Mixture-of-Personas Language Models for Population Simulation

Introduction

Large Language Models (LLMs) have shown impressive capabilities in generating human-like text and are increasingly used to simulate human behavior in social science, economics, and machine learning. But pretrained LLMs often produce generic and repetitive responses, failing to reflect the diversity of real-world populations shaped by age, culture, or demographics.

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.

Typical methods use “persona prompting” to steer LLM behavior—crafting prompts to mimic specific groups. However, even with persona prompts, responses still lack diversity and exhibit bias. Fine-tuning models with personal data could help, but it's costly, raises privacy concerns, and limits flexibility.

Our Solution: Mixture of Personas (MoP)

We introduce Mixture of Personas (MoP) — a probabilistic prompting framework that aligns LLM responses with the target population while preserving diversity. Instead of using one static persona, MoP treats simulation as a contextual mixture model, where each component is a language model agent defined by:

• A persona (a group-level identity or behavior pattern)
• An exemplar (an in-context example from the population)

During generation, both persona and exemplar are sampled probabilistically based on learned weights, allowing LLMs to produce more tailored and varied responses.

An example of MoP for simulating movie reviews.

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.

Experiments

To show the effectiveness of MoP, we use MoP to simulate movie and restaurant reviews from Yelp, SST2, IMDB and Agnews datasets. Table below shows that MoP can significantly outperform other baseline prompting methods. Specifically, our method can outperform the best baseline by 58.8% in terms of FID and by 27.9% in terms of MAUVE, averaged over all datasets. This result demonstrates that MoP can effectively steer LLM's outputs to the target responses of the target population. Besides the improvement in alignment scores, KL Cosine metrics also indicate that our responses are more diverse than other baselines.

Alignment and diversity of the generated datasets using our MoP with respect to the golden test set.

We train a DistilBERT model using 5,000 synthesized samples and evaluate its performance on the golden dataset, reporting the F1-score. Results show that the synthetic dataset generated by Llama3 with MoP prompting achieves up to a 2.68% improvement over the AttrPrompt baseline.

Performance on downstream classification task using generated datasets to train a sentiment classifier (Yelp, SST-2, and IMDB) or news topic classifier (AgNews). F1 scores are calculated on the golden test set.

To further investigate the advantage of MoP prompting, we plot the embedding space. The figure shows that our method matches the diversity of the golden datasets better than other baselines. Meanwhile, simulated responses of other baselines often collapse or cannot cover the diversity of the true responses.

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.