MIPO Improves LLM Personalization And Performance

Hyunji Alex Nam et al. (arXiv, Mar 10, 2026) propose Mutual Information Preference Optimization (MIPO), a contrastive post-training method that generates positive responses conditioned on correct prompts and negatives from unrelated prompts. Training with Direct Preference Optimization (DPO) maximizes pointwise conditional mutual information and yields 3–40% personalization improvements and 1–18% gains on math and multiple-choice tasks without human supervision.