Abstract

Large language models are increasingly deployed across diverse cultural contexts, yet they often encode and perpetuate subtle cultural biases. In this talk, I will explore how these biases manifest in practice, drawing on my work examining name-based cultural stereotyping and the challenges that arise when models encounter users with complex, multicultural identities. Moving beyond behavioural analysis, I will then discuss mechanistic approaches to understanding how cultural representations and biases are encoded within model internals, and how such analyses can inform more targeted mitigation strategies. Finally, I will present ongoing work on understanding how value-based post-training shapes model behaviour across cultures.

Bio: Siddhesh is a PhD student at the University of Copenhagen and currently visiting researcher at the Centre for Human-Inspired AI (CHIA) and the Language Technology Lab at the University of Cambridge through the ELLIS fellowship. His research interests lie at the intersection of cultural NLP , computational social science, and interpretability.