Technical practice
How would you measure whether an LLM is genuinely more energy-efficient without sacrificing performance?
What to coverSuggested preparation: Define comparable baselines and report energy use, compute, latency or throughput alongside task quality. Explain how you would control for hardware and workload differences so an apparent efficiency gain is credible.
What approaches would you investigate to make foundation-model training more efficient?
What to coverSuggested preparation: Compare plausible algorithmic or systems-level approaches in terms of compute, energy use, training stability and final model quality. Explain how you would test whether an improvement holds beyond a single experiment.
How would you reduce the cost of LLM inference while preserving useful output quality?
What to coverSuggested preparation: Discuss model and serving trade-offs, including memory use, latency, throughput and output quality. Describe measurements that would reveal whether a change saves energy under realistic workloads.
How would you deploy a foundation model on a resource-constrained platform?
What to coverSuggested preparation: Start with the platform’s memory, compute and power limits, then explain how you would choose and evaluate model or system-level adaptations. Address what you would do if an optimization meets the resource limit but degrades performance too far.
How would you evaluate an optimized model’s NLP performance?
What to coverSuggested preparation: Choose task-relevant quality measures and compare the optimized model against an appropriate baseline. Discuss how you would look for uneven quality losses that an aggregate score might conceal.
How would you diagnose an unexpectedly expensive training or inference run?
What to coverSuggested preparation: Describe how you would separate model, data-processing and system bottlenecks using profiling and reproducible measurements. You could frame the investigation in Python and modern AI frameworks such as PyTorch or Hugging Face, which the posting names.
How would you frame an AI research project for a sustainability application?
What to coverSuggested preparation: Define both the application outcome and the AI system’s own resource costs. Explain how you would assess whether the proposed technology delivers a meaningful sustainability benefit rather than merely shifting computational costs.
What would make a Green AI result convincing enough to publish?
What to coverSuggested preparation: Explain the research question, novelty, strong baselines and experiments needed to support claims about efficiency and model performance. Address reproducibility, limitations and the distinction between a measured result and a broader claim.
Behavioural practice
Tell me about a time you collaborated across disciplines on a research problem.
What to coverSuggested preparation: Give the situation, your specific actions to reconcile different priorities or terminology, and the resulting research outcome. Keep your contribution distinct from the team’s collective work.
Describe a time you mentored a junior researcher.
What to coverSuggested preparation: Explain the junior researcher’s goal, how you adapted your guidance and what changed as a result. Use a real example that shows both support and room for independent work.
Tell me about a research direction you pursued independently.
What to coverSuggested preparation: Set out the problem, how you decided what to investigate and the actions you took when evidence challenged your initial approach. End with the result and what you learned.
Describe your contribution to a research proposal or strategic initiative.
What to coverSuggested preparation: Explain the situation, the specific research case or plan you helped develop, and the outcome. Distinguish your own contribution from any proposal or initiative led by others.
Tell me about a time you communicated a research finding for publication.
What to coverSuggested preparation: Describe the finding, the evidence you assembled and the decisions you made to present its contribution and limitations clearly. State the outcome without implying that every submission was accepted.