Personalized Large Language Models

TL;DR

Universal LLMs excel at broad tasks but face limitations in aligning with individual user preferences for subjective text interpretation. The paper systematically compares personalized fine-tuning against zero-shot and few-shot reasoning, using user IDs as persistent context and evaluating on emotion recognition and hate speech datasets. It shows that personalized fine-tuning methods (CLS-P, LM-P) provide substantial performance gains across diverse models, with task and dataset characteristics moderating the magnitude of improvement. These findings support targeted personalization as a practical path to more user-aligned LLM outputs and include open-source resources to enable reproducibility and further research.

Abstract

Large language models (LLMs) have significantly advanced Natural Language Processing (NLP) tasks in recent years. However, their universal nature poses limitations in scenarios requiring personalized responses, such as recommendation systems and chatbots. This paper investigates methods to personalize LLMs, comparing fine-tuning and zero-shot reasoning approaches on subjective tasks. Results demonstrate that personalized fine-tuning improves model reasoning compared to non-personalized models. Experiments on datasets for emotion recognition and hate speech detection show consistent performance gains with personalized methods across different LLM architectures. These findings underscore the importance of personalization for enhancing LLM capabilities in subjective text perception tasks.

Figures (3)

• Figure 1: Non-personalized vs. personalized setups.
• Figure 2: Performance gains of personalized vs. non-personalized methods on the GoEmotions dataset.
• Figure 3: Performance gains of personalized vs. non-personalized methods on the Unhealthy Conversations.