Towards realistic evaluation of cultural value alignment in large language models: Diversity enhancement for survey response simulation

Haijiang Liu, Yong Cao, Xun Wu, Chen Qiu, Jinguang Gu, Maofu Liu, Daniel Hershcovich

Abstract: Highlights•Our diversity-enhanced framework improves cultural value alignment and representativeness in large language models.•We design multi-aspect evaluation metrics for understanding cultural value misalignment in large language models.•Our study reveals divergence and preference biases across models when simulating survey responses in US and Chinese contexts.