The Finishing-PhD Playbook: Market Pages to a Dated Candidate Slate

This playbook is for in-house recruiters, sourcers, talent leaders, and founders who need to hire a research scientist or research engineer straight out of a PhD. It takes you from a role brief to a ranked slate of at least 15 finishing candidates in ten working days. Every candidate on the slate has an earliest start date, a stated preference, and a first touch timed to land before faculty offers do. A finishing candidate here means a doctoral student whose defense is close enough to plan a start around. It also covers the postdocs that department pages list alongside those students.

Why does the finishing-PhD pool need its own pipeline?

The finishing-PhD pool runs on three things ordinary sourcing ignores: a fixed annual supply, a calendar set by the faculty job market, and advisors who act as gatekeepers. If you treat it like any other candidate search, you reach people after their faculty offers arrive.

The supply is real and growing. CRA's Taulbee Survey, the annual census of North American computing departments, recorded 2,352 computing PhDs awarded in 2023-24, up 8.2 percent from 2,173 the year before. In the 2025 Taulbee sample, CS doctorates reached 1,909, up 51 percent over five years and 16 percent in one year.

Most of these graduates go to industry, but not all of them, and the split moves from year to year.

The rows are different cuts from the CRA Taulbee reports, so compare a row with its neighbours with care. In the 2025 cross-section, industry hires outnumber academic hires about 1.9 to 1. In the longitudinal series, academia peaked at 40.1% in 2024, the same year industry fell to 54.8%. When industry hiring cools, more finishing candidates aim for academia. That matters for how you classify them later.

Two more facts shape the brief:

• Most industry hires take research roles. About two-thirds of doctoral graduates who went to North American industry took research positions, according to the 2022 Taulbee report.
• AI dominates. AI/ML was the most popular specialty in the 2025 report, at more than a quarter of degrees where the area was known. Historically, AI accounts for 29 percent of new computing PhDs employed in industry.

The industry-versus-academia split for AI/ML graduates specifically sits in Taulbee Table D4. I have not verified those cells. If your brief depends on that ratio, pull Table D4 directly from the current Taulbee report.

Which departments and labs are worth working?

Work the 15 to 25 departments that score best on three things: degree output in your specialty, how often their students appear on your target venues' accepted-paper lists, and whether they sit in a city where the pool clusters. Everything after this step is per-department work, so a bad list costs you days.

Translate the brief first

Before scoring anything, turn the role brief into one or two Taulbee specialty areas and three to five venues. A venue here means a peer-reviewed conference whose accepted-paper list you can download. The venue list is your best filter, and it's worth knowing how big each one is:

A venue can't have more first authors than accepted papers, so NeurIPS 2025 has at most 5,290 first authors, and many of those are not students. Nobody publishes the number of unique student first authors per venue. Growth also makes counts less meaningful: NeurIPS submissions rose from 9,467 in 2020 to 21,575 in 2025. Two first-author papers now say less than they used to.

Check where the pool actually sits

Refolk's index of professional profiles gives a rough size for each slice:

Don't read the NLP row as a real headcount. It almost certainly reflects how people label the skill, not how many NLP students exist (see failure modes below). The geography is more useful. In the index samples, Toronto and Montreal held 11 of the 25 sampled Canadian ML profiles, the University of Toronto led the Canadian sample, and Stanford accounted for 5 of the 25 sampled US NLP profiles. Pick departments by city as well as by reputation.

Score each department on these signals

For degree output by department, an open thesis index gives you a local proxy: count last cycle's deposited theses by university and subject. Treat it as approximate. The lag between defense and deposit is not documented.

What do department job-market pages publish, and what does each field prove?

A department job-market page is a public list of the PhDs and postdocs a department is putting on the market, and many already publish the exact fields a sourcer needs. Copying those fields word for word is the core of the harvest.

Two examples show the range:

• UChicago CS, "On The Job Market": Seeking, Research Area, Advisor, Finishing Date, plus email and website.
• Northwestern CS, "Outgoing PhDs and Postdocs": Advisor(s), Research Area(s), Thesis Topic, Planned PhD Defense Date, Employment Preference.

Northwestern's page already showed defense dates running to Summer 2027, which suggests pages for a given cycle go live by about November. That is one observation, not a rule. Nobody has established how many top departments publish such a page, or the usual month they go live. Check every department on your list in early fall, then again in November.

These pages mix postdocs into the student list. Flag every postdoc in a separate column, because the OPT dating rule below does not apply to them.

