Explain the difference between Encoder-only, Decoder-only, and Encoder-Decoder architectures. Also explain in what research scenario would you choose an Encoder-only model over a Decoder-only model for a classification task, and why?✱

Synthetic Data Generation: Describe a time you generated synthetic data for a project. How did you ensure diversity (preventing the model from repeating patterns)? How did you validate the quality of the generated data without manually checking every row? If you have no experience with this, please specify.✱

Agentic System Architecture: You are building an agent that must call set of external API. How do you ensure the LLM outputs valid JSON for the API? How do you handle "State"? (e.g., The user provides the Order ID in message 1, but confirms the cancellation in message 3).✱

Latency vs. Quality in Real-Time Systems: We need a chatbot that processes voice transcripts in real-time. The goal is low latency (Time-To-First-Token) without sacrificing accuracy. How would you architect this? How does your design change if the input text is noisy (ASR errors like "uhm", "ah", or misspelled entities)?✱

Evaluation & Benchmarking: Have you trained an LLM for a specific task (e.g., extracting ticket details). How did you go about training the model? What was the data? How did you evaluate this model beyond just loss curves? Were standard benchmarks (like MMLU) relevant there? And if not, what specific metrics would you design for this task? If you have no experience with this, please specify.✱

Multilingual Adaptation: An LLM is trained on English and French. You need to adapt it to Japanese. What specific challenges arise regarding tokenization for Japanese compared to European languages? How would you efficiently adapt the model without full re-training?✱