Demographic Pluralism Models Diverse Human Preferences at Inference Time

Summary

Large language models are increasingly used in culturally sensitive settings, where alignment may require representing distributions of preferences rather than a single population average. The paper introduces Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions without opinion-distribution training data or task-specific fine-tuning. It generates multiple perspectives within demographically grounded groups, explicitly accounting for variation among people in the same group. Across four model backbones evaluated on GlobalOpinionQA and VITAL, the method reduces Jensen-Shannon distance by 8.4% to 26.4% compared with Modular Pluralism. The authors compare weighted, equal-weighted, and inverse-weighted aggregation and find that equal weighting performs best overall. They also report that group-level error rises as a group receives more weight, offering an explanation for the weaker results of weighted aggregation.