How to prepare for a machine learning oral exam

Posted on Oct 12 • Originally published at four-leaf.ai

A machine learning oral exam tests whether you can explain a model out loud and then defend that explanation when the examiner pushes on it. The authors of a 2026 paper on AI-run oral exams at NYU Stern argue that a scripted answer without real understanding breaks down within one or two follow-ups. That makes the useful preparation spoken, done under questioning, and started weeks before the exam rather than the night before.

There's no single format. The three published examples below have one thing in common, which is that you explain something out loud to an examiner who can question you about it.

Read your own syllabus before anything in this post. The format decides what you practice.

Reading builds recognition. You see the bias-variance tradeoff on a slide, it looks familiar, and familiarity feels like knowing it. An oral exam asks for something else, which is producing the explanation from nothing and then extending it in a direction you didn't choose.

The follow-up is the part you can't prepare a script for. In the NYU Stern paper on the AI/ML Product Management exam, the examiner's questions are calibrated to what the student just said, and the authors argue that "a scripted answer that does not reflect actual understanding breaks down within one or two follow-ups." The project phase made this concrete by asking about each student's "goals, data choices, modeling decisions, evaluation approaches, and failure modes," which nobody can improvise convincingly about work they didn't do.

Expect follow-ups from a few families. Why does this work? When does it fail? What would you change if the data looked different? How would you know it was wrong? Each one starts from your previous sentence. That's why a student who can recite the derivation of logistic regression still stalls when asked what happens to it on perfectly separable data.

The practical lesson is that every topic on your list needs two layers. The first is the explanation. The second is the answer to "why" asked twice in a row.

The best published data says students in one AI-run oral exam found it more stressful than a written one, and offers a small hint that experience didn't help. The safe plan is to prepare to perform while nervous rather than to stop being nervous.

The NYU Stern study surveyed both cohorts of its undergraduate AI/ML Product Management course after their AI-run oral exams. In the first cohort, 83% of respondents found the exam more stressful than written exams. In the second cohort, which sat the exam after several fixes to the examiner, 63% did. In the first cohort, 70% agreed the exam tested their actual understanding. Those numbers come from one course at one institution, so they say nothing reliable about human-run oral exams elsewhere.

One more detail in the NYU Stern data is about experience. You'd expect students who had sat oral exams before to be calmer. They weren't. Every respondent with prior oral-exam experience (all 5 in Fall 2025 and all 4 in Spring 2026) reported the AI exam more stressful. The groups are tiny, so this is a hint rather than a finding, but it's a hint against expecting familiarity alone to calm you down.

What practice does change is how well you think while your heart rate is up. If you've explained gradient descent out loud twenty times, the twenty-first time under stress still has a path to follow. Four-Leaf's guide to interview anxiety covers the in-the-moment techniques, and they transfer directly to an exam room.

Start from the syllabus, then turn each topic into four questions you can answer in under two minutes without notes:

The fourth question is the one to spend the most time on, because it's where follow-ups naturally go. For a linear model, it's what happens with correlated features. For a decision tree, it's overfitting and what pruning trades away. For a neural network, it's vanishing gradients, overfitting and what each regularizer actually does to the weights.

If your exam centers on a research article, Aalborg University's Machine Intelligence course description makes a good checklist, since it asks students to present an article's premises, problem, theory, methods, results and conclusions. Can you state the paper's premise, the problem it addresses, the theory behind it, its method, its results and its conclusions? Then can you say where the method would break and what the authors didn't test? That last step is what "assess their qualities and practicality" means in practice.

For a project-based exam, rehearse your own decisions. Why that dataset, why that model, why that metric, and what you'd do with another month. Questions about why you built it that way test understanding in a way the finished report can't.

Practice the way the exam runs, which means speaking, being interrupted and answering one question at a time.

If you don't have a practice partner, a voice mock interview can stand in for the follow-up drill. Four-Leaf's voice mock interviewer is instructed to ask one question per turn and usually follows up on the answer you just gave before moving on. It's designed for job interviews rather than coursework, so bring your own topic list. Four-Leaf's guide to practicing interviews alone covers the recording and mirror methods in more detail.

Rewriting your notes. It feels productive and it builds recognition, which is the skill the exam doesn't test. Time spent tidying a summary sheet is time not spent saying the material out loud.

Memorizing model answers is overrated too, for the reason the NYU Stern paper on AI oral exams gives. A rehearsed paragraph covers the first question and none of the ones after it. Memorize definitions, key equations and the names of things. Then practice saying everything else in your own words, slightly differently each time, so you aren't depending on one phrasing.

Closely. The authors of the NYU Stern paper on AI oral exams extend their argument past the classroom to "a job interview, a professional licensing board" and anywhere else a polished artifact stands in for understanding that ought to be tested directly. An interviewer asks you to explain regularization, then asks what happens when you double the penalty, then asks how you'd detect that in production.

That overlap is useful if you're finishing a degree and heading into the job market. The habit of explaining a model and then defending it is the same habit the ML fundamentals and system design rounds test. Four-Leaf's breakdown of MLOps and ML engineer interview questions maps the four rounds those loops usually run, and the machine learning engineer interview prep page covers the technical, system design and behavioral questions those loops ask.

The difference is stakes and scope. An exam tests a syllabus you were given. An interview tests whatever the team cares about, plus production questions no course covers. Treat the oral exam as the rehearsal it is.

The NYU Stern paper on AI oral exams opens with an argument every ML student should take seriously. "Written work no longer certifies that a student understands it," the authors write, and "oral examinations restore that evidentiary link." Their system graded a full cohort's oral exams for under one dollar per exam, on top of a voice subscription that covered the speaking time. When a format that resists ghostwriting gets that cheap to grade, expect to meet it more often.

That makes explaining your work out loud, under questioning, a skill worth building on purpose. Before the exam, get asked "why" twice on every topic, and make sure you have an answer both times.

Four-Leaf is the all-in-one AI job search assistant covering every stage from application to signed offer: voice mock interviews across 24 roles, resume tailoring, cover letters, salary negotiation, email drafting, AI job search, and LinkedIn optimization. 3-day free trial, no credit card. Pricing: $20/mo or $5 / 5 Day Pass.