2026 update
A few things have changed in 2026. AI is now allowed in coding rounds at Canva and Meta, detection has improved at companies that still ban it, pay has split at staff+, and the post-onsite wait got longer. Read what changed in 2026 →
What you'll be expected to do
What they're grading- Design, run and analyse experiments on large models with researchers, from the first script to the write-up
- Build and scale training infrastructure: data pipelines, distributed training, checkpointing, throughput and reliability
- Build evaluations, find failure modes, and turn them into training environments or datasets
- Own the dev tooling and the experiment tracking the team runs on
- Implement and optimise model components, in PyTorch or JAX, at the level of tensor shapes and kernels
- Pair with researchers daily and take on work outside the job title when the project needs it
What does the interview loop look like?
8 rounds · 3–8 weeksMost companies follow a similar pattern for Research Engineer interviews. Total calendar time is 3–8 weeks from recruiter screen to offer.
Which team, which problems, why this lab, and a calibration of what you have built. Anthropic and OpenAI both ask the mission question here before any code
Anthropic: a 90-minute CodeSignal test, a class exposing a public API built up over four levels, where candidates report that speed and passing every case matters more than elegance. OpenAI and others: a practical problem such as an LRU cache or a rate limiter, judged on clean working code
What they're scoring
- Working code fast, under a time limit, with the edge cases covered
- Readable structure, since the reviewer reads the code after the run
- For the live version, talking while typing without stalling
Implement a substantial piece of a neural architecture from memory: scaled dot-product attention with masking, a Transformer block, a training loop with gradient clipping. Tensor shapes, contiguity and memory are the grading, and the Transformer is the expected default
What they're scoring
- Correct shapes at every step, said out loud before the code
- Masking and numerical stability handled without prompting
- Knowing what the framework does under the hood, contiguity and broadcasting included
OpenAI sends a paper two or three days ahead; you present its contribution, method, results, limits and what you would try next. At Anthropic the same hour is a deep dive on a project of your own, and the questions go at the decisions you made and the ones you did not
What they're scoring
- Separating what the paper showed from what it claimed
- One concrete, cheap extension you would run first
- Owning the decisions in your own project rather than the team's
Open-ended questions from a research lead: what would you try, why, what would tell you it is working. One candidate account describes 15 minutes, two questions, and elimination after three minutes of silence
What they're scoring
- Ideas within the first minute, refined out loud rather than delivered finished
- Saying what result would change your mind
- Cost awareness: which experiment is cheap enough to run first
Design the training or serving system for a stated model and budget, or debug a run that has gone wrong: a loss that plateaus, a throughput that halves, a gap between two evaluation numbers. DeepMind adds a rapid-fire fundamentals quiz where industry candidates report failing on formal definitions
Anthropic research and research-engineering roles usually include one: a problem set or a small dataset, or an analysis of a model API as a black box, with findings presented afterwards. Done without AI assistance unless the brief says otherwise
Motivation, how you work with researchers, and at Anthropic how you reason about safety. Reference checks run during the cycle rather than after the offer. Candidates who describe being dismissive of safety, working alone by preference, or being in it for the money report ending here
Sample questions you should be ready for
9 of the ones that decide itRepresentative of what companies ask at this level. Every question here can be practised out loud, which is the fastest way to find out whether your answer holds up under follow-ups. Calibrd adds voice practice with coaching on every answer, and a full voice mock interview: a live round with an AI interviewer who has read the role and your CV, then an honest debrief.
- 01“Implement multi-head attention from scratch in PyTorch, with a causal mask. Walk through every tensor shape as you go, and say where the memory goes at sequence length 32k.”
- 02“Write a training loop with gradient accumulation and gradient clipping. Where does mixed precision change what you wrote, and what breaks first at scale?”
- 03“Our evaluation number moved by two points between two runs with the same config. List what you would check, in order, and what each check costs.”
- 04“Design the data and training pipeline for a 7B-parameter run on 256 GPUs. Cover sharding, checkpointing, throughput targets and what you monitor.”
- 05“Design an evaluation harness the whole lab can add benchmarks to. How do you keep results comparable across model versions and prompt formats?”
