Where Culture Lives in LLMs: Neuron Localisation and Causal Interventions
A CORE Talk took on a hybrid format, connecting Aarhus and Tehran via Zoom for a cross-continental exploration of culture in multilingual language models.
We were pleased to welcome Lukas Galke Poech (University of Southern Denmark) and Danial Namazifard(University of Tehran), who presented their collaborative research on how cultural representations are encoded in large language models (LLMs).
Their talk introduced the concept of culture neurons — sparse neural populations whose activity both predicts and causally influences culturally conditioned behavior, while remaining disentangled from language-specific features. Using MUREL, an 85.2M-token corpus spanning six cultures, they combined localisation analysis with causal interventionsto identify and manipulate these neurons.
Their findings suggest that cultural tendencies in LLMs tend to cluster in upper network layers and can be selectively adjusted with minimal impact on overall model performance. This opens new possibilities for understanding and guiding cultural dynamics in language technologies.
Many thanks to Lukas and Danial for an insightful and boundary-crossing session — and for advancing the conversation on culture, computation, and critical AI research.