Large language models (LLMs) are increasingly deployed in culturally sensitive contexts, where alignment must capture the diverse preferences that exist within populations. Existing approaches typically model at coarse demographic or community levels and ignore intra‑group variation.
We introduce Demographic Pluralism, an inference‑time framework that estimates population‑level opinion distributions without requiring opinion‑distribution training data or task‑specific fine‑tuning. The key idea is to generate multiple perspectives grounded in demographic groups and aggregate them during inference.
Experiments on the GlobalOpinionQA and VITAL benchmarks using four backbone models show that Demographic Pluralism reduces the Jensen‑Shannon distance by 8.4%–26.4% compared to Modular Pluralism. The Jensen‑Shannon distance is defined as $JS(P\|Q)=\frac12 KL(P\|M)+\frac12 KL(Q\|M)$ with $M=\frac12(P+Q)$.
We evaluate weighted, equal‑weighted, and inverse‑weighted aggregation strategies; equal weighting yields the best overall performance. Further analysis reveals that group‑level error increases with group weight, which helps explain the weaker performance of weighted aggregation.
Review: Demographic Pluralism captures intra‑group diversity by introducing demographic perspectives at inference time, offering a practical route to improve alignment without extra annotation or fine‑tuning.