HyPerAlign: Interpretable Personalized LLM Alignment via Hypothesis Generation

We introduce HyPerAlign, a hypothesis-driven personalization approach for large language models (LLMs) that generates user-specific outputs rather than generic responses aligned to average-user preferences. Given few-shot examples from a user, the method first infers hypotheses about their communication strategies, personality, and writing style, then uses these hypotheses to generate customized outputs. The approach is evaluated on two tasks: authorship attribution and deliberative alignment, using datasets from diverse domains (news articles, blog posts, emails, jailbreaking benchmarks). Results show superiority over preference-based fine-tuning methods, with win-rates exceeding 90% for authorship attribution and up to 70% improvement in helpfulness for deliberative alignment. The inferred hypotheses generalize across models and out-of-distribution datasets, providing an interpretable and sample-efficient strategy for personalizing LLMs to individual users.