Title: Moving Past Discrete Actions: A Hierarchical Framework for Tailoring LLMs to Users

Abstract:

Despite the impressive performance of Large Language Models (LLMs) in various fields, customizing their responses to fit individual user preferences remains a significant hurdle. Current methods largely rely on a flat behavioral approach, which aggregates user actions without considering how these actions form deeper, structured patterns. To address this, we introduce PHF (Practice-Habitus-Field), a framework rooted in Pierre Bourdieu’s Theory of Practice. This approach redefines LLM personalization through a three-tiered hierarchy: individual actions are viewed as "practices," their long-term accumulation forms stable dispositions known as "habitus," and commonalities among similar user groups are identified as "fields." We demonstrate the practical application of PHF via $\mathrm{PHF}_{\text{Compass}}$, a lightweight, model-agnostic tool built upon a frozen LLM. Testing on the Language Model Personalization (LaMP) benchmark shows consistent performance gains across a variety of tasks. Additional analyses confirm that the behavioral structures learned by this method are both interpretable and extensible.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC