AI research interviews are unlike most technical hiring processes. They test research thinking, not just technical knowledge — and the ability to critically evaluate your own work is valued as highly as the ability to produce it. Here's how to prepare for each stage.
Stage 1: The Research Talk
Almost all senior research interviews include a research talk — typically 30–45 minutes presenting your most significant research contribution, followed by questions. This is often the highest-stakes element of the process.
What interviewers are assessing: Not just whether the work is good, but whether you understand why it's good, what its limitations are, and what you'd do differently with hindsight. Researchers who can only present their work positively signal lower research maturity than those who can critically engage with it.
Structure your talk around: The problem and why it matters (2–3 minutes), what existing approaches miss (2 minutes), your approach and key insight (10 minutes), experimental evidence (10 minutes), limitations and open questions (5 minutes), and implications for future work (5 minutes). Spend more time on the key insight and less on implementation details.
Prepare for these questions:
- "Why did you choose this baseline? What would happen if you used X instead?"
- "What do you think is the most significant limitation of this approach?"
- "How would this scale to [larger problem / different domain]?"
- "What would you do differently if you were starting this project again?"
- "What's the most important open problem this work points toward?"
Honest, thoughtful answers to limitation questions are more impressive than defensive ones. Researchers know no work is perfect; they're testing whether you know it too.
Stage 2: The Paper Deep-Dive
Many research interviews include a discussion of a paper — sometimes one you're given in advance, sometimes chosen by the interviewer during the session. The format varies: some labs send you a paper a week before the interview; others present one you haven't seen and give you 30 minutes to read it.
Prepare by developing a standard reading framework:
- What problem are they solving and why is it important?
- What's the key idea that makes their approach work?
- What are the experiments and what do they prove?
- What do the experiments not prove that the claims imply?
- What are the limitations the authors acknowledge? What limitations did they not acknowledge?
- What would you do next based on this work?
Practice this framework on 2–3 papers per week in the months before your interview. The ability to assess an unfamiliar paper quickly and critically is a genuine skill that can be trained.
The most important thing to demonstrate
Original thinking. Interviewers want to see that you can identify problems, generate hypotheses, and propose experiments that haven't been done. Anyone can summarise a paper. Not everyone can identify a non-obvious limitation and propose an experiment to address it.
Stage 3: Coding Problems
Research coding problems at UK AI labs typically test ML fundamentals through implementation rather than algorithmic puzzle-solving. Common formats:
- Implement from scratch: "Implement an attention mechanism in NumPy" / "Implement a variational autoencoder in PyTorch without using the torch.nn.Module shortcut for the KL divergence"
- Debug a training loop: "This training code isn't converging — why not?" Tests deep understanding of optimisation, gradient flow, and common failure modes
- Extend a paper: "Here's a PyTorch implementation of [paper]. Implement the ablation the authors mentioned they hadn't done"
Preparation: implement foundational algorithms from scratch — attention, transformers, variational autoencoders, policy gradient — without using high-level library abstractions. This builds the depth of understanding that research coding questions are testing.
Stage 4: The ML Theory Deep-Dive
Senior research interviews often include a conversational deep-dive on ML theory in your area. Expect:
- PAC learning bounds and statistical learning theory
- Information-theoretic concepts: mutual information, KL divergence, ELBO
- Optimisation: convergence theory, loss landscape analysis, second-order methods
- Probabilistic ML: Bayesian inference, variational inference, MCMC
- Domain-specific theory in your research area
The depth expected scales with seniority. For Research Scientist roles, expect genuinely deep theoretical discussion. For Research Engineer roles, a solid working understanding of the theory behind the models you'd work with.
Before the Interview: Specific Preparation
Read the lab's recent papers. Go to arXiv, filter by the research area and authors at the lab you're interviewing with. Read the last 12 months of output. Know what they're working on and what you'd want to contribute to.
Prepare your "future work" narrative. "What would you work on here if you joined?" should have a thoughtful, specific answer that connects your prior work to their open problems. This is often the most memorable part of a research interview.
Know your papers cold. Every paper you've co-authored should be presentable in depth. For papers you've cited in your talk, understand them well enough to defend why you cited them.
See the full AI Researcher career guide
Salary data, UK labs hiring, and the full career path from PhD to principal scientist.
Frequently Asked Questions
What makes a research talk strong?
Clear problem motivation, rigorous experimental design, honest acknowledgement of limitations, and compelling implications for future work. Critical self-evaluation impresses more than purely positive framing.
What coding problems come up?
Implementing ML algorithms from scratch, debugging training loops, and extending implementations from papers. The focus is ML understanding, not algorithmic puzzles.
How important are conference publications?
For Research Scientist roles: very important. Papers at top venues are the primary evidence of research ability. For Research Engineer roles, engineering ability matters more. Workshop papers and preprints are meaningful while still in a PhD.
Should I reach out to researchers before interviewing?
Yes, thoughtfully. Discussing their work at conferences is best. Cold outreach about their research (not asking for jobs) can build useful connections. Having a named supervisor at the lab who knows your work is a significant advantage for internship applications.
What questions should I ask?
About research planning processes, publication freedom, compute access, and the biggest open problems the team hasn't addressed. Questions that signal genuine research engagement rather than generic interest.