Thesis records do a narrower job than people expect. OATD (Open Access Theses and Dissertations) takes metadata from over 1,100 institutions, indexes 7,410,865 theses and dissertations, and searches titles, author names, abstracts, subjects, and university or publisher. A finishing student's doctoral thesis is not deposited until after the defense. For a live candidate, the index mostly turns up earlier work and the lab's recent graduates, which tells you what the group produces. Nobody documents how complete the advisor and defense-date fields are. Spot-check a few known graduates from one lab before relying on either field.

When does the faculty market move, and where is the industry window?

The faculty market runs on a fixed seasonal cycle: applications in the fall, remote interviews in winter, offers from about February with short deadlines. The industry window is the stretch before those offers land, when splitters are still weighing their options.

Three candidates' published accounts describe the same shape:

• Applications in the fall, with deadlines clustered in November and December, remote interviews in winter, and in-person interviews through spring.
• Applications generally due in early December, with responses most often in January or February.
• One CS candidate's dates: first screening call Dec 18, first interview Jan 21, first offer Feb 9.

Faculty offers often carry deadlines of 2 to 4 weeks. Aaditya Ramdas notes that Stat departments "start pressuring you to decide" while CS interviews are still in flow. On the industry side, one ML PhD's own job search had interviews in January through early February. A robotics candidate called starting industry applications on Dec 3 "too late". How much earlier an industry offer can land than a faculty offer is not established as a number. Plan as if it has to arrive before February.

An industry offer that arrives after a faculty offer is competing against a deadline, not a decision.

Advisor first or candidate first?

The sources disagree. Ed Lazowska's "The Ph.D. Job Hunt" says "the advisor must bear the major responsibility" for placement. Another account notes that students "post on social media, present at workshops or conferences" to signal for themselves that they're on the market. My rule is to follow the signal. Write to the candidate first when they have publicly said they're on the market or listed an industry preference. Ask the advisor first when the candidate is academia-first or the lab has no page. Advisor asks also bring up students who never listed themselves anywhere.

The ten-day procedure

The procedure runs over ten working days from brief to ranked slate, with each step ending in a concrete artifact. The durations are desk estimates. Nobody has published how long each step takes a sourcer, so time your own first run and adjust.

Steps 3 and 4 take the most time. Market pages give you the people who listed themselves, and venue lists add the people who didn't. Matching a first author on a venue list back to a department, an advisor, and a current title by hand is slow, so most sourcers stop at the market pages. Refolk removes that bottleneck: describe the slice in plain English and it returns matching people from GitHub, LinkedIn, and the open web. Then check each one against the market page and the paper's affiliation string.

Run that kind of query once for each department or lab on your scored list, and add any name that isn't already in the sheet. Those additions tend to be the most valuable rows, because they are the candidates nobody else found on a market page.

Use one row per person and keep a source URL for every field you fill in.

Name
Department
Lab and advisor
PhD or postdoc
Research area / thesis topic
Matches brief specialty (Y/N)
Finishing or defense date
Date source URL
International, OPT clock applies (Y/N/unknown)
Earliest start date
Stated preference (verbatim)
Class (industry-first / splitting / academia-first)
First-author papers at target venues
Affiliation string matches department (Y/N)
Current title checked on (date)
First touch sent (date)
Advisor ask sent (date)
Next touch due (date)
Rank

The first touch to a candidate should show you read their work and name the timing problem directly. The advisor ask should cost the advisor almost no effort.

--- Candidate first touch ---
Subject: Research role after your {defense term} defense

Hi {first name},

I read {paper title} at {venue}, and saw your {thesis topic} work on {department}'s job-market page. I'm hiring a {role} at {company} to work on {problem, one line}, and {specific result} is close to what the team needs.

I know you're planning around a {defense term} defense and may be on the faculty market. I can run our process in time to give you a written offer before February faculty offers arrive, so you'd have a real option in hand while you decide.

Open to 20 minutes on {date one} or {date two}?

{your name}

--- Advisor ask ---
Subject: Students in your group finishing by {term}

Professor {last name},

I'm hiring a {role} at {company} working on {problem, one line}. Your group's work on {topic} is the closest fit I've found.

Are there students or postdocs in your group finishing by {term} who would want an industry research option alongside faculty applications? I'm happy to send the role description and timeline for you to forward, so nobody feels pressed.

Thank you,
{your name}

How do you sort candidates into industry-first, splitting, or academia-first?

Sort each candidate by the preference they stated word for word, then adjust for timing. Industry-first means the page or the candidate says industry. Splitting means they list both, or "open". Academia-first means they list only faculty or tenure-track positions. Never drop academia-first rows. They are the pool most likely to turn into splitters once faculty offers start arriving.

