This work was done in collaboration between
@copenlu.bsky.social and UILab. Cannot thank the co-authors, Seogyeong Jeong, @siddheshp.bsky.social, Jisu Shin, Jiho Jin, Junho Myung, and supervisors @iaugenstein.bsky.social, Alice Oh enough!
🌐 What we found
1️⃣ LLMs encode Western-dominance bias and cultural flattening in their internals.
2️⃣ Internalized cultural biases do not necessarily align with extrinsic biases.
3️⃣ Low-resource cultures are less affected by these biases, likely due to limited training data.
🙋 How do Large Language Models internally process cultural knowledge? 🌐
Happy to share our new preprint "Entangled in Representations: Mechanistic Investigation of Cultural Biases in Large Language Models"
📃 Paper: arxiv.org/abs/2508.08879
