Evals And Observability medium
You compare your production model against a candidate model from a different provider using pairwise LLM judging. The judge is from the same model family as production, and the candidate loses 38% to 62%. What is the main concern with this result?
- A Position bias alone; randomizing which answer appears first fully removes any systematic advantage held by the production model's outputs.
- B Self-preference: judges tend to favor outputs resembling their own, so use a neutral judge or check the result against human pairwise labels.
- C No real concern if the judge runs at temperature zero, since a deterministic judge gives reproducible verdicts that cannot favor either side.
- D Length bias alone; truncating both answers to the same token count before judging is enough to make the comparison fair and unbiased.
Answer: B. Self-preference: judges tend to favor outputs resembling their own, so use a neutral judge or check the result against human pairwise labels.
**Self-preference bias** means a judge tends to rate outputs from its own model family higher, so using production's model family as the judge tilts the comparison. A neutral judge or a human-labeled subset costs more but removes the conflict of interest. Randomizing order (A) is necessary for pairwise judging, but it addresses position bias, not self-preference. Running the judge deterministically (C) makes a bias reproducible, not absent.
One system design question a day, with spaced repetition so the ones you miss come back.
Get the appRelated questions
- In one release, a RAG assistant got a new embedding model and a new chunking strategy, and the end-to-end judge score dropped 6 points. The team only tracks the end-to-end score and has spent a week tweaking the generation prompt with no recovery. What is the best next step?
- Over two months, your team iterated through 40 prompt versions, each tuned against the same 150-case golden set. The offline score rose from 72 to 91. Online task completion and retry rate have not changed. What is the most likely problem, and what is the fix?
- A prompt change produced an impressive answer in a live demo for leadership. The PM wants to ship it today. What should you do?
- You plan to replace human graders with an LLM judge for your answer-quality eval. Before trusting its scores in a CI gate, what should you do first?
- After you swap to a smaller model, p95 latency falls from 900 ms to 350 ms and the ops dashboard is all green. Leadership asks whether answer quality is still fine. What is the right response?
- Your CI gate fails any prompt change that drops the score 3 or more points on a 100-point rubric over 60 cases. Production generates at temperature 0.7. Running the eval twice on the same commit gives scores that differ by up to 4 points, and engineers have started rerunning failed gates until they pass. How do you fix the gate?