9 questions that keep showing up in Google MLE interviews (2025–2026)

I spent the last few months going through recent Google MLE interview reports from US loops. Same questions, different people, over and over. Here are the 9 that show up most.

1. Bias-variance tradeoff — but the real question is: when would you deliberately add bias?
   They don't want the textbook definition. They want to hear when you trade a little bias for a lot less variance, and how you know you made the right call.

2. Why cross-entropy instead of MSE for classification?
   Say "because MSE is flat when the model is confidently wrong." Then walk through the gradient.

3. Attention: why divide by sqrt(d_k)?
   Without the scaling, dot products grow with dimension and softmax saturates. One line of math, that's the whole answer.

4. ROC-AUC vs PR-AUC — which one, and when?
   Imbalanced data: PR-AUC. Balanced data: either works. If you say "always ROC-AUC" you lose points.

5. How do you catch data leakage in a training pipeline?
   Name one real check: train on data up to time T, validate only on data after T. Leakage hides in features that quietly contain the answer.

6. Training/serving skew — how do you find it before users do?
   Log features at serving time, compare distributions against training. If they drift, your offline metrics are fiction.

7. L1 vs L2 regularization — when does L1 actually zero out features?
   Only when the data agrees. L1 doesn't magically do feature selection; it does it when features are truly useless.

8. Cold start in a recommender — what do you do on day one with zero user history?
   Popularity + content features first, then switch to learned embeddings as data arrives. Say how you know when to switch.

9. Vanishing gradients — name two fixes that actually work in practice.
   Residual connections and proper initialization. Batch norm helps, but know why it helps.

I wrote up all 9 with full answers, plus the 80+ question Google checklist I used to prep: https://mleplaybook.gumroad.com/l/ewryk

$19. If you're interviewing at Google in the next few weeks, start with questions 3, 4, and 5 — those three alone came up in almost every report I read.