Mixture-of-Personas Language Models for Population Simulation

TL;DR

MoP introduces a two-level hierarchical mixture of LLM agents, combining learnable persona prompts and exemplar-driven conditioning to align generated content with target population distributions without fine-tuning. The method formalizes a gating-based mixture over personas and exemplars, with objective optimization on population data and a plug-and-play transfer capability across base models. Empirical results show substantial gains in alignment (FID) and distributional diversity (MAUVE, KL Cosine), along with improved downstream task performance and strong transferability. The work offers a practical, privacy-conscious approach to population-scale simulations with LLMs and highlights considerations around biases and model-access constraints that guide future research and deployment.

Abstract

Advances in Large Language Models (LLMs) paved the way for their emerging applications in various domains, such as human behavior simulations, where LLMs could augment human-generated data in social science research and machine learning model training. However, pretrained LLMs often fail to capture the behavioral diversity of target populations due to the inherent variability across individuals and groups. To address this, we propose \textit{Mixture of Personas} (MoP), a \textit{probabilistic} prompting method that aligns the LLM responses with the target population. MoP is a contextual mixture model, where each component is an LM agent characterized by a persona and an exemplar representing subpopulation behaviors. The persona and exemplar are randomly chosen according to the learned mixing weights to elicit diverse LLM responses during simulation. MoP is flexible, requires no model finetuning, and is transferable across base models. Experiments for synthetic data generation show that MoP outperforms competing methods in alignment and diversity metrics.

Figures (14)

• Figure 1: Sampling from foundational LLM agents frequently yields repetitive and generic responses. Meanwhile, prompting with personas can create more tailored, specific responses. The highlighted words in the figure correspond to the prompted personas.
• Figure 2: The generation pipeline for the Exemplar-based Mixture of Personas (MoP) operates as follows: Given a movie review $x$, MoP first samples a persona based on the learnable mixing weight $\pi$. Next, MoP selects an exemplar randomly from the observation pool according to the mixing weight $\Omega$. The selected persona and exemplar are then concatenated with the input context to create a personalized prompt used to sample from a base LLM agent. The dashed block indicates the process of persona synthesis.
• Figure 3: Diversity comparisons on Agnews of different prompting methods. The embeddings are computed using 'all-mpnet-base-v2'. The scatted points are synthesized samples with the colors indicating the corresponding labels. The circle lines indicate 2-std confidence ellipses of the golden test set. It can be seen that MoP offers synthesized samples that are diverse and aligned with the golden test set.
• Figure 4: Mauve scores with varying the number of personas and number of examples.
• Figure 5: MoP Prompt for Agnews