Evals And Observability easy
You use an LLM judge to compare your current model against a candidate model. The judge is the same model as the candidate, and the candidate wins 75% of comparisons. What is the main concern?
- A Nothing significant; a capable model grades its own outputs as fairly as it grades outputs from other models.
- B The 75% win rate is too low to be meaningful, since a judge needs 95% agreement before it can be trusted.
- C LLM judges tend to prefer outputs that resemble their own, so the win rate may be inflated by self-preference.
- D Pairwise comparison is unreliable for model selection, so the judge should switch to assigning absolute 1 to 10 scores.
Answer: C. LLM judges tend to prefer outputs that resemble their own, so the win rate may be inflated by self-preference.
Self-preference bias means a judge tends to favor text in its own style, so using the candidate as judge is cheap but can inflate its win rate. Assuming a strong model grades fairly is the tempting mistake. Use a judge from a different model family, or verify on a human-labeled calibration set; a same-family judge is acceptable only if that calibration shows no skew.
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?