Aligning AI with cultural values requires not only model tuning but also a target grounded in real‑human evidence. Three fundamental questions remain: whose values should be captured, where those values should be observed, and which signals reveal them.
Re3Culture addresses these questions with a large‑scale dataset. Over 700 participants from four cultures were recruited using cultural‑affiliation screening and demographically aware sampling to ensure representativeness. Scenarios were selected by community members and focus on culturally sensitive domains such as family and kinship. Value signals comprise more than 1,400 surveys, over 20,000 topic‑grounded behavioral principles, and explicit value‑priority responses for more than 6,000 situated questions.
Our analysis shows that estimating cultural values demands both community affiliation and broad internal coverage. Substituting cultures or limiting coverage to sub‑groups shifts value estimates, and LLM simulations disproportionately reflect particular sub‑groups. Values must be observed in concrete, community‑guided situations; abstract questionnaire profiles weakly recover situated priorities, and communities differ on which scenarios they deem value‑revealing.
Value priorities should be elicited explicitly rather than inferred from preference or final answers alone. Preferences conflate value agreement with other response qualities, and similar answers can arise from different underlying value hierarchies. By meeting these requirements, Re3Culture becomes the first human‑grounded resource that clearly informs target definition for cultural value alignment.
Review: Re3Culture offers a systematic, verifiable framework for cross‑cultural value alignment, highlighting the importance of community involvement and concrete contexts, and sets a solid foundation for future model evaluation and fine‑tuning.