In this article
- How the four rounds fit together
- Start here: the learning path
- The paper-discussion round: read in three passes
- Research brainstorm and problem-framing rounds
- Research taste questions
- Defending your own research
- What the rest of the loop looks like
- Questions interviewers ask on research scientist interviews
- All 3 guides in Research scientist interviews
- Follow-up questions
- Sources
How the four rounds fit together
A research loop tests one skill from four sides: whether you can reason about evidence you did not collect yourself, and about evidence you did. Formats vary by team and employer, so ask your recruiter which rounds apply. What stays constant is the vocabulary of the people on the other side of the table. They are researchers, and they judge a paper or an idea the way a reviewer does.
That gives you a free rubric. The NeurIPS 2025 reviewer guidelines ask reviewers whether a submission is technically sound, clearly written, significant and original, and whether the authors adequately addressed its limitations (NeurIPS 2025 Reviewer Guidelines, read 2026-10-24). An interviewer who asks "why is this paper important?" is asking about significance; "what would you do differently?" is about soundness and limitations. Prepare in those terms and most probes will map to one of them.
Text description of the diagram
- Research loops often include four rounds: paper discussion, research brainstorm, research taste and defending your own work.
- Paper discussion is the starting point of the path.
Start here: the learning path
Start with the paper-discussion round, because it teaches the reading and critique habits that the other three rounds reuse. The brainstorm round comes next because it asks you to turn a critique into an experiment. Research taste follows, because taste is what you show when nobody gave you the paper. Defending your own research comes last since it points the same critique at your own work. The first step, the paper-discussion round, is on this page.
- Research scientist interview
- A research scientist interview is a hiring loop for a role that produces new knowledge, in which interviewers judge how you read, question and extend research, usually through paper discussions, open-ended experiment design and a defence of your own past work.
- Paper-discussion round
- A paper-discussion round is an interview in which you explain, critique and extend a published paper, either one you choose or one the interviewer sends, and answer follow-up questions on its claims, evidence and limits.
- Research brainstorm round
- A research brainstorm round is an interview in which you are given an open problem and asked to propose hypotheses and experiments on the spot, and to say what result would prove you wrong.
- Research taste
- Research taste is the judgement that tells you which problems are worth working on and which explanations are most plausible, and in an interview it shows as specific, falsifiable opinions rather than general enthusiasm.
- Ablation
- An ablation is an experiment that removes or replaces one component of a method, keeping everything else fixed, to measure how much that component contributes to the result.
- How do research brainstorm and problem-framing rounds work?
- Research taste questions and answers
- Defending your own research: how do you do it in an interview?
The three rounds after the paper discussion
Three rounds follow the paper discussion: research brainstorm and problem-framing, research taste questions, and defending your own research. Each is covered in its own section below.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The paper-discussion round: read in three passes
The paper-discussion round rewards a reading method more than a good memory. S. Keshav's method reads a paper in up to three passes instead of from start to finish. The first pass takes about five to ten minutes: read the title, abstract and introduction, the section headings and the conclusions, and glance at the references. It ends with the "five Cs": category (what type of paper), context (which papers it relates to), correctness (do the assumptions look valid), contributions and clarity (Keshav, 2007).
The second pass takes up to an hour and studies the figures: are the axes labelled, and are results shown with error bars? After it you should be able to summarise the main thrust of the paper, with its supporting evidence, to someone else. The third pass tries to "virtually re-implement" the paper, which takes about four or five hours for beginners and about an hour for an experienced reader, and it exposes implicit assumptions, missing citations and weaknesses in the experimental or analytical techniques. If you choose the paper yourself, do all three passes. If the lab sends a paper, do the third pass on the claims you expect to be probed.
Text description of the diagram
- The first pass takes about five to ten minutes.
- The second pass takes up to an hour.
- The third pass takes about an hour for an experienced reader and four or five hours for a beginner.
To rehearse the round, prepare answers to the probes that follow from the reviewer rubric above: why the paper matters, its strongest claim and whether the evidence supports it, its limitations, how you would improve it, and what you would run next. This is a practice list built from the rubric, not a report of what any employer asks, so treat it as a rehearsal list and not a script. "I read it but I do not remember the details" is the answer to avoid; a one-page note per paper, with the claim, the evidence and your two doubts, prevents it. Keep the note to your own words.
Research brainstorm and problem-framing rounds
A research brainstorm round asks you to design an investigation on the spot, and it is scored on the discipline of the plan more than on the idea. A strong plan states a hypothesis, the experiment that would separate it from the most plausible alternative, the metric, and what result would prove the hypothesis wrong. A weak plan asks for more compute or adjusts parameters without a question behind it.