- 06“Serve a 70B model for internal researchers with a latency budget and a fixed GPU pool. Batching, KV cache, quantisation, and what you give up.”
- 07“Tell me about an experiment you designed. What did the result change, and what did you run next?”
- 08“Describe a time a researcher wanted something the infrastructure could not do yet. What did you build, and what did you refuse to build?”
- 09“Why this lab, and which of its recent papers would you have wanted to work on?”
These are the general ones. Paste a real posting and Calibrd predicts the questions that company asks for that exact role, then interviews you on them.
Predict my questions →Compensation benchmark
US majors · USD · medianTypical pay for Research Engineer at major US tech companies, headline numbers in USD. Typical pay in London, Berlin and Singapore is meaningfully lower, and equity varies a lot by company stage.
Levels.fyi for Research Engineer at OpenAI, read on 11 September 2026: median total $880K, range to $1.08M, and the one level with data, L4, at $980K made of $380K base and $600K a year in stock, four-year vest at 25% a year. Anthropic's own posting for a pre-training Research Engineer or Scientist in Seattle lists a salary of $350,000 to $850,000. Google DeepMind research scientist medians are reported at $336K to $406K plus Google stock. The samples are small, the equity is in private companies for two of the three, and the numbers move by the month.
How to prep
6 tactical tipsLead behavioural answers with the STAR method: Situation, Task, Action, Result. The tips below build on that structure for this specific role.
- 01Implement attention and a Transformer block from memory until you can do it in 25 minutes with the shapes narrated. It is the default coding question, and tensor shapes are the grading
- 02Rehearse the paper round on a paper you did not write: contribution, method, evidence, limits, the cheapest extension. Two or three days of notice is what OpenAI gives
- 03Practise thinking out loud on open research questions with a friend for 15 minutes at a time. The brainstorm round eliminates people who need silence before they speak
- 04Read Goodfellow, Bengio and Courville's 'Deep Learning' for the fundamentals quiz, and Hamming's 'The Art of Doing Science and Engineering' for what research taste means in a room
- 05Prepare one project of your own to the level of every decision: why that architecture, why that data, what you measured, what you would redo. The deep dive goes at the decisions
- 06Say why this lab in terms of its work. A layoff explains why you are looking; it is not a reason to want this loop, and the mission question is asked twice
Where do Research Engineer candidates fail?
Spot it in a mock firstA few common mistakes that get Research Engineer candidates rejected even when they are otherwise strong. Worth catching in a mock interview before they show up in a real one.
Solving the coding problem correctly with the wrong tensor shapes narrated, or none, and fixing the mask only after the interviewer points at it.
Why it fails
The ML coding round grades on whether you can be trusted with a training run, and shape errors are how training runs die at 3 a.m. Correct output with silent shape reasoning tells the interviewer you got there by trial, and masking or contiguity fixed on a hint counts against you rather than for you.
The losing answer
“Here is the attention function, it passes the test case.”
The winning answer
“Q is batch by heads by sequence by head-dim, K transposed gives sequence by sequence per head, I mask the upper triangle with negative infinity before the softmax so padding cannot leak, and at 32k tokens this score matrix is the memory problem, which is why I would reach for a fused kernel.”
Fix
Say every shape before you write the line that produces it, from the batch dimension to the head dimension, and write the mask and the numerical guard before the interviewer asks. Treat the code as something that will run on 256 GPUs tonight.
Going silent in the research brainstorm, or offering one polished idea after a long pause instead of several rough ones fast.
Why it fails
The round grades taste and the speed of the first idea, because that is what a researcher gets from you in a hallway. A candidate's own account of an Anthropic round describes three minutes of silence, the interviewer's expression, and an elimination at that stage despite passing the coding rounds.
The losing answer
“Let me think about that for a moment. [long pause] I suppose you could try a bigger model.”
The winning answer
“The cheap first check is whether the effect survives a change of prompt format, one afternoon on the existing eval. If it does, the interesting direction is whether it is a data effect or an architecture effect, and I would split the training data before I touched the architecture.”
