LLM-Based Offline Learning for Embodied Agents via Consistency-Guided Reward Ensemble F i n d i n g s - E M N L P