Beyond Isolated Behaviors: Hierarchical User Modeling for LLM Personalization
Summary
A new framework, PHF (Practice-Habitus-Field), is introduced for personalizing Large Language Models (LLMs) by modeling user behavior hierarchically. Developed from Pierre Bourdieu's Theory of Practice, PHF organizes user interactions into three levels: individual behaviors as practices, their accumulation into stable dispositions as habitus, and shared patterns among similar users as fields. This approach moves beyond flat behavioral paradigms that aggregate user data without deeper structural understanding. The framework is instantiated as PHF_Compass, a lightweight and model-agnostic implementation built upon a frozen LLM. Experiments conducted on the Language Model Personalization (LaMP) benchmark demonstrate consistent performance improvements across various tasks, while further analysis confirms the interpretability and extensibility of the learned behavioral structures.
Key takeaway
For Machine Learning Engineers developing personalized LLM applications, consider adopting a hierarchical user modeling approach. Your current flat behavioral paradigms may limit personalization depth and effectiveness. Implementing frameworks like PHF, which models user practices, habitus, and fields, can yield consistent performance improvements. This method, exemplified by PHF_Compass, allows for efficient, model-agnostic personalization using frozen LLMs, enhancing both output relevance and interpretability of user structures.
Key insights
LLM personalization benefits from a hierarchical user model based on practices, habitus, and fields.
Principles
- User behavior is hierarchical, not flat.
- Stable dispositions (habitus) emerge from practices.
- Shared regularities define user groups (fields).
Method
PHF is instantiated as PHF_Compass, a lightweight, model-agnostic system using a frozen LLM to learn hierarchical behavioral structures.
In practice
- Apply PHF to improve LLM personalization tasks.
- Utilize frozen LLMs for efficient personalization.
- Interpret learned user structures for insights.
Topics
- LLM Personalization
- Hierarchical User Modeling
- Practice-Habitus-Field
- Frozen LLMs
- LaMP Benchmark
- Sociological Theory
Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.