Cross-Cultural Reasoning in Large Language Models
Hybrid Seminar Session
Join us in room Mänty & Honka, Agora (Turku) or join the online broadcast via Zoom.
Abstract
Large language models (LLMs) are now used across countries, languages, and communities. Though multilingual LLMs show remarkable proficiency in different languages, it remains unclear what it actually means for these models to understand the culture of the corresponding language. Is recognising familiar foods, festivals, or expressions enough? Or should a culturally competent model also be able to identify the assumptions, social norms, and values that give those expressions meaning, and adapt them appropriately in another cultural context?
This talk will address the difference between cultural knowledge and cross-cultural reasoning. The speaker will try to discuss benchmarks that test whether language models can recognise culturally specific meaning, infer culturally appropriate expressions, and adapt them across various cultural contexts.
The broader discussion will consider how cultural competence should be evaluated, why conventional closed-form benchmarks may provide an incomplete picture, and what is needed to build AI systems that communicate more reliably across cultural boundaries.
About the Speaker
Md Mohsinul Kabir is a second-year ELLIS PhD student at the University of Manchester, supervised by Prof. Sophia Ananiadou and co-supervised by Prof. Shaoxiong Ji from the University of Turku. His research focuses on culturally adaptive Natural Language Processing, combining ideas from NLP, Computational Social Science, and machine learning.
He studies how language models learn, represent, and reason about culture, and how they can be made more reliable across diverse cultural contexts. His work spans cultural representation in autoregressive language models, the adaptation of LLMs for culturally sensitive tasks such as mental health, irony, and sarcasm, and the analysis of cultural bias through model interpretability.