Rethinking LLM-based Preference Evaluation