R
Rosie
Text
Retrieval-Augmented Cultural Values Alignment for Large Language Models
A framework (ValuesRAG) that improves the cross-cultural alignment of large-language-model outputs by retrieving demographic-conditioned summaries of survey-measured human values and injecting them into the prompt through in-context learning. Per-individual value summaries are built from the World Values Survey, retrieved by demographic features and reranked to a top-k set, after which a reasoning model answers grounded in those values. It is presented as an alternative to role-assignment and few-shot prompting for reducing Western-centric bias.