Last updated: November 4, 2026
An AI research scientist job description should name the open problem the hire will work on, what counts as success at month twelve, the publication and patent policy, the compute they get, and the pay range. Most postings list a PhD and a stack instead. The researchers you want read that as a company that hasn’t decided what it’s buying.
Three questions. She asked all three on the second call.
The company was a medical imaging software firm in San Diego, around 200 people, and the req was for a research scientist in computer vision. The posting was four lines long in the part that mattered. PhD required. Publish at top venues. Experience with PyTorch and transformers. Competitive salary. Their finalist was a postdoc at UC San Diego with two first-author CVPR papers on low-dose CT reconstruction, which was close to the exact problem the company had, and she was interviewing at two other places.
Who approves a paper before it goes out, and how long does that take? How many GPUs would she have, and would she share them with anyone? And who owns something she builds on a Saturday on her own laptop?
The VP of engineering didn’t know any of the answers. He found out over the next week. Legal said every external paper needed review, with no timeline attached. The GPUs were 16 A100s shared with the product team’s nightly retraining jobs. The offer letter carried the company’s standard invention assignment agreement, and nobody had attached the written notice that California requires with one. She took an offer from a Seattle lab whose posting had said, in plain words, 30-day pre-publication review and a dedicated allocation of 32 H100s. Was the Seattle offer bigger? About $15,000 on base. She told the VP it wasn’t the money.
I’m not a researcher. I came up through sales leadership, and at KORE1 I sit in on intake calls where a hiring manager tells us what the job actually is, which is rarely what the draft posting says. On research reqs the gap is wider than anywhere else we recruit. Our AI research scientist staffing team earns a fee only if you hire a candidate we sent you. Keep that in mind as you read. You can use all of it without us.
Two things are covered elsewhere, so I’ll skip them. Whether you need a frontier researcher, an applied scientist, or an engineer with a fancy title is the first question in our 2026 hiring guide for AI research scientists, along with how to read a publication-heavy CV. If the honest answer is engineer, start from the ML engineer job description template instead. This post is about the sentences a research posting needs that an engineering posting doesn’t.
What a Research Scientist Looks for in a Posting
A research scientist job description is read by people who review papers for a living. They read it the way they read a methods section. Skeptically, looking for the thing that’s missing. An engineer scans a posting for the stack and the level. A researcher scans for whether the company has a problem worth two years of their life, and whether the company will let them work on it.
The federal description of the occupation is much broader than what you’re hiring for. O*NET’s profile for computer and information research scientists puts it in Job Zone Five, extensive preparation needed, and its survey respondents split between bachelor’s, master’s, and doctoral degrees as the requirement. That range covers everyone from a federal systems researcher to a reinforcement learning specialist at a frontier lab. Your posting has to narrow it down. Fast.
Here’s what the strong candidates check for, roughly in the order they check.
- The problem. Named. Not “advance the state of the art in AI.”
- What month twelve looks like, and whether papers count toward it.
- Publication policy, including who signs off and how long sign-off takes.
- Compute. GPU type and count, and whether the allocation is theirs.
- Who they’ll argue with. Other researchers, a reading group, an academic collaborator.
- The invention assignment terms, though most won’t ask until the offer stage, and the ones who ask early have usually been burned before.
- Pay, including what share of it is equity.
Seven items. The typical draft I see covers one of them, pay, and only as “competitive.”
Name the Problem, Then Say What Done Looks Like
Write the open question into the first paragraph of the posting. One sentence. Something like “We’re trying to reconstruct diagnostic-quality CT images from scans taken at a quarter of the usual radiation dose, and the published methods break down on the scanner models our hospital customers actually use.” A researcher who has worked on that problem will stop scrolling. One who hasn’t will know within ten seconds that it’s not their area, and that saves both of you a screen.
If you can’t write that sentence, don’t post yet. Seriously. The hire guide makes the case that a lot of research reqs are really engineering reqs, and the fastest way to find out which one you have is to try writing the problem sentence and see whether it comes out as a question nobody has answered or as a feature request.
The second sentence is harder, and it’s the one most postings skip entirely. How will you judge this person a year from now? Research output comes in different shapes, and the shape you reward decides who applies.
| What you’ll measure at month 12 | Sentence for the posting | Who it pulls in |
|---|---|---|
| Peer-reviewed papers | “We expect one or two first-author submissions a year to venues such as NeurIPS, ICML, ICLR, or CVPR.” | Postdocs and frontier researchers who want to keep a public record |
| A capability in the product | “By month 12, your method runs in production on [workload] and beats our current baseline on [metric].” | Applied scientists who like seeing work ship |
| Internal technical reports that change the roadmap | “You’ll write the internal reports that decide which approaches we fund next quarter.” | Senior researchers comfortable with less public credit |
| Patents | “We file on novel methods before publishing, and inventors are named on every filing.” | Industry researchers from hardware, medical device, and semiconductor backgrounds |
| A mix | “Roughly 70% of your time goes to [product capability], 30% to work you’re free to publish.” | The widest pool, if the split is honest |
That last row is the one I’d push most companies toward, and it’s also the one that goes wrong most often. A 70/30 split that turns into 100/0 by the second quarter is the most common reason I hear from researchers who leave inside eighteen months. Write it only if your VP will defend the 30 when a product deadline slips. If they won’t, write 100/0 and hire an applied scientist who wants exactly that. Plenty do.