The NeurIPS paper checklist is a useful template, because it lists what a finished study must address: its claims, limitations, experimental settings, statistical significance and compute (NeurIPS Paper Checklist, read 2026-10-24). Walk the interviewer through those headings out loud. A good practice prompt is an open question such as how you would detect a model behaviour like hallucination or deception; it shows why a failure condition stated in advance matters.
Research taste questions
Research taste questions ask what you would work on and what you believe that others do not, and they are scored on specificity. "What are the open problems in the field?" has no wrong answer, only vague ones. A good answer names one problem, gives your reason for thinking it matters, states what evidence would change your mind and says what first experiment you would run.
Richard Hamming's 1986 talk to a research colloquium included three questions he put to colleagues, in order: "What are the important problems of your field?", then "What important problems are you working on?", then, if what they were doing was not important, why they were at Bell Labs working on it (Hamming, 1986). He defines an important problem as one where you have "a reasonable attack". Expect a version of these questions. Prepare two or three positions you can defend and one you hold loosely.
Defending your own research
A defence of your own research is a questioning of decisions you made, and it is scored on honesty about what you know. Interviewers ask what gap you saw, which ablation was the strongest, what surprised you, which assumption would hurt most if wrong and what you would run with more time. Answer each in the same order: the claim, the evidence, the weak point.
If you have not published, use a project with real measurements: a thesis chapter, an open-source experiment or a reproduction. The reviewer's own question is a good practice prompt: does the main claim in your abstract match what your experiments show? That is the first item on the NeurIPS checklist, which asks whether the claims in the abstract and introduction accurately reflect the paper's contributions and scope.
What the rest of the loop looks like
A research loop is usually one part of a longer process, with coding, machine learning and behavioural rounds around it. The Anthropic interview process page lays out one AI-lab loop stage by stage, and OpenAI interview questions collects the questions reported for that lab. For the coding side, see machine learning coding rounds; for how rounds fit together, see interview loops. Behavioural answers matter in research loops too, so build your stories with the behavioral interview guide. If a paper you discuss is about transformers or agents, the transformers guide and the AI agents guide explain the background in plain language.
Questions interviewers ask on research scientist interviews
These are recurring prompts, in our own words, with what each tests. We make no claim about how often each is asked.
| Question (our wording) | What it tests |
|---|---|
| Why does this paper matter, and to whom? | Significance |
| What is the strongest claim, and is the evidence enough? | Soundness |
| What would show the central hypothesis is wrong? | Falsifiability |
| What is the most plausible alternative explanation? | Critical reasoning |
| What would you run next with more compute? | Experiment design |
| Which problem would you work on here, and why? | Research taste |
| Which ablation mattered most in your own work? | Depth of understanding |
| Where would a replicator get stuck? | Reproducibility |
Follow-up questions
What do research scientist interviews ask?
They ask you to explain and critique papers, design experiments for open problems, say what you would work on and defend your own work. Each prompt maps to a reviewer's question about soundness, significance, originality or limitations.
How do you prepare for a paper-discussion interview?
Read the paper in three passes, write a one-page note with the claim, the evidence and two doubts, and rehearse the probes about importance, limitations and next experiments. If the lab chooses the paper, ask how long you have and what format to expect.
What would falsify the central hypothesis here?
Name the observation that would contradict it, not a vague failure. For example, if a method claims a gain from one component, an ablation that removes the component and keeps the gain would falsify the claim.
How do you present a paper in an interview?
State the claim first, then the evidence, then one limitation, and finish with the experiment you would run next. Pick a paper you can explain without notes, and avoid the most famous paper in your field.
What if I have not published a paper?
Use the best research-like work you have done, such as a reproduction, an open-source experiment or a thesis chapter, and be ready to explain every decision in it. Interviewers score the reasoning, not the venue.
- Research rounds test one skill from four sides: judging evidence you did or did not collect.
- Use reviewer questions as your rubric: soundness, clarity, significance, originality and limitations.
- Read a paper in three passes and keep a one-page note in your own words.
- A brainstorm answer needs a hypothesis, a discriminating experiment and a result that would prove you wrong.
- Taste is shown by specific, falsifiable positions; defend your own work claim first, evidence second, weakness third.
Sources
- How to Read a Paper — S. Keshav, ACM SIGCOMM Computer Communication Review 37(3), pp. 83-84, 1 Jul 2007
- NeurIPS Paper Checklist Guidelines — NeurIPS, 24 Oct 2026
- NeurIPS 2025 Reviewer Guidelines — NeurIPS, 24 Oct 2026
- You and Your Research — R. W. Hamming, Bell Communications Research colloquium (transcript read 2026-10-24), 7 Mar 1986
