Question 760
A preference-tuned assistant performs well on prompts similar to the preference dataset but produces unhelpfully terse answers to new customers. What should evaluation expand?
Explanation
Correct Answer
Preference optimization can overfit a narrow reward signal; broad held-out tests detect degradation in real-world helpfulness.
Incorrect Answers
Job age does not establish generalization.
Serving capacity is not language behavior quality.
Training preference accuracy does not measure unseen task completion.