A newly published academic study focuses on the issue of value-selection bias in large language models (LLMs) when presented with multiple cultural identity cues. The study was posted on the arXiv preprint server on October 24, 2026 (ID 2610.31199v1), employing a dual-identity framework: the system prompt predefines a national “persona,” while the user input introduces a second cultural cue—either by explicitly naming a country or implicitly through the query language.
The study combines the World Values Survey (WVS) multiple-choice questions with an open-ended socio-political generation task for evaluation. The results show that, in the multiple-choice task, linguistic cues induce answer shifts roughly three times as strong as those triggered by explicit country naming; however, the direction of these shifts diverges from the actual distribution of values associated with the countries linked to those languages. By contrast, explicit country naming does not consistently steer the model’s outputs toward the empirical distributions of the corresponding countries.
In an analysis covering 1.6 million open-ended generated responses, explicitly naming a specific country can reduce the gap between the model’s baseline response and that country’s domain-specific response by 30–44%; moreover, compared to linguistic cues, explicit cues elicit more pronounced and distinct shifts in values. Based on these findings, the study underscores that national identity and language should not be treated as interchangeable proxies for culture, but rather modeled and evaluated as independent signals.