Researchers introduce PLURAL, a large-scale preference dataset designed to help large language models reflect diverse global values rather than predominantly Western ones. The dataset is grounded in the Integrated Values Survey, which spans 92 countries, and contains approximately 500,000 synthetic preference triplets representing people from 20 diverse nations.
- PLURAL transforms nationally representative survey responses into realistic scenarios that preserve normative value signals.
- Dataset-level validation confirms it preserves both cross-country differences and within-country diversity from the original survey.
- Automated evaluation shows training on PLURAL reduces mean absolute error in cultural profile alignment by up to 27.7% compared to strong baselines.
- Blind human evaluations with 176 participants in India, Brazil, and Japan judged PLURAL-aligned responses as more representative of their national values.
The authors consider this important because it offers a scalable resource for pluralistic alignment, enabling models to better represent diverse value systems globally.