Question bank

4,310 interview questions, answered.

Reference answers, what the interviewer is really testing, how it is graded, and the follow-ups that come next.

Easy 190Medium 2,498Hard 982
FiltersDifficulty, topic, company

4,310 questions

  1. 4297A vectorized masked loss matches output shape but changes training. Design a minimal reference fixture.MediumCodingAI ML
  2. 4298A 120 ms step spends 80 ms waiting for data and 40 ms in compute. A kernel halves compute time. What ideal speedup follows, and what trace check is necessary?MediumAI MLTechnical
  3. 4299Loss scale is 512, true gradient norm is 4, and clip threshold is 1. Explain correct processing and the clip-before-unscale error. Use idealized norm arithmetic and ignore numerical epsilon in clipping.HardAI MLTechnical
  4. 4300Two ranks have gradient sums 8 and 4 over two and four valid tokens. Compute the global mean and the average of local means.HardAI MLTechnical
  5. 4301A reproduction matches sample IDs and logits, but loss and gradient norms are exactly three times larger for a three-example batch. What do you inspect before another full run?HardAI MLTechnical
  6. 4302Your optimization wins a benchmark only after excluding odd-length inputs. Those inputs are in the original contract. How should the result be reported?MediumAI MLTechnical
  7. 4303A method improves one benchmark after doubling context and parameter count. What claim is supported and what control tests the proposed memory mechanism?HardDataAI ML
  8. 4304You have 600 predictions from six independently trained models. The claim concerns variability across trained models. What is the resampling unit?HardDataAI ML
  9. 4305The best of 40 hyperparameter settings scores 88% on validation. A team wants to report 88% as final generalization. What should happen?MediumDataAI ML
  10. 4306A result estimates +0.3 points with an interval from -2 to +2.6, and useful improvement means at least +1. What can you conclude?MediumDataAI ML
  11. 4307A factorial study scores 60 for neither component, 63 for X, 64 for Y, and 70 for both. Calculate the interaction and state its limits.MediumDataAI ML
  12. 4308Method A scores 81 at 10 GPU-hours; B scores 83 at 30; C scores 80 at 20. Which is dominated, and how do tuning costs affect the report?MediumDataAI ML
  13. 4309A model is better within easy and hard groups but worse overall. What data do you request and how do you analyze it?HardDataAI ML
  14. 4310Your five-seed study supports a method on one language. An interviewer asks you to deploy it in a new language with a different label process. How do you revise the recommendation?HardBehavioralAI ML