Six LLM debugging scenarios for 2026 interviews and how to structure answers

Applied LLM interviews increasingly open with a production symptom, such as truncated answers, a sudden cost spike, or overnight behavior changes, rather than conceptual questions. The recommended answer structure is to list likely causes, explain how to distinguish between them, and then state what to change. Among the scenarios, the guide covers the "Lost in the Middle" effect, where information placed mid-context is recalled poorly, and advises checking stop reasons such as length or max_tokens before guessing. It also recommends bucketing malformed-output failures before applying fixes, and breaking cost down by requests, tokens, and price to find the driver. Finally, it notes that prompt caching requires an exact prefix match and a live cache within its lifetime, and that unpinned model aliases can change behavior silently.