Research Project Deep Dive and Your AI-Agent Research Workflow

This is a 45-minute technical phone interview for a research internship. After about 10 minutes of introductions on both sides, the interviewer spends about 20 minutes on a deep dive into one of your research projects. Then they ask extension questions tied to the problem the intern would work on, and finally they ask how your day-to-day research workflow looks now that AI agents are available: writing papers, running experiments, and coming up with ideas. Treat the last question as a real probe of how you work, not as small talk. ### Clarifying Questions - Should the deep dive cover the project you consider strongest, or the one closest to the team's work? - How technical should it be: loss functions, training details and ablations, or a higher-level narrative with the option to go deeper? - What problem would the intern work on, and what does the team's current approach to it look like? ### Part 1 — Project deep dive Pick one project and walk the interviewer through its motivation, the research question, the training pipeline, the experimental results, and the major takeaways, in about 20 minutes. Expect to be interrupted and asked to go deeper on any of these. ```hint One question, one thread Choose the single question your project answers and make every later section (pipeline, experiments, results) point back to it; know the numbers you will be asked to defend. ``` #### What This Part Should Cover - A motivation and research question that are precise and show why the problem matters - A training pipeline whose design choices are justified, not just listed - Results against baselines, with ablations, variance and honest limitations - Takeaways that go beyond this project ### Part 2 — Extending your work to the internship problem The interviewer describes the problem the intern would work on and asks how your research would extend to it, or what you would try first. ```hint Transfer, change, test Separate what carries over from your project (methods, infrastructure, lessons) from what would have to change, then name a first experiment and how you would judge it. ``` #### What This Part Should Cover - Questions that pin down the team's problem, data and success metric - A specific mapping from your methods and lessons to that problem - A scoped first experiment with a success criterion, risks and a fallback ### Part 3 — Research workflow with AI agents Describe your day-to-day research workflow now that agents exist: how you use them to write papers, run experiments, and come up with ideas, and where you do not use them. ```hint Delegate, verify, and know where it breaks For each activity, be concrete about what you hand to an agent, how you check the output, and a failure you have seen or would expect. ``` #### What This Part Should Cover - Concrete uses across writing, experiments and ideation - Verification habits that protect correctness and reproducibility - Limits and failure modes of agents in research, and integrity rules such as citations and disclosure - An opinion on how automation changes what a researcher spends time on ### What a Strong Answer Covers - A deep dive that can go two or three levels down on any design choice without losing the storyline - Quantitative, honest results, including negative results and what they taught you - Good use of the time slot: a clear headline first, detail on demand - A realistic, specific account of working with agents, with verification built in - Curiosity about the team's problem and a credible plan for contributing to it ### Follow-up Questions - What was the most important idea in this project that failed, and how did it change the direction of the work? - With ten times the compute or data, what would you do differently, and what would you expect to learn? - Which part of your research workflow would you never hand to an agent, and why? - How would you measure whether agents actually make you a more productive researcher, rather than just faster at producing text and code?