REAL: Response Embedding-based Alignment for LLMs