The 60-second answer
Start with the behavior/data gap and compare prompting, RAG, SFT/PEFT, preference optimization, and full fine-tuning before choosing the most expensive option. For SFT/LoRA, define the trainable parameters, data format, objective, and what remains frozen; for preference methods, define chosen/rejected or reward signals.
Build the answer in this order
Start with the behavior/data gap and compare prompting, RAG, SFT/PEFT, preference optimization, and full fine-tuning before choosing the most expensive option.
For SFT/LoRA, define the trainable parameters, data format, objective, and what remains frozen; for preference methods, define chosen/rejected or reward signals.
Evaluate capability, factuality, safety, regression slices, latency, and cost on held-out prompts rather than training loss alone.
Version adapters/checkpoints and maintain a rollback path because alignment/fine-tuning can regress unrelated capabilities.
A useful interview mental model
This is the shape of a strong answer—not a script to memorize.
Senior-level signal
- Discuss data contamination, preference bias/reward hacking, and catastrophic forgetting as separate risks.
- Choose PEFT rank/target modules or preference constraints based on measured quality-vs-memory/throughput trade-offs.
What the interviewer is really testing
Likely follow-up questions
Common weak-answer patterns
- Reciting a definition without mechanism or assumptions.
- Claiming one technique is always better without a data regime.
- Stopping before failure modes, validation, or deployment implications.