最大化用户上下文与响应之间的互信息以提升大型语言模型个性化，无需额外数据

A novel self-improvement framework, Mutual Information Preference Optimization (MIPO), is proposed to enhance the personalization of large language models (LLMs) without relying on external supervision or human-labeled data. MIPO utilizes a contrastive data augmentation method to construct preference pairs, generating a positive response based on the correct prompt and a negative response based on a random, unrelated prompt. This approach effectively leverages existing data, circumventing the high costs associated with collecting new high-quality data, while transcending the limitations of traditional verifiable tasks. Experimental results demonstrate that MIPO significantly improves the personalization performance of models across various domains, showcasing the potential for self-improvement. Through this method, LLMs can optimize response quality and user adaptability without additional data.