Publication and Patent Policy, in Writing
Every research candidate asks about publishing. Most companies answer verbally, during the third round of research scientist interviews, with something warm and vague. Put it in the posting instead, because the vague version costs you candidates you never hear from.
Four policies show up in nearly every research org we recruit for. Any of them can work. “Case by case” doesn’t, because researchers have learned it usually means no.
| Policy | How to write it | How it reads to a researcher |
|---|---|---|
| Open, with a review window | “Papers go through a 30-day review for confidential information, then you submit.” | The best case for anyone who wants a public record |
| File, then publish | “We file a provisional patent first, which adds about six to eight weeks before submission.” | Fine, as long as the delay is stated and deadlines are planned around it |
| Internal only | “This role does not publish externally. Work is shared through internal reports and talks.” | Honest, and it filters out people who would leave over it |
| Case by case | “Publication is evaluated on a case-by-case basis.” | No |
A robotics startup in Pittsburgh we worked with had the file-first policy and was nervous about printing it. Their lawyers wanted provisional filings on anything touching their grasping work before it went to a conference. The founder assumed researchers would hate that. We wrote it in anyway, with the six-to-eight-week number. Two of the three finalists mentioned it on their first call, and both said the same thing, more or less. Good, at least you know. The candidate they hired had left a previous job after a paper sat in legal review for five months with nobody able to tell her when it would come out.
Patents bring up the invention assignment agreement, and in some states that has a legal edge. Under California Labor Code section 2870, an employment agreement can’t make an employee assign an invention they developed entirely on their own time, without the employer’s equipment, supplies, facilities, or trade secrets, unless it relates to the employer’s business or its actual or demonstrably anticipated research, or came out of work they did for the employer. Section 2872 then requires the employer to give written notice of that limit at the time the agreement is made. A few other states have similar statutes. I’m not your lawyer, and none of this is legal advice. But a researcher who builds side projects will read your agreement, and the San Diego candidate did, and the missing notice told her more about the company than the posting had.
What goes in the posting is one sentence. “Our invention assignment agreement follows [state] law, and we’ll share it with finalists before the onsite.” It costs nothing, and it signals that someone thought about the question before she had to ask.
Compute and Data, Down to the Card
Researchers ask about GPUs the way sales candidates ask about territory. It’s the resource the job runs on.
“Access to significant compute” means nothing. Every posting says it. Write the hardware and the allocation instead. How many cards, which generation, whether they’re dedicated to research or shared with production training, and how jobs get scheduled. A line like “a dedicated allocation of 32 H100s on our Slurm cluster, plus burst capacity on AWS p5 instances with a monthly budget” tells a candidate more than a paragraph of adjectives. If the honest answer is “eight A100s shared with the product team, scheduled through a Kubernetes queue that gets busy before releases,” write that too. Some researchers will pass. The ones who don’t will arrive with accurate expectations, and that matters more at month nine than it does on day one.
The Stanford HAI 2025 AI Index found that nearly 90% of notable AI models released in 2024 came from industry, up from 60% in 2023, and that training compute for frontier models was doubling about every five months. Researchers know those numbers better than we do. They know a smaller company can’t match a frontier lab’s cluster. Most aren’t asking you to. They’re asking whether the work they’d be hired to do is possible on what you have, and only you can answer that.
Data gets one more sentence. Name the datasets they’ll touch, roughly how large, and whether labeling has a budget. A medical imaging researcher who learns at the onsite that the de-identified scans are still stuck in a data-use agreement negotiation with two hospital systems will do the arithmetic on how long that takes. So will her next employer’s recruiter.
Who They’ll Argue With
A lone researcher at a company with no other researchers tends to go quiet around month six, and then tends to go. The hire guide gets into why. The posting’s job is smaller than that. It just has to tell the truth about the room.
So say how many people are in it. “You’ll join two research scientists and a research engineer, reporting to our head of applied research, who publishes on the same problem” is one honest version. “You’ll be our first research hire, reporting to the CTO, who has a PhD in signal processing and hasn’t published since 2017” is another, and I’ve watched that second version recruit well, because the people who want to build a function from scratch go looking for exactly that sentence. The posting that gets in trouble says nothing about the team at all, so the candidate pictures a lab and then meets one tired CTO at the onsite.
