Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Styles + Persona-plug = Customized LLMs

Yutong Song, Jiang Wu, Shaofan Yuan, Chengze Shen, Jian Wang, Amir Rahmani, Nikil Dutt, Yu Wang

TL;DR

The paper reframes user personalization in LLM generation as a residual between a user’s true linguistic behavior and a neutral base model, enabling style-aware control without per-user fine-tuning. It introduces PsPLUG, a lightweight soft-prompt plug-in that injects a trainable prefix into a frozen backbone, learned via style-conditioned preference contrasts to separate persona from style. A unified style-persona balancing mechanism with an inference-time alpha control allows continuous trade-offs between personalization strength and adherence to explicit style instructions. Experiments on LaMP demonstrate improved persona alignment and robust style fidelity with minimal computational overhead, supporting scalable, style-aware personalization for LLMs.

Abstract

We discover a previously overlooked challenge in personalized text generation: personalization methods are increasingly applied under explicit style instructions, yet their behavior under such constraints remains poorly understood. To balance implicit personalization and explicit style, we formulate personalization as a distributional residual and propose PsPLUG, a lightweight soft-prompt plug-in trained with style-conditioned preference contrasts. Across LaMP benchmark, our framework improves persona alignment, maintains stylistic fidelity, and outperforms retrieval-based and soft-prompt baselines with minimal computation. These results show that residual modeling provides a simple and principled foundation for controllable, style-aware LLM personalization.

Styles + Persona-plug = Customized LLMs

TL;DR

The paper reframes user personalization in LLM generation as a residual between a user’s true linguistic behavior and a neutral base model, enabling style-aware control without per-user fine-tuning. It introduces PsPLUG, a lightweight soft-prompt plug-in that injects a trainable prefix into a frozen backbone, learned via style-conditioned preference contrasts to separate persona from style. A unified style-persona balancing mechanism with an inference-time alpha control allows continuous trade-offs between personalization strength and adherence to explicit style instructions. Experiments on LaMP demonstrate improved persona alignment and robust style fidelity with minimal computational overhead, supporting scalable, style-aware personalization for LLMs.

Abstract

We discover a previously overlooked challenge in personalized text generation: personalization methods are increasingly applied under explicit style instructions, yet their behavior under such constraints remains poorly understood. To balance implicit personalization and explicit style, we formulate personalization as a distributional residual and propose PsPLUG, a lightweight soft-prompt plug-in trained with style-conditioned preference contrasts. Across LaMP benchmark, our framework improves persona alignment, maintains stylistic fidelity, and outperforms retrieval-based and soft-prompt baselines with minimal computation. These results show that residual modeling provides a simple and principled foundation for controllable, style-aware LLM personalization.
Paper Structure (57 sections, 13 equations, 4 figures, 7 tables)

This paper contains 57 sections, 13 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Task Formulation
  4. PsPLUG Framework
  5. System Instruction Embedding.
  6. Personalization Embedding Projector.
  7. User Query Encoder.
  8. PsPLUG Training: Persona--Style Balancing
  9. Construction of Style-Conditioned Pairs.
  10. Residual Based Training Objective.
  11. Inference Time Persona Strength Control.
  12. Special Case: Personalization without Style Instructions
  13. Neutral Baseline and Preference Pairs.
  14. Residual Objective without Style.
  15. Experimental Settings
  16. ...and 42 more sections

Figures (4)

  • Figure 1: PsPLUG is a lightweight plug-in framework that injects a learnable user-specific prefix into a frozen LLM to enable controllable personalization. The fire logo indicates the trainable model and the snow logo indicates the frozen model.
  • Figure 2: Personalization performance conducted at different strengths and styles.
  • Figure 3: A case study for personalized-style outputs for LaMP 7. The yellow background highlights content overlapping with the user history.
  • Figure 4: LLM-based judgments and human evaluations for LaMP 7.