Isolating Culture Neurons in Multilingual Large Language Models

2025-08-04

AI Summary

This study investigates the neural encoding mechanisms of culture-specific knowledge in multilingual large language models (LLMs), specifically addressing the distinction, localization, and interaction between culture-specific and language-specific neurons. Method: We propose a neuron localization framework based on causal attribution and cross-layer intervention, and introduce MUREL—the first large-scale, multiculturally annotated evaluation benchmark—enabling precise isolation of culture neurons. Contribution/Results: Empirical analysis reveals that cultural representations are predominantly encoded in higher transformer layers; distinct cultures occupy separable, independently controllable neural subspaces; and cultural and linguistic representations exhibit both statistical and causal decomposability. These findings establish a foundation for fine-grained, non-invasive editing of cultural biases, thereby significantly improving cross-cultural fairness and value alignment in LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Language and culture are deeply intertwined, yet it is so far unclear how and where multilingual large language models encode culture. Here, we extend upon an established methodology for identifying language-specific neurons and extend it to localize and isolate culture-specific neurons, carefully disentangling their overlap and interaction with language-specific neurons. To facilitate our experiments, we introduce MUREL, a curated dataset of 85.2 million tokens spanning six different cultures. Our localization and intervention experiments show that LLMs encode different cultures in distinct neuron populations, predominantly in upper layers, and that these culture neurons can be modulated independently from language-specific neurons or those specific to other cultures. These findings suggest that cultural knowledge and propensities in multilingual language models can be selectively isolated and edited - promoting fairness, inclusivity, and alignment. Code and data is available at https://github.com/namazifard/Culture_Neurons .

Problem

• Identify how multilingual LLMs encode cultural information
• Disentangle culture-specific neurons from language-specific ones
• Enable selective editing of cultural knowledge in LLMs

Innovation

• Extend methodology to isolate culture-specific neurons
• Introduce MUREL dataset for culture experiments
• Modulate culture neurons independently in upper layers