Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalization

TL;DR

Proposes DPS, a method that identifies 'Preference Heads'—attention heads responsible for user preferences in LLMs—and leverages them to control personalization at inference time without training.

Problem

Existing LLM personalization methods rely on prompt engineering or fine-tuning, operating as black boxes with limited understanding of how personalization occurs internally. They also suffer from high computational cost or low interpretability.

Approach

The authors hypothesize the existence of a sparse set of attention heads (Preference Heads) that encode user-specific stylistic and topical preferences. DPS (1) identifies Preference Heads via causal masking analysis and (2) amplifies the difference between predictions with and without these heads during decoding to strengthen preference-aligned outputs. A Preference Contribution Score (PCS) measures each head's causal impact on personalization.

Results & Contribution

Across multiple LLMs and personalization benchmarks, DPS consistently improves personalization fidelity while preserving content coherence and low computational overhead. It also provides a mechanistic explanation of where and how personalization emerges within transformer architectures, enhancing interpretability.