MPG: Multi-Personality Generation of Large Language Models at Decoding-time

ACL ARR 2025 May Submission7520 Authors

Abstract: Multi-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a key challenge. Existing retraining methods are costly and unscalable, while decoding-time methods often rely on external models or heuristics, limiting flexibility and robustness. We propose MPG, a novel decoding-time framework addressing these issues. MPG formulates multi-personality generation as sampling from a weighted mixture distribution of individual preference models. It leverages the density ratio principle, where the target distribution's ratio relative to a reference model is proportional to a weighted sum of individual density ratios. And MPG employs rejection sampling for efficient generation. A core advantage of MPG is universality: a unified, probability-ratio-based framework capable of composing heterogeneous models from diverse sources, allowing simple personality addition without costly combined model retraining. Experiments on MBTI personality and role-playing demonstrate the effectiveness of MPG, showing improvements up to 16.36%–17.57%.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: personalization; alignment

Languages Studied: English

Submission Number: 7520