All 46 questions

  1. What is the basic difference between RAG and fine-tuning, and when would you use fine-tuning or RAG?

    Asked in 3 of 440 interviews · last asked Oct 21 · Concept · Mid-level

    Show answer

    RAG retrieves relevant information and gives it to the model as context, without changing its weights. Fine-tuning trains the model on examples and changes its weights. Use RAG for current or private knowledge that needs source references. Use fine-tuning for consistent style, output format, or task behavior.

    • RAG lets you update source documents without retraining the model.
    • Fine-tuning needs good training examples; it is not a reliable way to store changing facts.
    • RAG depends on retrieval quality and can still produce unsupported answers.
    • You can combine RAG for knowledge with fine-tuning for behavior.

    What the interviewer checks: The interviewer is checking whether you can separate knowledge access from learned behavior and choose the right approach.

    Other ways it was asked
    • “When would you choose RAG over fine-tuning?”
    • “Can you briefly explain the difference between RAG and fine-tuning?”
  2. How would you measure whether fine-tuning improved the model?

    Asked in 3 of 440 interviews · last asked Sep 27 · Scenario · Mid-level

    Show answer

    I would compare the original and fine-tuned models on the same held-out test set that reflects real use. I would use task-specific metrics and blind human review, with the same prompts and generation settings. I would check whether gains are reliable, test for regressions, and compare latency and cost.

    • Keep test data separate from training and model selection to avoid leakage.
    • Choose metrics that match the task, such as accuracy, factual correctness, or instruction following.
    • Use enough test cases and confidence intervals to separate real gains from noise.
    • Check important user groups, safety, and general skills, not just the average score.

    What the interviewer checks: The interviewer is checking whether you can design a fair comparison and show useful gains without hiding regressions.

    Other ways it was asked
    • “How would you determine whether fine-tuning actually improves a model?”
    • “How would you systematically compare a model's performance before and after fine-tuning?”
  3. What is model fine-tuning, and when would you use it?

    Asked in 2 of 440 interviews · last asked Jul 29 · Concept · Mid-level

  4. How do you fine-tune small language models?

    Asked once · last asked Sep 25 · Concept · Mid-level

    Other ways it was asked
    • “What are the different ways to fine-tune small language models?”
  5. How would you handle an agent that uses both static company-policy information and dynamic patient or insurance data, and where would fine-tuning fit?

    Asked once · last asked Feb 5 · Design · Senior

  6. What is LoRA (low-rank adaptation), and how does it work?

    Asked once · last asked Jun 12 · Concept · Mid-level

  7. How do you format prompts and training examples for fine-tuning a model to produce structured outputs?

    Asked once · last asked Jul 16 · Scenario · Mid-level

  8. What is the difference between full fine-tuning and LoRA?

    Asked once · last asked Jul 16 · Concept · Mid-level

  9. What is the difference between training a model and changing its system prompt?

    Asked once · last asked Jul 16 · Concept · Mid-level

  10. What are the main considerations when training or retraining an LLM?

    Asked once · last asked Jul 16 · Concept · Senior

  11. When do you choose fine-tuning versus RAG versus prompt engineering?

    Asked once · last asked Jul 22 · Concept · Senior

  12. How would you approach a model customization process from a high level, and what information would you gather before starting?

    Asked once · last asked Jul 27 · Scenario · Senior

  13. Why would you fine-tune an LLM, and what process would you follow?

    Asked once · last asked Aug 13 · Concept · Mid-level

    Other ways it was asked
    • “Why did you fine-tune the LLM, and what was the process?”
  14. How would you fine-tune an LLM for code generation, and when would you use a fine-tuned model instead of a base LLM?

    Asked once · last asked Aug 13 · Concept · Mid-level

  15. An engineering team is fine-tuning a model using quantized LoRA (QLoRA) on limited GPU infrastructure and repeatedly encounters out-of-memory errors. How would you mitigate these errors?

    Asked once · last asked Aug 20 · Scenario · Senior

    Other ways it was asked
    • “The engineering team is doing a quantized LoRA for a model on limited GPU infrastructure and keeps getting out-of-memory errors. How could we help them mitigate these errors?”
  16. What is the difference between fine-tuning and in-context learning for document processing?

    Asked once · last asked Aug 25 · Concept · Mid-level

  17. How would you fine-tune a model when no training dataset is available?

    Asked once · last asked Sep 5 · Scenario · Senior

  18. In which scenarios should we use supervised fine-tuning, reinforcement learning, GRPO, or PPO for small language models?

    Asked once · last asked Sep 25 · Concept · Senior

  19. For an LLM answering questions about frequently changing internal policies, would you fine-tune it or use another approach?

    Asked once · last asked Sep 27 · Scenario · Senior

  20. When building an enterprise AI feature, what key factors help you decide between using advanced RAG and fine-tuning?

    Asked once · last asked Oct 20 · Design · Senior

  21. How would you adapt a six-item pass/fail rubric to an RLHF preference task with two candidate responses so annotators consistently choose A or B?

    Asked once · last asked Oct 21 · Scenario · Senior

    Other ways it was asked
    • “How would you adapt those same six pass or fail items to an RLHF preference task where the model gives two candidate responses? What would annotators look for to choose A or B consistently?”
  22. How would you design a preference tool and workflow for RLHF to keep rater consistency high over time, including rubric, calibration, goldens, adjudication, and measuring drift in rater behavior?

