SourceDo LLMs have Consistent Values?
The paper administered the 57-item PVQ-RR questionnaire to six LLMs under five prompting strategies and two temperatures, then computed Spearman rank correlations between each model's value ranking and the human benchmark from Schwartz and Cieciuch (2022, 49 cultural groups, N=53,472). Most prompt-model combinations yielded correlations above 0.8 with p < .001. The value anchor prompt produced correlations ranging from 0.75 to 0.85. A notable exception was GPT-4-0314 under the basic prompt, which showed very low correlation and stark deviations (conformity-rules ranked first vs. humans' 16th; benevolence-care ranked 18th vs. humans' 1st).