; an institutional pattern is a last resort you must verify.

ORCID is used by more than 20 million researchers, many of whom link their institutional email, employment history, and publication list to their profile, which makes it a strong second source. Funder databases such as NIH Reporter list PI names, institutions, and project details for active grants, confirming current institution and department. Where none of these yields an address, most universities follow predictable email patterns, but pattern inference is a guess you must confirm against a stated corresponding email.

1. Corresponding-author email in the most recent paper PDF or version of record
2. Email linked on the researcher's public ORCID profile
3. Institution and department confirmed via a funder database (e.g. active grant record)
4. Lab or group page listing a direct contact
5. Verified institutional email pattern (only if confirmed against a stated address)

On lawful basis: for sourcing candidates and matching them to roles, the correct basis in the UK and EU is legitimate interest, supported by a documented Legitimate Interest Assessment. Under GDPR you must inform a sourced candidate that you hold their data and intend to process it, typically within 30 days of collecting it and before you use it for any purpose. This is not optional caution. In October 2024 the Irish Data Protection Commission fined LinkedIn EUR 310 million over invalid consent and legitimate-interest claims, the exact profiling-without-basis failure this step prevents.

Before you call the shortlist ready

Run this checklist before the list leaves your hands. It catches the errors that survive the pipeline and look correct.

Keeping the shortlist current

A publication-based shortlist decays the moment a researcher moves labs or a new preprint drops, so treat it as a living artefact, not a one-time export. Re-run the employer-resolution step on a cadence, because that is the field that goes stale first.

Set a recheck rhythm tied to the two things that change: affiliation and output. For affiliation, re-resolve the current employer on any candidate you have not contacted within a quarter, since a mover will show a new latest paper or an updated ORCID employment record. For output, re-run your retrieval query against the most recent submission window; October 2024's record of 24,226 new submissions in a single month shows how fast fresh first-author work appears. When a candidate publishes a new top-venue first-author paper, their rank should rise, and when they move from academia to a lab you were targeting, their reachability and interest both change.

Keep the raw provenance for every field: which paper gave the name, which source gave the employer and its date, and which source gave the contact path. That record is what lets you re-verify quickly instead of rebuilding, and it is what your Legitimate Interest Assessment points to when you have to show where the data came from. A shortlist you can re-audit in an hour is worth more than one you have to regenerate from scratch every time someone asks how you know.