Small things count here too. A Thursday reading group. Two conferences a year, paid for, with the travel budget written down. An adjunct appointment someone on staff holds at Carnegie Mellon, or a funded collaboration with a lab at UC San Diego. Research is a small world, and candidates will look up every name you print.
Requirements That Don’t Filter Out the Wrong People
“PhD required” is the line almost every research posting starts with. It’s not wrong. It just filters on the credential when what you need is the evidence the credential usually comes with.
Write the evidence instead. “A record of original research in [area], typically shown through first-author papers at peer-reviewed venues, usually earned through a PhD.” That keeps the doctoral candidates in. It also keeps the occasional researcher with a master’s degree and three strong first-author papers, and it drops the PhD from an unrelated field who would otherwise clear the filter. The BLS describes the occupation as typically needing at least a master’s degree, by the way, not a doctorate, though frontier research is a narrow slice of that occupation and the doctorate is far more common there.
Be specific about the area. “Machine learning” covers everything from causal inference to speech recognition. “Generative models for medical imaging, ideally diffusion or flow-based reconstruction” covers far fewer people, and those people will recognize themselves. Tools belong lower in the list. PyTorch or JAX, sure. Everyone in the pool already uses one of them, so they barely discriminate.
The nice-to-have list is where you put the things that separate a researcher who can work inside a company from one who can’t. Experience taking a method from a paper to a system someone else maintains. Having mentored a junior researcher or an intern. Having written for a non-research audience. Those are the traits our recruiters spend the most time probing for, and the ones that most often explain why our placements hold. Across our technical desks, 92 out of every 100 placements are still employed by the client twelve months later.
Pay, Equity, and the Visa Line
Print a range. Colorado, California, Washington, New York, and more states each year require a pay range in job postings, with thresholds that vary by state, and researchers compare offers line by line wherever they live.
Bureau of Labor Statistics data for May 2025 shows a $140,300 median annual wage across computer and information research scientists, with the top tenth earning more than $230,630, and the bureau projects 22% employment growth from 2025 to 2035 with about 2,900 openings a year. Those figures cover a much wider occupation than AI research, so treat them as a floor. Our hire guide puts applied research scientists at roughly $180,000 to $260,000 in base and frontier researchers at $200,000 to $300,000, with frontier-lab total packages running from $600,000 to well past $1.5 million once equity is counted. Before you set your own number, compare it with the lab filings in our AI research scientist salary guide and check it against our salary benchmark tool.
If equity is a big share of the package, say so in the posting and give the form. Options or RSUs, the vesting schedule, and whether there’s a refresh. Researchers coming from big labs know how to value RSUs. Startup options are harder for them to price, and a posting that explains the basics earns trust before the offer call.
Then the visa line. Many of the strongest candidates in AI research came through U.S. doctoral programs on student visas, and plenty of them hold OPT, H-1B, or O-1 status today. Say plainly whether you sponsor, which visa types, and whether you support green card filings. “We sponsor H-1B and O-1 visas and support employment-based green card applications” is a complete sentence. So is “We are unable to sponsor visas for this role.” Leaving it out wastes everyone’s first call. If green card sponsorship is part of your offer, our read on the October PERM suspensions covers what changed for employers and what didn’t.
AI Research Scientist Job Description Template
Replace every bracket. Some brackets are notes to you rather than words for the candidate, so delete those before you publish.
Job Title
[Research Scientist, Computer Vision / Senior Research Scientist, Language Models / Applied Scientist, Speech] [Put the area in the title. If the role doesn’t publish and mostly ships product, “Applied Scientist” is the more honest title.]
About the Role
[Company] builds [product, in one plain sentence] for [customers]. We’re hiring a research scientist to work on [the open problem, in one sentence a specialist would recognize]. Published methods [fall short because / haven’t been tested on / don’t scale to] [your specific constraint]. You’ll report to [title], who [publishes in the area / leads applied research / is our CTO].
What Success Looks Like at Twelve Months
[By month 12, we’d count it a success if [a method beats our current baseline on [metric] and is running on [workload] / you’ve submitted [one or two] first-author papers / your internal reports have decided which of [two approaches] we fund].] [If it’s a split, give the percentages and who defends them.]
Publication and IP
[Papers go through a [30]-day review for confidential information, then you submit.] [Or: We file provisional patents before submission, which adds about [six to eight] weeks.] [Or: This role does not publish externally.] Our invention assignment agreement follows [state] law, and we’ll share it with finalists before the onsite.
Compute and Data
[A dedicated allocation of [number] [GPU model] on [cluster or scheduler], plus [cloud burst capacity / a monthly compute budget of $[amount]].] [Or: [number] [GPU model] shared with [team], scheduled through [queue].] You’ll work with [datasets, rough size, and whether labeling has a budget].
