Generate evaluation criteria for any question using LLMs (scenarios, perspectives, criteria).
More expansion rounds → more comprehensive criteria, but longer processing time per question.