20 LLM Interview Questions to Test Your Readiness
If you can’t answer half, you’re not ready 👇
LLM Fundamentals
1. Explain tokenization beyond “splitting text”.
2. Why do decoder-only models dominate? When would you prefer encoder-decoder?
3. Walk me through attention. Where do queries, keys, values actually get used
4. Sampling strategies: top-K vs top-P vs temperature. When do you pick which?
Prompting & Context Engineering
5. Give me an example of a zero-shot prompt failing. How did you fix it?
6. How do you version and test prompts for stability?
7. Explain “context window waste.” How do you mitigate it?
Fine-Tuning & Alignment
8. LoRA vs QLoRA vs PEFT - where do you trade off memory, speed, accuracy?
9. What are the limits of RLHF? Give one real failure mode.
10. When would you choose open-source over proprietary LLMs for fine-tuning?
RAG (Retrieval-Augmented Generation)
11. Walk through how embeddings + similarity search actually work.
12. Chroma vs Pinecone vs Weaviate - what’s your POV?
13. When does hybrid retrieval outperform vanilla vector search?
Agentic AI
14. Define an agent without buzzwords.
15. How do you handle tool-use errors in an agent loop?
16. What’s the hardest part of multi-agent orchestration?
MLOps / LLMOps
17. How would you monitor hallucinations in production?
18. What metrics do you track in an eval pipeline?
19. How do you deploy a GenAI service (FastAPI + Docker + K8s) with rollback safety?
Scaling & Risk
20. Your system just blew the budget because of model size. Walk me through trade-off options in 3 minutes.
👉 Self-test: Can you answer these out loud, clearly, without hand-waving? If not, that’s your prep gap.