Anthropic Research Scientist Interview Questions (2026)

1. 1

   Why do you want to work on this Anthropic research team, rather than on AI research in general?

   What they're checking: Whether you've engaged with the team's agenda. The Interpretability application asks this in writing and wants "deeper and more specific engagement" than interest in AI.

   A strong answer: Name one paper from the team, the open question in it you'd push on, why its safety case matters to you, and what in your background equips you to do it.

   Common miss: Enthusiasm for AI safety in general, with no paper, result or open question named.

   Likely follow-up: “What would you work on in your first three months?”

   Show how to answer

“Tell me about yourself”, “Why here?” and the other common questions, and how to prep each →

1. 1

   Tell me about a research project where you prioritized safety considerations over performance metrics

   What they're checking: Whether you'll give up a headline number for a result you trust, and can say plainly what it cost.

   A strong answer: STAR. The metric you could have pushed, the risk you saw, what you did instead, what it cost the project, and whether you'd make the same call now.

   Common miss: A trade-off that cost nothing, or a hypothetical instead of a real project.

   Likely follow-up: “Who disagreed with that call, and what did you tell them?”

   Show how to answer
2. 2

   Tell me about a result you believed that turned out to be wrong. How did you find out, and what did you change?

   What they're checking: Scientific honesty: how fast you update, and whether you fix the process as well as the result.

   A strong answer: STAR. A real result, not a trivial bug; the check that exposed it; how quickly you told collaborators; and the change to your workflow that would catch it earlier next time.

   Common miss: An error someone else made, or a story that ends at the fix with no change in how you work.

   Likely follow-up: “Had anyone already built on the result? What did you do about that?”

   Show how to answer
3. 3

   Tell me about a project where you did as much engineering as research. What did the engineering unlock?

   What they're checking: Whether you'll build your own tools. The Interpretability posting says "every team member writes code, designs and runs experiments, and interprets results."

   A strong answer: STAR. The tooling or infrastructure you built, the experiment it made possible, and how the result depended on it. Be specific about what you wrote yourself.

   Common miss: Treating engineering as support work someone else did while you had the ideas.

   Likely follow-up: “What would you build differently if you started that project today?”

   Show how to answer

The six topics behind behavioral questions, and how to prep a story for each →

1. 1

   Give us a concise overview of one research project and defend the most consequential methodological choice you made

   Senior

   What they're checking: Depth and ownership in the research discussion, which the Fellows posting names as its own stage.

   A strong answer: Problem, prior work, hypothesis, method, evidence, limits, next experiment, in about three minutes. Then go deep on one choice: the alternative you rejected, why, and the result that would have made you switch. Make your own contribution explicit.

   Common miss: A tour of the whole paper at one depth, with "we" hiding what you personally did.

   Likely follow-up: “What is the strongest alternative explanation for your main result?”

   Show how to answer
2. 2

   Implement multi-head self-attention from scratch in PyTorch or NumPy, then walk through each tensor's shape.

   What they're checking: Fluent ML implementation without a library call. One research candidate was told their Colab round would require QKV attention.

   A strong answer: Write the query, key and value projections, scale by the square root of the head dimension, apply a causal mask, use a numerically stable softmax, and state each reshape between batch, head and sequence as you go.

   Common miss: Getting the math right but losing track of shapes, then debugging silently instead of narrating.

   Likely follow-up: “How would you add a KV cache for generation?”

   Show how to answer
3. 3

   Build an in-memory banking system in stages: record transactions, return the top k accounts by outgoing money, add scheduled transfers that can be cancelled, then merge two accounts.

   What they're checking: Clean, extensible code under time pressure. This mirrors a Blind description of the 90-minute CodeSignal assessment.

   A strong answer: Pick a data model you can extend, keep earlier levels passing as you add features, and write small helpers rather than one growing function. Don't over-engineer level one.

   Common miss: Designing for all four levels up front and running out of time on level two.

   Likely follow-up: “How would merging change if transfers could be scheduled between the two accounts being merged?”

   Show how to answer
4. 4

   How would you tell, before it answers, whether a language model knows the answer or is about to hallucinate?

   What they're checking: Whether you know the calibration literature and can turn it into a testable method.

   A strong answer: Frame the goal as predicting correctness before generation. Assume access to activations. Anthropic's "Language Models (Mostly) Know What They Know" trained models to predict P(IK), and its circuit-tracing work points to a default "can't answer" pathway. Name the trade-off between a probe and a behavioral test, pick one, and say how you'd validate it on held-out domains.

   Common miss: Proposing to check the answer afterward, which doesn't answer the question.

   Likely follow-up: “Your probe works in-domain and fails on a new task. What does that tell you?”

   Show how to answer
5. 5

   Make the case that a model's chain of thought is not a faithful account of how it reached its answer. What experiments would test that?

   Senior

   What they're checking: Whether you can argue a research position with evidence, then design the experiment that could prove you wrong.

   A strong answer: Cite the evidence: Anthropic's 2023 paper truncated, corrupted and paraphrased chains of thought and found larger models often less faithful; its 2025 study found Claude 3.7 Sonnet mentioned a hint it used only 25% of the time. Propose an intervention test, then argue the other side: what CoT monitoring is still good for.

   Common miss: Asserting unfaithfulness from intuition with no experiment that could falsify it.

   Likely follow-up: “If the chain of thought isn't faithful, is monitoring it useless?”

   Show how to answer
6. 6

   You find a feature that seems to fire on deceptive text. How would you show it causally drives the model's behavior?

   Senior

   What they're checking: Whether you separate correlation from causation in interpretability work.

   A strong answer: Frame the claim precisely. Top-activating examples only show correlation, so propose ablation and steering interventions, controls for confounded features, and a behavioral metric that should move. Say in advance which result would make you drop the claim.

   Common miss: Showing more activating examples as if they were causal evidence.

   Likely follow-up: “Ablating the feature changes nothing. Is the feature irrelevant?”

   Show how to answer

1. 1

   You have two weeks and a modest compute budget to make progress on an open question from a recent Anthropic paper. Pick the question, propose three experiments, and choose one.

   Senior

   What they're checking: Research taste: problem choice, cheap falsification, and committing to one path.

   A strong answer: Frame the question and why it matters. State what you're assuming about compute and access. Propose three experiments that could each fail fast, name what each trades off, pick one, and say what result in week one would tell you to switch.

   Common miss: An ambitious plan that needs a month of compute before it can tell you anything.

   Likely follow-up: “Your first experiment is inconclusive after four days. What now?”

   Show how to answer