• Industry-first: "Seeking" or "Employment Preference" names industry or research labs. Proves current intent. It lies when the page is stale and the candidate already has an offer.
• Splitting: both tracks listed, or the candidate is applying to both. These people decide around February and are your main industry window.
• Academia-first: tenure-track or faculty only. Proves where they are aiming, not where they will land. One candidate reported more than 700 applications for a single tenure-track CS position.

Expect the academia-first column to be the largest from market pages. Both example entries on the UChicago page say "Tenure-track positions". Market pages are written partly for faculty search committees, which skews the stated preferences. In a year when industry hiring cools, the academia-first share grows, as it did when academia peaked at 40.1% in 2024.

How do you date an international candidate's earliest start?

For a US international student, the earliest realistic start date is the program end date plus the post-completion OPT window and adjudication time, never the defense date itself. Post-completion OPT, short for Optional Practical Training, is the work authorization most F-1 students use after finishing a degree. Its rules, set out in the USCIS Policy Manual (Volume 2, Part F, Chapter 5), decide the earliest date on that candidate's row.

The rule:

• A student may apply no earlier than 90 days before the program end date and no later than 60 days after it.
• The requested start date cannot be more than 60 days after the program end date.
• All practical training must be completed within 14 months of completing study.

Adjudication sits on top of this. One secondary source puts OPT processing at roughly 90 to 120 days. If a candidate files on the first allowed day and processing takes that long, authorization arrives somewhere between the program end date and about a month after it. Processing times change, so check current USCIS processing times before you put a start date in an offer.

Date postdocs with a different rule. They are not on the post-completion OPT clock, so use their stated availability and record their work-authorization status as unknown until they tell you. This guide covers the US rule only. For candidates finishing in Canada or elsewhere, look up the local post-study work rule before dating their rows.

How this goes wrong: failure modes and false positives

Most bad finishing-PhD slates fail in the same seven ways. Each one looks like a valid row until you check one specific thing.

1. Skill labels skew pool counts. Refolk's index shows 30 US NLP PhD candidates against 285 in Computer Vision. That gap reflects how people tag the skill, not how many NLP students exist. What it looks like: a specialty that seems too thin to staff. Check: rerun the search on headline keywords such as "NLP" and "LLM" before concluding the pool is small.
2. Stale student titles. A profile still says PhD student after the person has founded a company or taken an industry job. What it looks like: a strong row that never replies. Check: compare the current affiliation against the market page and the latest paper's affiliation string, and date the check.
3. Industry employers listed as student affiliations. In Refolk's index, Apple appears among the top employers on PhD-student profiles. These are internships, not hires. What it looks like: a candidate who seems already taken by a big lab. Check: read the date on the internship. A finished summer internship means the person is still available.
4. Treating academia-seeking pages as a reason to exclude. Market pages lean toward faculty seekers. What it looks like: a slate of five people from a department with twenty listings. Check: classify those listings as academia-first, keep them, and send advisor asks.
5. Defense date treated as start date. OPT start can fall no more than 60 days after the program end date, and adjudication can push it later. What it looks like: a hiring manager planning onboarding for the week after a defense. Check: every international row's earliest start must show the OPT arithmetic.
6. The first author is not the student. Postdocs also lead papers, and they appear on the same market pages. What it looks like: a "student" with a decade of papers and no defense date. Check: read the author's affiliation and homepage, and mark postdocs in their own column.
7. Dating under an unverified grace-period rule. The reported 30-day grace period has not been confirmed. What it looks like: every international row shifts at once when someone checks. Check: confirm the current DHS rule, and record which rule you used in the sheet.

Failures 2, 3, and 6 have the same root cause: someone trusted a single layer of the evidence stack. A row is only ready to rank when at least two layers agree on department, status, and timing. Refolk helps here because it returns each person with a current affiliation, but the market page and the paper's affiliation string still settle disputes.

How do you keep the slate current after day ten?

A finishing-PhD slate goes stale on a known schedule: market pages update in the fall, faculty offers land around February, and visa rules change by publication. Put those dates in your calendar on day ten instead of rebuilding from scratch.

• Early fall and November: re-visit every department on the scored list. New pages and new defense dates appear, and Northwestern's page showed next-cycle defense dates by November.
• Late February: re-classify academia-first rows once first faculty offers have gone out, and follow up with anyone whose status changed.
• Each new venue cycle: pull the accepted-paper lists for your three to five venues and match new first authors to departments. A Refolk query for each lab, phrased the same way as on day ten, turns up people who never listed themselves.
• Each Taulbee release: update the destinations table, and pull Table D4 if AI/ML-specific splits matter to the brief.
• Before any offer to an international candidate: re-check the USCIS Policy Manual chapter, current processing times, and the status of the grace-period change.