AI Research Scientist interview prep, what to expect

You interview, and the talk is where the field gap closes or does not. Labs grade the talk on research judgment: motivation, the gap, the design decisions, the ablations you ran and the ones you should have. Those transfer whole from vision to language. What does not transfer is the paper discussion and the open-problem round, where the interviewer will hand you a language-model question and watch you reason. Spend the weeks before the onsite reading five to ten recent LLM papers deeply enough to critique the experimental design, and pick your talk project for the decisions in it rather than the modality. In the why-this-lab answer, name a language problem at the lab that your method speaks to, such as a probing or evaluation technique, and propose the first experiment. A vision researcher who arrives with one concrete LLM experiment in hand is judged as a researcher; one who arrives with a vision talk and no bridge is judged as a transfer candidate.

Apply as a research engineer unless you have first-author papers at top venues. DeepMind's research scientist track requires a doctorate in practice, and every lab's talk and paper rounds assume a publication record to interrogate; without one there is nothing for the loop to grade. The research engineer loop at the same labs is where your profile wins: the coding bar is higher, the implementation and debugging rounds are the centre, and the research rounds ask whether you can run someone else's idea well rather than originate it. Pay at the labs is close between the two titles at the same level. If you want the scientist title later, the route that works is publishing from inside: research engineers at these labs co-author, and a first-author paper from a lab team is the credential the scientist loop wants. Prepare for the engineer loop with the research engineer guide, and keep the talk-style project narrative from this one, because the same interviewers ask both.

Two weeks of deliberate derivation, out loud. The quiz is oral, undergraduate level, and unaided, and DeepMind's own interviewers say candidates out of academia for years fail it on formal definitions they once knew. Work through Mathematics for Machine Learning with a notebook and a timer: the rank-nullity theorem, eigendecomposition and why it matters for PCA, the gradient of logistic regression and of a two-layer network by hand, the bias-variance decomposition, KL divergence between two Gaussians, why L1 gives sparsity with the subgradient argument. For each, practise saying the definition first and the intuition second, because the round rewards the definition and then asks for the picture. Pair it with the debugging round, which also tests what you retained rather than what you can look up: a script that runs but does not learn, and the three or four bugs that cause that. Five years of industry gives you the research talk and the judgment rounds; the quiz is the one place experience counts against you.

Anthropic's is a 45-minute conversational round, and candidates describe it as the one that decides. It is not a test of a safety publication record; it tests whether you have thought about the problem as a researcher. Prepare two considered positions: the strongest arguments against Constitutional AI as a method, with what evidence would settle them, and what you would do if a model behaved differently during evaluation than in deployment. Connect one to your own work, for instance how your evaluation method would detect that gap. The failure the lab names is a brilliant candidate who treats the question as a formality. OpenAI has no dedicated round and folds alignment into the research and behavioural discussions; DeepMind puts mission and values into the final round with team leads. The same two prepared positions serve all three, and at OpenAI and DeepMind the question tends to arrive as "how would this affect what you would build" rather than as a separate hour.

The loop described here is the frontier-lab shape (OpenAI, Anthropic, Google DeepMind), with the differences between them named per round. Meta FAIR, Microsoft Research and the applied labs run the same spine of a research talk, a paper deep dive and coding, with Meta adding two LeetCode-style problems per 45-minute coding session and an ML system design round, and Microsoft Research leaning further toward publication record and the talk. The pay figures are for the three frontier labs; big-tech research groups sit on their company's engineering ladder.

The process takes four to eight weeks, and team matching can add a month. Give the research talk the first two weeks of prep and rehearse it against interruption; give the fundamentals and the from-scratch coding the next two, because those are the rounds that remove strong researchers.

Preparing the talk and nothing else. The talk is where a strong record shows, and the rejections come from the rounds around it: a CodeSignal screen failed on correctness, an oral quiz on undergraduate mathematics answered with intuition, a safety question treated as a formality, and a known limitation defended instead of owned.

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.

This guide covers the general bar at L5 / IC4 equivalent (lab RS). 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.