Mixture-of-Personas Language Models for Population Simulation

Ngoc Bui, Hieu Trung Nguyen, Shantanu Kumar, Julian Theodore, Weikang Qiu, Viet Anh Nguyen, Rex Ying · 2025

Read Mixture-of-Personas Language Models for Population Simulation by Ngoc Bui, Hieu Trung Nguyen, Shantanu Kumar, Julian Theodore, Weikang Qiu, Viet Anh Nguyen, Rex Ying (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-05019.txt.

Description

Ngoc Bui and co-authors propose this 2025 prompting scheme for making language models imitate a whole population, not one average voice. Simulated people are used to augment social science data, but a stock model flattens human diversity. Their Mixture of Personas treats the population as a blend of subpopulations, each voiced by a persona paired with an example response.

For each simulated answer, the system draws a persona and example at random, weighted by learned mixing proportions. No retraining of the base model is needed, and the scheme ports across models. In synthetic data tests, the blend beat rival methods on both matching the target population and preserving its spread. The approach suits survey simulation and training data generation where diversity is the point.

Catalog record

CREATORNgoc Bui, Hieu Trung Nguyen, Shantanu Kumar, Julian Theodore, Weikang Qiu, Viet Anh Nguyen, Rex Ying
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-05019

Browse the full library · About the Tower of Babel