The Team
[You’ll join [number] research scientists and [number] research engineers.] [Or: You’ll be our first research hire.] [Reading group, conference travel (for example, two conferences a year, funded), and any university collaboration.]
What You’ll Do
- Frame [the open problem] into experiments that can succeed or fail clearly, and run them.
- Read the literature in [area] and tell us which published results are worth reproducing on our data.
- Build prototypes that a research engineer or ML engineer can take to production, and stay involved through the handoff.
- Write up results, including the ones that didn’t work, in a form [product leadership / the research team] can act on.
- [Publish and present at [venues], within the policy above.]
- [Mentor [interns / junior researchers].]
What You Bring
- A record of original research in [area], typically shown through first-author papers at peer-reviewed venues and usually earned through a PhD.
- Hands-on depth in [the specific methods, for example diffusion models, reinforcement learning from human feedback, or self-supervised representation learning].
- Fluency in [PyTorch / JAX] and experience running experiments at [scale, for example multi-node training on [number] GPUs].
- The ability to explain a result, and its limits, to people who don’t read papers.
Nice to Have
- [Experience taking a method from paper to a production system someone else maintains.]
- [Work on [domain-specific data, for example medical imaging, speech, or robotics sensor data].]
- [Mentoring or teaching experience.]
Pay, Equity, and Sponsorship
The base salary range for this role is $[low] to $[high], depending on [level and location]. [Equity: [options / RSUs], vesting over [four] years with [a one-year cliff], [refresh policy].] [We sponsor [H-1B / O-1] visas and support employment-based green card applications.] [Or: We are unable to sponsor visas for this role.] [Benefits, in one line.]
How to Apply
[Send a CV and links to [two or three] papers or projects you’re proudest of.] [Tell candidates how many rounds there are and whether one includes a research talk.] [We aim to give every candidate a decision within [three] weeks of the first call.]
What Research Leads Ask Us Before They Post
Can we just write “PhD required” and leave it there?
You can, but you’ll filter on the credential instead of the evidence. Write “a record of original research in [area], usually earned through a PhD” and you keep the right people in.
The difference shows up in the edges of the pool. A PhD in an unrelated field clears the “PhD required” filter. A master’s-degree researcher with three first-author papers in your exact area doesn’t. Neither outcome is what you wanted.
Is promising publication rights in writing risky?
Less risky than staying vague. A written policy with a review window protects your confidential information and tells candidates exactly what they’re signing up for.
How specific should the compute line get?
GPU model, count, whether it’s dedicated, and how jobs are scheduled. Four facts, one sentence.
If the allocation is modest, write it anyway. Researchers would rather know at the posting stage than at the onsite, and the ones who apply knowing you have eight shared A100s are the ones who’ll make that work. Dressing it up as “access to significant compute” just moves the disappointment to week three of employment, which is a much more expensive place for it.
Research scientist, applied scientist, or research engineer on the title line?
Match the title to the twelve-month measure. Papers or novel methods point to research scientist, shipped capability to applied scientist, and turning other people’s methods into systems to research engineer.
Titles carry pay expectations too, and candidates search by them. A posting titled research scientist for a job that’s really applied science attracts people who’ll be disappointed by the work. One titled applied scientist for a job that’s genuinely novel research loses the people who could actually do it. When the honest answer is research engineer, our LLM engineer job description and the ML engineer template are closer starting points than this one.
Do we print the salary range if most of the package is equity?
$190,000 to $240,000 base, plus equity described in a sentence, beats a posting that only says “competitive.” Print the base range, then describe the equity’s form and vesting.
In states with pay transparency laws, the base range is the part the law cares about. The equity sentence is the part that wins candidates from bigger companies, because it shows you understand how they’ll compare offers.
Realistically, how long until someone’s in the seat?
Usually four to eight weeks once the role is scoped properly, which is longer than our 17-day average across IT searches. Research candidates get evaluated on judgment, and that takes real conversations with people who can read the work.
Most of the delay we see comes before sourcing, not after. If the problem sentence, the twelve-month measure, and the publication policy are settled before the posting goes up, the search itself tends to run on schedule. If they’re settled in the middle of interviews, add a month.
Let a Researcher Read It First
Before it posts, hand the draft to someone who has published in the area. A collaborator, an advisor you know, the researcher you already have. Ask them one thing. Would you apply? Then ask what they’d want to know before answering, and check whether the posting already says it. The answer is nearly always one of the seven items near the top of this post.
We’ve recruited technical people since 2005. The recruiters on these searches average over 15 years each, we cover more than thirty metro areas, and the intake call is where most of a research search gets won or lost. If you want a research req read before it goes live, send the draft to one of our recruiters. That works whether you’re building a permanent team through direct hire, bringing in a contract researcher for one hard problem, or not using an agency at all. And when you want candidates, our research scientist recruiting team already knows who’s publishing in your area this year.
