How ML theory interviews work
The ML theory round, also called ML breadth or ML fundamentals, is a run of short spoken questions: bias and variance, regularization, losses and optimizers, evaluation metrics, data leakage, training problems in deep networks and, more and more, the basics of language models. The guides we read agree that a textbook definition is only the starting point. The signal is applied judgment: when you would use a method, which knob you would turn for a given symptom, and how you would check that it helped.
Key points
- Give the definition in a sentence, then say when you would use it and how you would know it worked.
- Expect “why?” after every answer: why L1 gives sparse weights, why the loss diverges, why this metric here.
- Name the trade-off, not only the method. Every fix for overfitting costs something.
- Choose metrics for the problem, not by habit. Accuracy on imbalanced data is the classic trap.
- Some rounds turn into a short math quiz: a derivative, a Bayesian update or a probability worked by hand.
What the round looks like
ML theory shows up in several shapes, sometimes as its own round and sometimes folded into a coding or design round. Guides and practitioner write-ups describe these:
| Format | What happens |
|---|---|
| Breadth questions | A run of short questions across the field, each followed by one or two “why” or “what if” probes. |
| Applied scenarios | A symptom to explain: the model overfits, the loss will not converge, training accuracy is high and validation accuracy is not. |
| Math quiz | Short derivations by hand, such as the gradient of softmax, a Bayesian update or why one divergence is not symmetric. Reported mostly at finance firms. |
| Depth on your area | Questions on the domain of the team you would join, such as ranking or language models, or on the models you have shipped. |
What interviewers look for
No company publishes its rubric for this round. The guides describe the same signals:
- Accuracy: correct definitions and the conditions they hold under.
- Mechanism: why a method works, not just its name, such as how regularization trades variance for bias.
- Applied judgment: why you would choose a method for this data and this goal, and how you would check that it helped.
- Data awareness: missing values, scaling, outliers and leakage, not only the model.
- Metric fit: a metric chosen for the problem and the cost of each kind of error.
- Clear explanation: walking through the reasoning, in a short answer a colleague could follow.
The topics that come up most
- Generalization: bias and variance, overfitting and underfitting, regularization, cross-validation and early stopping.
- Losses and optimization: common loss functions, gradient descent and its variants, learning rates and schedules, convergence.
- Classical models: linear and logistic regression and their assumptions, decision trees and ensembles, k-means, principal component analysis.
- Evaluation: precision and recall, ROC and precision-recall curves, imbalanced data, choosing a threshold, tuning hyperparameters.
- Deep learning: backpropagation, activations, vanishing and exploding gradients, initialization, normalization and dropout.
- Data: missing values, encoding categories, scaling, feature selection and leakage.
- Language models: attention and the transformer, tokenizers, fine-tuning and retrieval.
The mistakes guides report most
- Shaky fundamentals: not being able to explain overfitting, the bias-variance trade-off or the common metrics clearly.
- Naming an algorithm without saying why it fits this data and this goal.
- Memorised answers to a list of common questions that break at the first “why?”.
- Leaving the data out: missing values, scaling, outliers and leakage.
- Jumping to the answer without walking through the reasoning.
How the questions chain
Questions tend to come in chains rather than one by one: how you know training has converged, then how you tell overfitting from underfitting, then how you would regularize, then why that works. One practitioner’s advice is to prepare by revisiting the foundations rather than a list of answers, because the chain soon goes past any list.
In the math-quiz style, one essay written from the interviewer’s side warns that these questions reject many good candidates who were not trained for them, and recommends a mix of theory, calculation and explanation questions. If your loop includes one, practise derivations out loud, not only on paper.
Depth on the team’s domain
Many ML roles hire into a specific team, so one practitioner reports that the theory round often goes deep into that team’s area, such as search, recommendations, ads or language. The same write-up reports that language-model questions mostly stay on the basics: the original transformer, common tokenizers, well-known model families, and widely used techniques such as preference tuning, retrieval-augmented generation and low-rank adapters. The paper from last week is rarely the point.
Sources on how the bar moves with level are thin for this round, so we give no level table here.
Go deeper on this round
Practise the round out loud
Mockich’s ML theory round runs 35 min: six questions at regular and five at lead across core ML, deep learning, evaluation, language models and debugging models, asked out loud one at a time. Vera probes each answer once or twice for depth and gives at most one small hint. The report scores accuracy, depth, practical judgment, communication (plus breadth and connection at lead) and quotes what you said.
Sources
We read these public write-ups on 2026-11-02 and summarised where they agree, in our own words. Sources 1 and 2 are one practitioner and count as one voice; source 4 is mostly a question list and is used for topics only. They report typical practice; no company publishes how it scores this round, and formats differ between companies and teams.
- [1] Yuan Meng, (Opinionated) Guide to ML Engineer Job Hunting
- [2] Yuan Meng, MLE Interview 2.0: Research Engineering and Scary Rounds
- [3] Brian Kihoon Lee (moderndescartes.com), Interviewing for ML/AI Engineers
- [4] DataJourney (Substack), Acing Machine Learning Breadth Interviews
- [5] MeetAPro, Common Pitfalls in ML Interviews and How to Avoid Them