    Asked once · last asked Oct 22 · Design

  23. When training a reward model from human preference labels, how would you detect and mitigate rater bias or position bias so the reward does not overfit annotator quirks?

    Asked once · last asked Oct 22 · Scenario · Senior

    Other ways it was asked
    • “When training a reward model from those preferences, how would you detect and mitigate rater bias or position bias so the reward does not overfit annotator quirks?”
  24. In RLHF, how would you build an evaluation suite to detect reward hacking and regression in safety beyond just tracking average reward score?

    Asked once · last asked Oct 22 · Scenario · Senior

  25. When training sparse autoencoders to decompose activation spaces within transformer models, how do you distinguish meaningful features from dead features?

    Asked once · last asked Oct 26 · Scenario · Senior

  26. Why would you choose RAG over fine-tuning for a document-heavy company?

    Asked once · last asked Oct 27 · Concept · Mid-level

    Other ways it was asked
    • “Why did you choose RAG over fine-tuning for a document-heavy company?”
  27. When a fine-tuning job fails on a GPU pod, how do you use the logs to identify the cause?

    Asked once · last asked Nov 2 · Scenario · Mid-level

  28. What factors can make failures in GPU fine-tuning jobs difficult to reproduce?

    Asked once · last asked Nov 2 · Scenario · Mid-level

  29. How would you define objectives and metrics for reducing a model's psychopathy-related behavior while preserving reasoning, and design the data split, training, and steering-intervention workflow?

    Asked once · last asked Nov 4 · Scenario · Senior

    Other ways it was asked
    • “How would you define an objective for reducing psychopathy-related behavior while preserving reasoning, and design the data split, training, and steering-intervention workflow?”
  30. How would you structure a run that includes a data split, sparse autoencoder training, and an inference-time intervention? What would you hold fixed, and how many trials or examples would you evaluate at each stage?

    Asked once · last asked Nov 4 · Scenario · Senior

  31. How would you shape the reward to avoid shortcuts, such as deleting tests or overfitting to a narrow test suite, while still providing useful intermediate signals over a long horizon?

    Asked once · last asked Nov 4 · Scenario · Senior

  32. How would you design a synthetic data generation and grading pipeline to train an RLHF model to write safe code patches, ensuring the grader is hard to game and correlates with real-world correctness?

    Asked once · last asked Nov 4 · Design · Senior

  33. When automated graders disagree—for example, tests pass but static analysis or security checks raise flags—how do you combine their signals into a single reward while preserving safety and minimizing false positives?

    Asked once · last asked Nov 4 · Scenario · Senior

  34. How would you measure and monitor reward gaming over time? What concrete metrics or canary tasks would you add to detect reward hacking and distribution shift?

    Asked once · last asked Nov 4 · Scenario · Senior

  35. You see clear reward gaming; what's your step-by-step mitigation plan in training, for example, reward redesign versus data filtering versus constraint hardening? And how do you verify it does not reduce helpfulness?

    Asked once · last asked Nov 4 · Scenario · Senior

  36. What role do small language models play in enterprise agents?

    Asked once · last asked Nov 4 · Concept · Senior

  37. What is reinforcement learning from human feedback (RLHF), and when would you use it?

    Asked once · last asked Nov 11 · Concept · Mid-level

  38. What data preparation or evaluation workflows are used in LLM post-training?

    Asked once · last asked Nov 13 · Tooling · Mid-level

  39. What does user-side polishing and training of an AI model look like in practice?

    Asked once · last asked Nov 18 · Scenario · Senior

  40. When would you use supervised fine-tuning (SFT) or LoRA?

    Asked once · last asked Nov 18 · Concept · Mid-level

  41. How do you compute the loss and backpropagate gradients when fine-tuning a model, and how do frozen versus trainable parameters affect the process?

    Asked once · last asked Nov 19 · Concept · Mid-level

  42. What does fine-tuning mean in the context of an LLM?

    Asked once · last asked Feb 5 · Concept · Junior

  43. What is LoRA?

    Asked once · last asked Jul 21 · Concept · Junior

  44. What is agentic reinforcement learning?

    Asked once · last asked Sep 25 · Concept · Senior

  45. Why would you choose RAG rather than fine-tuning for frequently changing internal policies?

    Asked once · last asked Sep 27 · Concept · Senior

    Other ways it was asked
    • “What are additional reasons to use RAG rather than fine-tuning for frequently changing internal policies?”
  46. Do you know what SFT and LoRA stand for?

    Asked once · last asked Nov 18 · Concept · Junior

More Generative AI and LLM interview questions

How this list is made

Every question here was asked by a real interviewer in a real job interview that a candidate ran MeetAssist in. After the call, the interviewer's questions are pulled from the transcript. Practice sessions, tests, exams and anything that was not a job interview are dropped.

Questions are rewritten so they make sense on their own: names of people, employers and clients are removed, and speech-to-text mistakes are fixed. Small talk, logistics and questions about one candidate's own résumé are left out. Rephrasings of the same question are merged, and the count is the number of different interviews it came up in.

The interviews come from candidates who use MeetAssist, so they lean towards software, data and AI roles: the shares describe these interviews, not every Generative AI and LLM interview. Answers to the most asked questions are written and checked by an editor; a "common mistake" is shown only when reviews of real answers in these interviews show it, with the count.

440 interviews, 243 candidates, Feb 2026 to Nov 2026. Updated every month with new interviews.