Data Science Interview Topics in 2026: Classic Stats vs GenAI

Last updated: October 2026.

Quick answer: Data science interviews in 2026 test two parallel tracks. The classic track covers SQL, Python, statistics, and ML model design. The GenAI track covers transformer fundamentals, embeddings, RAG pipelines, fine-tuning trade-offs, and LLM evals. Finance firms and product analytics roles weight the classic track heavily. Companies building AI products add the GenAI track, especially for senior and ML-forward roles. If you don't know which track your target company weights, it's easy to spread your prep too thin.

The interview content has shifted, but not evenly. A statistics-first finance firm and a GenAI-forward tech startup will weight very different subjects. Most prep guides hedge by covering everything, which doesn't help much if your interview is in three weeks.

What Are the Two Tracks in a 2026 DS Interview?

Think of them as two skill sets that coexist, with different companies weighting them differently.

Track 1: Classic data science. This has been stable for years: SQL and Python fluency, probability and statistics, A/B testing and experiment design, and ML model selection and evaluation. These skills are tested at nearly every company.

Track 2: GenAI engineering. For roles at companies actively building AI products, interviewers also test: how LLMs are trained at a conceptual level, what embeddings are and how vector search works, how to design a RAG pipeline, when to fine-tune instead, and how to evaluate an LLM feature in production. A 2026 data scientist interview guide from KORE1 notes that LLM and MLOps questions now appear regularly in senior loops, and TechInterview calls choosing between RAG, fine-tuning and prompting one of the most-asked AI/ML interview questions of 2026.

The tracks are not mutually exclusive. A senior DS role at a tech company will test both.

What Does the Classic Stats Track Cover?

This is the part most candidates prep for. It is still the majority of the interview at many companies.

Hypothesis testing and A/B testing draw the most questions, and the questions tend to be practical rather than definitional. Knowing what a p-value is gets you through the screen. Knowing how to design a test, set guardrail metrics, handle unbalanced traffic, and correct for multiple comparisons gets you through the final rounds.

A/B testing questions almost always extend into edge cases. What do you do when you're running 40 tests simultaneously? How long do you run a test before calling it? What happens when your primary metric improves but a guardrail metric degrades? Those edge cases are where candidates who prepped only the mechanics fall short.

Probability and distributions show up at every level. The depth scales with seniority: early-career questions test distributions, expected value, and basic combinatorics; senior questions test experimental power calculations, Bayesian inference, and causal identification.

ML model design covers model selection, bias-variance trade-offs, regularization, cross-validation, and evaluation metrics. These are applied: given a business problem and dataset, which approach makes sense and why? A good answer weighs the trade-offs instead of naming the most sophisticated option.

SQL and Python remain a baseline expectation everywhere. Some companies use SQL as a filtering round that happens before any modeling conversation. Getting screened out on window functions before you reach the statistics discussion is an avoidable way to lose the loop.

What Does the GenAI Engineering Track Cover?

For roles at companies building with LLMs, a second set of topics has appeared. For most roles, they sit on top of classical statistics rather than replacing it.

LLM fundamentals. You should be able to explain how a transformer processes input, what embeddings represent, and why cosine similarity is used to find similar content. Interviewers want conceptual understanding; you don't need to memorize architectures.

RAG pipelines. A typical senior interview question asks you to design or critique a retrieval-augmented generation system. The standard pipeline runs: chunk documents, generate embeddings for each chunk, index the vectors, retrieve the nearest neighbors for a query, optionally rerank them, then pass the retrieved context to the model. Interviewers follow up on failure modes: what breaks when the chunks are too large or too small, how you handle access control so users cannot retrieve content they shouldn't see, when this approach fails and fine-tuning would work better instead.

Fine-tuning approaches. You do not need to have fine-tuned a model in production, but you should be able to explain when it makes sense to fine-tune instead of using RAG, and what parameter-efficient fine-tuning methods (like LoRA) offer over full fine-tuning.

Evals. How do you judge whether an LLM feature works? Common evaluation approaches include labeled reference sets, LLM-as-judge (using a stronger model to grade outputs against a rubric), and task-specific benchmarks. Interviewers ask candidates to design an eval suite for a new feature, explain what makes a judge prompt unreliable, and detect when a deployed model drifts.

MLOps and model drift. For senior roles, you should be able to walk through how you would diagnose a degradation in model performance: did the input distribution shift, did the feature pipeline break, or did data quality degrade upstream?

How Do DS Interview Topics Vary by Company Type?

Finance interviews, especially at banks and quantitative firms, weight statistics heavily. A/B testing questions focus on methodological rigor rather than product framing. Risk modeling and causal inference appear in ways that differ from tech-product roles. Treat the finance DS track as close to quant interview preparation in its emphasis on statistical correctness.

Startup loops are often the least predictable. The take-home case tends to carry the most weight, and strong candidates frame their solutions around business impact rather than algorithmic sophistication.

Which Track Should You Prep First?

Prep the track your target company emphasizes. Look at the job description: does it mention LLMs, RAG, or evals? Or does it emphasize experimentation, metrics, and statistical modeling? That tells you more than any general guide.

If you are targeting Big Tech, both tracks matter. Classical statistics and metrics remain the core of the analytics and experimentation rounds; GenAI system design appears in technical and design rounds for senior roles. Do not skip either.

If you are targeting finance, go deep on classical statistics and probability. GenAI topics come up much less often here, unless the role sits on an AI product team.

For SQL and Python: these are never optional. Build fluency with window functions, CTEs, and pandas data manipulation before the interview loop starts.

How Do You Practice Both Tracks?

The classical statistics track responds well to deliberate case practice: pick an A/B testing scenario with an edge case (low power, multiple comparisons, conflicting metrics) and work through it out loud. Behavioral interview prep with the STAR framework helps for the rounds that test how you communicate decisions.

The GenAI track is harder to practice with flashcards. The questions are applied: design a pipeline, critique an eval setup, debug a drift scenario. Practice them in mock technical interviews where you explain your reasoning out loud and get pushed on trade-offs. In PokeBot's Interview Practice, you can run Technical and DS Project Deep-dive sessions out loud and get scored feedback, with optional follow-up questions that push on what your answer left out.

Where Should You Start If You Have Three Weeks?

Three weeks is enough to cover the essentials if you are focused. Week one: SQL fluency and Python data manipulation, plus hypothesis testing mechanics. Week two: A/B testing edge cases, ML model selection, and LLM fundamentals. Week three: RAG pipeline design, evals, and mock interviews that force you to apply everything under time pressure.

Aim to reason through an applied problem in any of these areas out loud, without backtracking. That takes practice as well as reading.

Treat the three-week schedule above as your starting point, and adjust the time on each area based on what your target company emphasizes.

Score your resume free, create your PokeBot account