Fix
Start talking within the first minute with the obvious idea and say why it is obvious, then improve it out loud. Name the experiment you would run first because it is cheap, and the result that would make you stop.
Presenting the paper round as a summary and the project deep dive as a team story.
Why it fails
Both rounds exist to find out whether you can judge research and whether you made the decisions in your own work. A summary shows reading; the round wants the limits, the missing baseline and the extension you would run. A deep dive told as "we" leaves the interviewer unable to tell what you did, and at the labs a hire has to clear the bar on that one point.
The losing answer
“The paper proposes a new attention variant and shows improvements on three benchmarks.”
The winning answer
“The gain is real on the two long-context benchmarks and within noise on the third, and the baseline was not tuned at the same compute. The extension I would run first is the same comparison at matched tokens, one day of compute, because if the gain survives that the method is worth adopting.”
Fix
For the paper, spend a third of the time on what it did not show and one concrete extension. For your project, use "I" for every decision you owned, name the one you got wrong, and say what you measured after changing it.
Answering "why this lab" with the market, the brand or the pay, and treating the safety or values round as a formality.
Why it fails
Interviewers at the labs name the same rejection reasons: candidates who are dismissive of safety, who prefer to work alone, or who want the offer for the money. The mission question is asked on the first call and again in the last hour, and a mismatch between the two answers is noticed.
The losing answer
“You are the leading lab and the compensation is competitive, and I want to work on frontier models.”
The winning answer
“Your interpretability work on feature attribution changed how I read my own evals, and I have a view on where it breaks at scale. I want to build the tooling that makes that kind of work routine rather than heroic.”
Fix
Name the lab's recent work you would have wanted to be part of and say what you would do differently. On safety, bring an opinion and a trade-off you have made, and let the interviewer disagree with it.
Recommended resources
No affiliate linksBooks, courses, and tools that come up most often in Research Engineer prep.
- 01Deep Learning (Goodfellow, Bengio, Courville) →
The fundamentals the DeepMind quiz round and the OpenAI ML rounds assume. Free online. Chapters 6 to 10 before any lab loop.
- 02The Art of Doing Science and Engineering (Richard Hamming) →
The book behind the research-taste round: how to choose problems worth working on and how to tell a promising direction from a dead end.
- 03Sundeep Teki, AI Research Engineer interview guide: OpenAI, Anthropic, DeepMind →
The most specific public write-up of the three lab loops: the 4 to 6 hour final loop, the paper sent 2 to 3 days ahead, the CodeSignal gate and the quiz round.
- 04Designing Machine Learning Systems (Chip Huyen) →
For the training and serving systems round. The data and pipeline chapters transfer directly to a lab's infrastructure questions.
- 05Anthropic, guidance on candidates' AI usage →
Linked from every Anthropic posting: use Claude to prepare and to polish an application, none in live interviews, and no AI on a take-home unless the brief says so.
Common scenarios
Situations that come up a lotI'm a senior backend engineer with 6 years of experience and no ML research background. Can I get a Research Engineer role at a frontier lab?
Yes, and the posting says so: Anthropic's pre-training Research Engineer listing asks for significant software engineering experience and someone eager to learn ML research, in that order. What you cannot skip is the ML coding round. Spend eight to twelve weeks implementing attention, a Transformer block, a training loop with clipping and accumulation, and a data loader that streams, until the shapes come out of your mouth before the code. Read Goodfellow for the fundamentals a quiz will check. In the systems round you are the strong candidate: distributed training, checkpointing, throughput and reliability are your home ground, so lead with the largest system you kept alive. For the paper round, pick three recent papers from the lab and practise the contribution, limits, extension structure out loud. Frame your story as the engineer who makes research runs boring and repeatable, and say why this lab in terms of its work.
I have a PhD in ML and I'm applying for Research Engineer rather than Research Scientist. What changes in the loop, and is it a step down?
