Mohammadamin Shafiei
M.Sc. student · University of Milan
Human-centered AI · University of Milan
I work on natural language processing, large language model evaluation, and the social and cultural context of language technology. I study misleading yet superficially correct model outputs, social norms and dehumanization, gender bias, alignment variability across languages, and what LLMs memorize. A recurring theme is to stress-test models in realistic, multilingual, and value-laden settings rather than in isolation. My work includes benchmarks on comparative reasoning, false-premise multihop QA, and workplace humor.
Benchmarks where models look correct but are misleading, including bilingual and false-premise multihop settings.
Directional skew when models compare two sides of an issue—testing reasoning beyond a single “correct” answer.
Directional and demographic bias, social norm classification, and harms beyond classic hate speech.
Physical commonsense and social reasoning in multilingual and cross-lingual settings (including Farsi and Iranian norms).
Alignment variability, evaluation of model behavior in nuanced, subjective tasks.
For the most up-to-date list, see my Google Scholar profile. * equal contribution where noted.
