The short answer: Research engineers build the training, evaluation and research systems that frontier models depend on, and at these labs they often co-author the papers. The interviews mix general coding, ML coding and debugging, ML fundamentals, research discussion and behavioral rounds. OpenAI and Google DeepMind publish their interview stages for all roles, and Anthropic publishes general facts such as Python-only interviews for several research engineer roles. What happens inside each technical round comes mostly from candidate reports, so ask your recruiter for the exact mix. The three labs also set different rules on AI tools during interviews. Last reviewed October 2026.
Research Engineer interview: quick answers
| Question | Short answer |
|---|---|
| What does a research engineer do? | Builds and scales the systems that train, evaluate and study large models, runs experiments, and debugs large ML codebases. At all three labs the work overlaps heavily with research. |
| What do the interviews test? | General coding, ML coding such as attention or a training loop, debugging ML code, ML fundamentals and maths, large-model system design, a research or paper discussion and behavioral questions. |
| Can I use AI tools? | It depends on the company. OpenAI sets it per interview, Anthropic allows AI only when it says so, and Google DeepMind allows AI for preparation only unless told otherwise. |
| How much do research engineers make? | Posted US base ranges run from $147K at Google DeepMind to $850K at OpenAI and Anthropic. Google DeepMind adds a bonus target and equity to its base. |
| How is it different from a research scientist? | Research scientists usually own a research agenda and, at Google DeepMind, normally hold a PhD. Research engineers build and scale the experiments. Many OpenAI and Anthropic postings combine the two titles. |
| Do I need a PhD? | No. Anthropic says about half of its technical staff had no prior ML experience, and Google DeepMind's Polaris posting requires no degree. Residency and fellowship programs also exist. |
What research engineers do
The job is engineering in service of research. OpenAI's general Research Engineer posting centers on designing and improving massive-scale distributed ML systems and writing bug-free ML code. Google DeepMind describes research engineers as experimentalists who bridge theory and implementation, building and scaling systems to test new ideas.
The line with research is thin. Anthropic says its engineers do lots of research and every Anthropic paper has engineers as authors, often as first author. Most OpenAI research openings are combined Research Engineer / Research Scientist postings for one team. Google DeepMind keeps separate tracks, with Research Scientists normally holding a PhD, and uses a combined title for some senior roles.
Teams shape the day-to-day work. Examples from the postings include post-training production Claude models with RLHF and Constitutional AI at Anthropic, building evals and RL environments for Codex at OpenAI, and hand-building long-horizon eval tasks for Gemini on Google DeepMind's Polaris team. Some Anthropic roles, such as Pretraining Scaling, include on-call during model launches.
How 3 companies run the research engineer interview
The short version of each process. Each company guide has the full detail, with links to the company's own pages, and says which parts come from candidate reports.
| Company | Screens | Main interviews | AI tools | What stands out |
|---|---|---|---|---|
| OpenAI | Intro calls; two 60-minute coding screens (candidate reports) | Finals: 4 to 6 hours, 4 to 6 interviewers | Varies by interview | One general posting considers you for every team |
| Anthropic | Recruiter, Python assessment, manager call (candidate reports) | Virtual onsite of 4 to 5 hours (candidate reports) | Not unless Anthropic says so | All interviews in Python; take-home rebuilt to resist AI |
| Google DeepMind | 30-minute recruiter call; 2 or 3 peer skills interviews | Finals with team leads and leadership | For prep only, unless told otherwise | Maths, ML and CS quiz in older reports |
The interview types in depth
ML coding and debugging
This is the round most specific to research engineers. You implement a piece of a model or training setup, such as attention, a loss and its gradient, or a training loop, or you find the bug in someone else's code. Candidate reports describe ML coding in NumPy and PyTorch and transformer debugging at OpenAI, and ML coding such as attention or training loops at Anthropic.
Interviewers look for code that is correct on the first run, shapes you can state at every step, and a clear debugging method. A structure that works: restate the inputs and output shapes, write the simplest correct version, test it on a tiny example you can check by hand, then make it fast. When debugging, start from the symptom, such as a loss stuck at ln(vocab size), and narrow it down with one check at a time.
General coding
Most loops also include a standard coding round. Candidate reports describe a multi-part practical problem using only the standard library at OpenAI, and medium to hard algorithm problems where the code must run at Google DeepMind. Several Anthropic research engineer postings say every interview is in Python.
Write complete, runnable code with tests, and keep a steady pace through problems that build in stages. Get a working version of each part before extending it.
ML fundamentals and maths
Expect factual and short-derivation questions on probability, linear algebra, optimization and how large models train. Candidate reports from 2021 and 2022 describe a factual maths, statistics, ML and computer science quiz at Google DeepMind, and reports disagree on whether it is still used.
Prepare quick, correct answers to standard questions (why attention scales by sqrt(d_k), why bfloat16, how Adam works) and back-of-envelope numbers: training compute of about 6 FLOPs per parameter per token, and memory per parameter for weights, gradients and optimizer state.
