LLM Evals · Turn quality into a signal
When do evals give false confidence?
Short answer
When the dataset is unrepresentative, labels are noisy, the metric is gamed, data leaked, or the judge is miscalibrated. A high score is only as trustworthy as those assumptions.
Evals are measurements, and measurements can be wrong. A suite can look comprehensive while covering only easy cases, or reward verbosity because the judge prefers long answers. The most dangerous outcome is a high score that hides a real production failure.
- Unrepresentative data: real users ask things the set never includes.
- Label noise: ambiguous criteria make the score unstable.
- Metric gaming: optimising length, format or judge preferences rather than value.
- Leakage: eval cases appear in prompts, training or retrieval.
- Miscalibrated judge: consistent but systematically wrong.
- Stale baseline: the product or distribution changed since the set was built.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you detect that a team is overfitting to evals?
- 2Why is a private holdout useful even with a large golden set?
- 3What would make you distrust a sudden score jump?
Say it in the interview
Evals lie when the data is unrepresentative, the labels are noisy, the metric rewards the wrong thing or the judge is miscalibrated. A sudden score jump is a warning sign, not automatically good news. I triangulate with a private holdout, human review and online behavior. If all three agree, I trust the result much more than a single eval number, however comprehensive the suite looks.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live LLM Evals mock interview and get scored feedback on depth, clarity and confidence.