RongxinChen/dpo_personality

Multi-Personality Generation of LLMs at Decoding-time

Paper | Code

This repository contains datasets used in the paper "Multi-Personality Generation of LLMs at Decoding-time".

Introduction

Multi-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a fundamental challenge. The proposed Multi-Personality Generation (MPG) framework enables Large Language Models to simultaneously embody multiple personalization attributes at decoding-time. Unlike retraining-based approaches, MPG leverages implicit density ratios in single-dimensional models to control multi-personality without relying on scarce multi-dimensional models or extra training.

Datasets

The released data includes several DPO (Direct Preference Optimization) datasets used for training and evaluation:

MBTI Datasets

These datasets focus on specific Myers-Briggs Type Indicator dimensions:

• E / I: Extraversion vs. Introversion
• J / P: Judging vs. Perceiving
• T / F: Thinking vs. Feeling
• S / N: Sensing vs. Intuition

RolePlay Datasets

Datasets designed for enhancing role-playing capabilities:

• Personality: DPO data for general personality alignment.
• Profile: DPO data based on character profiles.

Citation

If you find our work or datasets useful, please consider citing our paper:

@inproceedings{chen2026multipersonality,
  title={Multi-Personality Generation of LLMs at Decoding-time},
  author={Chen, Rongxin and Li, Yunfan and Yuan, Yige and Xu, Bingbing and Shen, Huawei},
  booktitle={Proceedings of the 19th ACM International Conference on Web Search and Data Mining (WSDM '26)},
  year={2026},
  publisher={ACM},
  url={http://arxiv.org/abs/2511.01891}
}