Large-model system design
Some loops ask you to design a system for training or evaluating large models. Candidate reports describe system design for large-model training at Anthropic, and a hard distributed systems problem and a coding task with light system design at OpenAI.
Cover how the model and data are split across devices, communication costs, checkpointing and recovery from failures, and how you would know the run is healthy. Use real numbers for memory and bandwidth.
Research discussion
You discuss a paper, your own project or a research question. Candidate reports describe research or paper discussions at OpenAI and Anthropic, and conversations with tech leads about ML experimentation at Google DeepMind.
Be ready to state the claim, the evidence and its weak points, then propose the next experiment, with the controls and the result that would change your mind.
Behavioral and values
Every loop has behavioral questions. Google DeepMind says many interviews are competency-based and recommends the STAR method. Candidate reports describe a culture and values round at Anthropic, and OpenAI says it evaluates communication and collaboration.
Prepare stories about a failed project, a result you could not reproduce, a change of direction and a disagreement you resolved, with numbers where you have them.
Pay in US job postings
US salary ranges from each company's own postings for this role, seen October 29, 2026. They are base salary unless the company guide says the posting lists total salary. Each company guide has the postings and links.
| Company | Titles in the postings | Posted US pay range |
|---|---|---|
| OpenAI | Research Engineer (one general posting plus team postings such as Codex and AI for Chip Design), and combined Research Engineer / Research Scientist postings for each team | $250K to $850K |
| Anthropic | Research Engineer, <team> (for example Pretraining Scaling, Interpretability, RL Engineering), combined Research Engineer / Research Scientist postings, and Staff or Staff+ Research Engineer | $300K to $850K |
| Google DeepMind | Research Engineer, <team>, DeepMind (for example Gemini Large Scale Pretraining, Robotics, Polaris), with Senior, Staff and Senior Staff titles and some combined Research Scientist/Engineer roles | $147K to $364K |
Entry-level paths
- OpenAI: OpenAI Residency is a six-month program offering a path to a full-time role for researchers and engineers who do not currently focus on AI.
- Anthropic: Anthropic does not offer internships; its Fellows Program is 4 months of full-time research with mentorship, and its Research Residency is for professors, technical founders or staff-level engineers moving into AI.
- Google DeepMind: The Student Researcher Program offers paid 12 to 24 week internships for enrolled BS, MS or PhD students, applied for through Google careers.
AI tools in research engineer interviews
Rules on AI differ by company, so ask your recruiter before every round.
Set per interview. OpenAI says expectations vary by interview: some formats intentionally allow AI tools, and others are designed to assess your independent problem-solving without them. Its interview prep materials say what is allowed, and it asks you to check with your recruiter if unsure.
Off unless the company says so. Anthropic says live interviews are "all you" with no AI assistance unless it indicates otherwise, and take-homes are completed without Claude unless it indicates otherwise. Its engineering blog says its performance engineering take-home explicitly allows AI. For applications, it asks you to write the first draft yourself and then use Claude to refine it.
Prep only. Google DeepMind says you are free to use AI to help you prepare for interviews, but unless told otherwise you should not use any AI tools during live interviews or interview tasks.
A 6-week prep plan
| When | Focus |
|---|---|
| Week 1 | Read the target lab's postings and guide. Write down the team, the title, the interview language and the AI rule. Pick two projects you can discuss in depth. |
| Week 2 | ML coding from scratch in NumPy and PyTorch: attention, a training loop, a loss and its gradient, and KV caching. Test each against a tiny example. |
| Week 3 | Debugging: take working model code, break it in small ways, and practice finding the bug from the symptom. Add general coding problems with tests. |
| Week 4 | ML fundamentals and large-model systems: probability, optimization, parallelism, memory and compute estimates. Practice quick factual answers and short derivations. |
| Week 5 | Research discussion: read three recent papers from your target team and practice stating the claim, the evidence and the next experiment. Prepare behavioral stories. |
| Week 6 | Two full mock loops with back-to-back technical rounds, following the lab's AI rule for each round. |
Courses for each round
- Anthropic Research Engineer Interview
- LLM Architecture
- PyTorch
- Deep Learning
- GPU Programming
- ML System Design
Practice questions
Questions from our bank, then practice prompts from the company guides.
Questions from our bank
From the DataInterview question bank, picked for research engineer interviews. Open one to answer it.
Self Attention in Transformers
LLM Scaling Laws
Data Parallelism Limits
Multi Head Attention Benefits
Attention Repetition Issue
Offline vs Online Evaluation
Generative vs Discriminative Models
Spanish Scrabble Solver
Tensor Parallelism Comparison
ML coding practice
One prompt per company, from the company guides. These are our prompts, not reported interview questions.
Add KV caching to a small decoder-only transformer so that generating each new token does not recompute attention over the whole prefix.
