{"id":"arxiv/2607.02368","url":"https://arxiv.org/abs/2607.02368","tags":["stat.ML","cs.AI","cs.LG","math.DG"],"score":0,"title":"Beyond Cultural Persona Bias: Response Geometry as a Measure for Language Models and People","domain":"stat.ML","source":"arxiv","authors":["Yuan Yuan"],"reviews":0,"subject":"stat.ML","summary":"Researchers are increasingly prompting language models with cultural personas so that they answer questionnaires as diverse human groups would. However, studies often report that these persona responses do not predict the models' behavior. We argue that whether the persona is Western or non-Western, the model's answers carry a cultural preference, and that this preference masks the variables that could predict its behavior. We examine the prompted preference at two levels: aggregate Big Five scores and response geometry, the correlation matrix between dimensions within each respondent,…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-07-02","reactions":{}}