Evaluating Regional Bias in LLMs From Abstract Stereotype to Concrete Social Decision-Making
Six tested LLMs showed consistent regional preferences across China, including in simulated education, job, and social choices.
The paper introduces S2D, a framework covering all 34 provincial-level regions of China and measuring both stereotype ratings and paired decisions. Models rated regions on warmth and competence, then made choices in education, occupation, and social interaction tasks. The authors report substantial regional differences, with especially strong agreement across models for competence and occupation outcomes. The patterns were linked to economic and digital development indicators and stayed largely stable across Chinese and English prompts. ArXiv · AI/CL/LG's note
The paper introduces S2D, a framework covering all 34 provincial-level regions of China and measuring both stereotype ratings and paired decisions. Models rated regions on warmth and competence, then made choices in education, occupation, and social interaction tasks. The authors report substantial regional differences, with especially strong agreement across models for competence and occupation outcomes. The patterns were linked to economic and digital development indicators and stayed largely stable across Chinese and English prompts. ArXiv · AI/CL/LG's note
score 4