Sample Answer
Store the keys and values of every layer for all past positions, shaped (batch, heads, seq, head_dim). On each step, compute the query, key and value for the new token only, append the key and value to the cache, and attend from the one query to all cached keys. The causal mask is no longer needed for a single new query, but position handling must use the cached length as the offset. A strong answer checks the output matches the uncached model token for token and notes that cache memory grows linearly with sequence length.
In PyTorch, capture the residual stream activations at one layer of a transformer during a forward pass, then steer the model by adding a fixed vector at that layer.
Write a numerically stable softmax cross-entropy loss and its gradient with respect to the logits, in NumPy.
Research discussion practice
One prompt per company, from the company guides. These are our prompts, not reported interview questions.
A reasoning model scores higher when it writes longer chains of thought. Design a simple, controlled experiment to test whether length itself causes the gain.
Sample Answer
Hold the model, prompts and sampling settings fixed and vary only the allowed reasoning length with a hard token budget. Compare accuracy at each budget on a fixed eval set, with enough problems to put a confidence interval on each point. Add a control where extra length is filler that carries no reasoning, to separate length from content. A strong answer states what result would change your mind before running anything.
You add a new principle to a Constitutional AI training setup. How would you check that it changes the intended behavior without making the model less helpful?
Polaris hand-builds a small number of realistic eval tasks instead of generating many synthetic ones. What are the trade-offs?
Fundamentals practice
One prompt per company, from the company guides. These are our prompts, not reported interview questions.
Estimate the training compute for a 7B-parameter model trained on 1 trillion tokens, and how long it takes on 1,000 GPUs that each sustain 400 TFLOP/s.
Sample Answer
Training costs about 6 FLOPs per parameter per token: 6 x 7e9 x 1e12 = 4.2e22 FLOPs. The cluster delivers 1,000 x 4e14 = 4e17 FLOP/s. Time is 4.2e22 / 4e17 = 1.05e5 seconds, about 29 hours. A strong answer adds that restarts, evals and lower real utilization usually make it longer.
You have a compute budget of 1e23 FLOPs. Using training compute of 6ND and about 20 tokens per parameter, roughly what model size and token count are compute-optimal?
A test for a rare condition has 99% sensitivity and 95% specificity, and 1% of people have the condition. If someone tests positive, what is the chance they have it?
Behavioral practice
Prompts we wrote for research engineer behavioral rounds, with what a strong answer covers.
Tell me about a project that failed.
Sample Answer
Name what you expected, what happened and when you knew it was not working. Explain what you salvaged, what you would do differently and how it changed the way you plan experiments.
Tell me about a result you could not reproduce.
Tell me about a time you changed direction on a project.
Tell me about a time you disagreed with a researcher about how to run an experiment.
Are you ready for research engineer interviews?
1 / 5Can you implement attention, a training loop or a tokenizer from scratch in PyTorch or NumPy, and debug it?
Company guides
Read the guide for each company you are interviewing with:
- OpenAI: The OpenAI research engineer interview: the steps OpenAI publishes, what candidates report about the coding and ML rounds, the AI rule and pay in its postings.
- Anthropic: The Anthropic research engineer interview: what Anthropic says about its interviews and AI use, what candidates report about each round, and pay in its postings.
- Google DeepMind: The Google DeepMind research engineer interview: the four stages DeepMind publishes, what candidates report about the coding and ML rounds, the AI rule and US pay.
Frequently Asked Questions
What is a research engineer?
An engineer who builds and scales the systems used to train, evaluate and study ML models, and who runs experiments alongside research scientists. At OpenAI, Anthropic and Google DeepMind, research engineers often co-author papers.
How is a research engineer interview different from a software engineer interview?
It adds ML coding and debugging, ML fundamentals and maths, and a research discussion to the usual coding and behavioral rounds. Candidate reports describe transformer debugging at OpenAI, ML coding in Python at Anthropic and maths and ML rounds at Google DeepMind.
Can I use AI tools in research engineer interviews?
It depends on the company. OpenAI says some formats allow AI tools and others do not, and its prep materials say which. Anthropic says no AI unless it indicates otherwise. Google DeepMind allows AI for preparation but not in live interviews or interview tasks unless told otherwise.
How much do research engineers make?
Posted US base ranges seen in October 2026 run from $250K to $850K at OpenAI, $300K to $850K at Anthropic and $147K to $364K at Google DeepMind, where postings add a bonus target and equity. Each company guide lists the postings.
Do I need a PhD to become a research engineer?
No. Anthropic says about half of its technical staff have PhDs and about half had no prior ML experience. Google DeepMind's Polaris posting requires no degree or prior work experience. Google DeepMind does say its Research Scientists normally hold a PhD.
What programming language should I use?
Python, in most cases. Several Anthropic postings say all interviews are in Python, and candidate reports describe ML coding in NumPy and PyTorch at OpenAI. Postings also name JAX at Anthropic and Google DeepMind.
How long does the research engineer interview process take?
Google DeepMind says 4 to 10 weeks. OpenAI says each stage usually turns around in about a week, and one candidate report describes about 7 weeks. Candidate guides for Anthropic range from about 20 days to 6 to 10 weeks.
