Beyond Mimicry: Preference Coherence in LLMs