It is a different job, and at the labs the pay bands overlap. What changes is where the bar sits. The coding rounds are heavier and graded on production quality: clean working code under a time limit on the screen, then from-memory architecture code where tensor shapes and memory are the grading. The paper round is easier for you than for anyone else, so do not spend your prep there. Spend it on the systems round, which asks how a 7B run survives on 256 GPUs, and on writing code outside a notebook: classes, tests, a training script that restarts from a checkpoint. In the project deep dive, describe the engineering you did on your own thesis code, the parts that broke and what you rebuilt, because that is the evidence the title needs. Say plainly why you want to build the systems rather than lead the research, and mean it; the values round listens for it.
How do I prepare for the Anthropic Research Engineer loop specifically, with the CodeSignal test and the take-home?
Treat the CodeSignal gate as its own exam: 90 minutes, a class exposing a public API built up in four levels, and candidates report that finishing every level with passing tests matters more than optimal complexity. Practise timed problems that grow in stages. The loop after it runs about 20 days: a live coding round that candidates describe as a medium problem with a twist, a project deep dive that goes at your decisions, a research brainstorm with a lead where the first idea has to come fast, a take-home of around five hours in a 48-hour window done without AI assistance, and a values conversation. Reference checks run during the cycle, so warn your references early. Read Anthropic's candidate AI guidance before you start: Claude to prepare and polish, none in live rounds, none on the take-home unless the brief says otherwise. Bring an opinion on safety with a trade-off you have made; a dismissive answer there ends the process.
I solved every coding question at two lab onsites and was rejected by both without feedback. What went wrong?
Almost certainly not the code. Accounts from candidates in the same position in 2026 point at the same two rounds: the project deep dive and the research or values conversation. Passing the technical rounds is the entry price at the labs; the hire decision is made on whether you judged research well and whether you owned the decisions in your own work. Three checks. Did your project story use "I" for the decisions and name one you got wrong? Did you start talking within a minute in the brainstorm and improve the idea out loud, or wait for a finished answer? Did your "why this lab" name their work, and did your safety answer carry a trade-off rather than agreement? Fix those and re-apply after the cooling period, which is usually several months. Next time, ask the recruiter after the loop which round was the concern; some will say.
Frequently asked questions
I'm currently a ML Engineer (L4 / IC3). Should I read this guide or the ML Engineer guide first?
Read the ML Engineer guide first. Companies calibrate IC4 / IC5 · lab L4 candidates against the L4 / IC3 bar with a clear scope-gap lens, they want to see where you stand today, then probe the gap up to IC4 / IC5 · lab L4. Read this guide AFTER you understand the L4 / IC3 baseline, so you know exactly which signals you need to demonstrate for the step-up.
How long should I prep before my Research Engineer onsite?
The loop runs three to eight weeks depending on the lab. Add eight weeks of prep: from-memory attention and Transformer code with shapes narrated, one paper a week presented out loud with limits and an extension, and 15-minute brainstorm reps with a friend. The coding is the entry price; the paper and the brainstorm decide.
What's the most common mistake candidates make at the Research Engineer bar?
Preparing for a product-company MLE loop. LeetCode grinding buys nothing here. The rounds that decide are the paper discussion, the brainstorm and the values conversation, where candidates who solved every coding question still report rejections.
What if my interview process is different from what's listed?
Most variation is at the edges. Major tech companies (FAANG, scale-ups, mid-size SaaS) follow processes within 1–2 rounds of what's described. Smaller startups often run fewer rounds (3–4) but the bar at each round is similar; less-tech-mature companies sometimes skip system design or behavioural rounds entirely. Read the posting and ask the recruiter on the screening call; they'll tell you what's coming.
How does this guide compare to running a free scan?
This guide covers the general bar at IC4 / IC5 · lab L4. The free scan reads your specific job description and returns predicted questions for that exact role + company, a pay benchmark matched to the role and, with your CV, a read on your experience gaps and an ATS resume check, emailed as a PDF.
Walk in ready
Walk into your Research Engineer interview ready.
Paste your actual job and Calibrd shows you exactly what that company asks, where your CV is thin, and what it should pay. Then rehearse the round out loud with honest feedback until you're confident. Any tech role. Free to start.
Free to start · No card · Your resume stays yours, encrypted, remove it anytime