Cross-Cultural Alignment of Llms: The Emergence and Mitigation of Risks
Cross-Cultural Alignment of Llms: The Emergence and Mitigation of Risks
Wednesday, 9 July 2025: 00:00
Location: FSE036 (Faculty of Education Sciences (FSE))
Oral Presentation
The development of large language models (LLMs) is predominantly anchored in regions with advanced technological infrastructure and widespread use of common languages. These models may exhibit reduced sensitivity to the nuances of diverse cultures, thereby presenting risks in cross-cultural alignment. Drawing on the principle of cultural relativism, this study delineates risks from three critical perspectives: (a) the simplification of model value cognition; b) misidentification of cross-cultural circumstances; and (c) scarcity in cross-cultural reasoning. By regarding cross-cultural alignment as the commensuration of cultural content and values across diverse linguistic and contexts, this study concludes three reasons from the gap between corpus and values. First, the capability of multilingual understanding is replaced with cross-cultural understanding. For the public content retained in the singular language on the Internet or specific corpus, there may be a problem of unbalanced representation. Second, the local authenticity is replaced with technical rationality. The rigid adaption for classic values theories such as Inglehart, Schwartz and Hofstede leads to contextual simplification of values and social norms in cross-cultural alignment. Third, the embodied value interpretations are replaced with institutional norms. Current LLMs tend to prioritize multilingual training parameters while neglecting the cultural connotations that emerge from the interactions among local actors in cross-cultural contexts. Textual data alone cannot encompass all forms of cultural representation. The absence of thick data that elucidates local cultures may hinder the model's capacity to comprehend cross-cultural situations effectively. To avoid a neo-technological colonialism under the spread of LLMs and reduce the dissemination of hegemonic technological paradigms across developing nations, this study proposes to bring indigenous multi-actors into the process of cross-cultural alignment.