Multi-Personality Generation of LLMs at Decoding-time

Multi-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a fundamental challenge. Existing retraining-based approaches are costly and poorly scalable, while decoding-time methods often rely on external models or heuristics, limiting flexibility and robustness.

In this paper, we propose a novel Multi-Personality Generation (MPG) framework under the decoding-time combination paradigm. It flexibly controls multi-personality without relying on scarce multi-dimensional models or extra training, leveraging implicit density ratios in single-dimensional models as a "free lunch" to reformulate the task as sampling from a target strategy aggregating these ratios. To implement MPG efficiently, we design Speculative Chunk-level based Rejection sampling (SCR), which generates responses in chunks and parallelly validates them via estimated thresholds within a sliding window. This significantly reduces computational overhead while maintaining high-quality generation. Experiments on MBTI personality and Role-Playing demonstrate the effectiveness of MPG, showing improvements up to 16%-18%.

• Promising Framework (MPG): A novel decoding-time paradigm that leverages implicit density ratios from single-dimensional models as a "free lunch", enabling flexible multi-personality control without extra training.
• Practical Method (SCR): Introduces Speculative Chunk-level Rejection sampling with parallel multi-preference scoring, which substantially reduces the computational overhead of large-model evaluations during decoding.
• Validated Effectiveness: Extensive experiments on MBTI simulation and Role-Play scenarios demonstrate significant performance gains, achieving 16%–18% improvement over baselines.

We release the MBTI and RolePlay DPO (Direct Preference Optimization) datasets used in our study.

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

Dimension	Description	Dataset Link
E / I	Extraversion vs. Introversion
J / P	Judging vs. Perceiving
T / F	Thinking vs. Feeling
S / N	Sensing vs. Intuition

Datasets designed for enhancing role-playing capabilities.

Type	Description	Dataset Link
Personality	DPO data for general personality alignment
Profile	DPO data based on character profiles

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}
}