Learning LLM Preference over Intra-Dialogue Pairs: A Framework for Utterance-level Understandings