Training Trains logo

LLM Engineering Internship Interview Questions for Freshers (with Answers)

By K. Boopathikumar, CEO & Founder · 16 November 2026 · 8 min read

LLM engineering interns work on adapting, evaluating and serving large language models. Interviews check whether you understand how transformers work at a high level and how to choose between prompting, retrieval and fine-tuning.

How LLMs work

  • What is a transformer? A neural network architecture built on self-attention, which lets every token weigh every other token in the input. It replaced recurrent networks because it trains in parallel and handles long context better.
  • What does attention do? For each token it computes how relevant every other token is (using queries, keys and values) and mixes their information accordingly.
  • What is tokenisation and why does it matter? Splitting text into sub-word pieces the model uses. It affects cost, context length and how well a model handles languages such as Tamil, which may need more tokens per word.
  • Pre-training vs instruction tuning? Pre-training teaches next-token prediction on huge text; instruction tuning and preference training teach the model to follow instructions and give helpful, safe answers.

Adapting a model

  • When do you fine-tune instead of using RAG? Use RAG when the model needs facts that change or are private. Fine-tune when you need a consistent style, format or a narrow skill that prompting cannot reach reliably.
  • What is LoRA? Low-rank adaptation: training small extra matrices instead of all the model's weights, which makes fine-tuning cheaper and lets you swap adapters.
  • What data do you need for fine-tuning? Clean, representative input-output examples, checked for errors, duplicates and personal data, with a held-out set for evaluation.
  • What is catastrophic forgetting? When fine-tuning on a narrow task makes the model worse at things it could do before. Mixing in general data and using small learning rates helps.

Evaluation and serving

  • How do you evaluate an LLM for a task? A fixed test set with expected answers or a rubric, automatic checks where possible, human review for quality, and comparing versions on the same set.
  • What is quantisation? Storing weights in fewer bits (for example 8 or 4) to cut memory and speed up inference, with a small loss in quality that you must measure.
  • What affects latency when serving an LLM? Model size, prompt length, output length, batching, hardware and caching. Streaming tokens improves how fast it feels.
  • What is a KV cache? Stored attention keys and values from earlier tokens, so generating each new token does not recompute the whole prompt.
  • How do you keep a deployed model safe? Input and output filters, rate limits, logging, red-team testing within policy and a clear way for users to report problems.

How to answer well in the interview

  • Explain a trade-off you made between cost, quality and speed.
  • Show an evaluation table comparing at least two approaches.
  • Be honest about what you ran yourself versus read about.
  • Know the licence of any open model you used and what it allows.

Next steps

Looking for an internship near Erode?

10 days to 6 months, online or at our Erode and Perundurai branches, with a verifiable certificate.

See internships →

More guides