Interview Questions

AI Researcher Interview Questions

10 questions you are likely to meet in an AI Researcher interview. For each one: what a strong answer covers, and the free lesson that prepares it.

  • 10 Questions
  • 8 Skills
  • Reviewed 16 November 2026

How the Interview Usually Runs

Usually a deep technical round on machine learning and math, a paper discussion, a coding or implementation round and a research talk.

Foundations

The concepts interviewers check first.

  1. Explain self-attention, and why transformers replaced recurrent networks.

    What a strong answer covers

    Each token builds queries, keys and values; attention weights come from query-key similarity and mix the values. Unlike recurrent networks, all positions are processed in parallel and long-range links are direct, which scales training on GPUs.

    Prepared InS7 · #80Transformers Explained, coming in Season 7

  2. What is the difference between pretraining, supervised fine-tuning and RLHF?

    What a strong answer covers

    Pretraining learns next-token prediction on huge text; supervised fine-tuning teaches instruction following from examples; RLHF or DPO tunes the model toward human preferences using comparisons.

    Prepared InS7 · #82How LLMs Are Made, coming in Season 7

  3. Explain backpropagation in plain words.

    What a strong answer covers

    After a forward pass computes the loss, the chain rule passes the error backward layer by layer to find how much each weight contributed, giving gradients that gradient descent uses to update the weights.

    Prepared InS7 · #73A Neural Network From Scratch, coming in Season 7

Hands-on

How you would actually build or do it.

  1. Why is cross-entropy the loss for language models?

    What a strong answer covers

    It measures how much probability the model gave to the correct next token; minimizing it is maximum likelihood. It pairs naturally with softmax outputs and gives useful gradients.

    Prepared InS5 · #61Loss, Softmax, Cross-Entropy, coming in Season 5

  2. How would you implement a small GPT from scratch?

    What a strong answer covers

    A tokenizer, token and position embeddings, a stack of transformer blocks with causal self-attention and residual connections, a final projection to the vocabulary, a training loop with cross-entropy, and sampling to generate.

    Prepared InS15 · #183Build an LLM From Scratch 1, coming in Season 15

  3. How do you read a research paper quickly and decide if it is worth reproducing?

    What a strong answer covers

    A first pass on title, abstract, figures and conclusion; a second on the method and experiments; a third only if it matters. Check baselines, ablations, code availability and whether the gains are large and robust.

    Prepared InS15 · #185Reading AI Papers, coming in Season 15

Scenario

Open problems: design, debugging and judgment.

  1. What are scaling laws, and what do they imply?

    What a strong answer covers

    Loss falls predictably as a power law with more parameters, data and compute. They guide how to split a compute budget between model size and data, though downstream abilities do not always follow smoothly.

    Prepared InS7 · #82How LLMs Are Made, coming in Season 7

  2. Your new method beats the baseline by one point. How do you convince yourself it is real?

    What a strong answer covers

    Run several seeds and report variance, tune the baseline equally, run ablations, test on more datasets, check for contamination and data leakage, and apply a significance test.

    Prepared InS15 · #185Reading AI Papers, coming in Season 15

  3. What is benchmark contamination, and how do you guard against it?

    What a strong answer covers

    Test questions leaked into training data, inflating scores. Guard with held-out or fresh test sets, overlap checks between training and test data, and evaluation on private tasks.

    Prepared InS15 · #185Reading AI Papers, coming in Season 15

Behavioral

Your experience, told with the STAR method.

  1. Describe a research idea that failed, and what you learned.

    What a strong answer covers

    A STAR story: the hypothesis, the experiments, the evidence that it did not work, how you decided to stop, and what the failure taught you or the team.

    Prepared InS17 · #224Behavioral Interviews and Negotiation, coming in Season 17

Skills Checklist

What AI Researcher interviews test, each with the free lesson that teaches it. Tick what you can already explain.

0 of 8 ready

Your Next Step

Learn Everything These Questions Test

The AI Researcher path takes you there lesson by lesson, from zero, for free.

Prepare for the Interview, Free

Every ai2hired lesson is built from real AI job ads, and the last season is all about getting hired.