CRANE: Causal Relevance Analysis of Language-Specific Neurons in Multilingual Large Language Models

Researchers introduce CRANE, a new framework for analyzing how multilingual large language models organize language capabilities at the neuron level. The method uses targeted interventions to identify language-specific neurons based on functional necessity rather than activation patterns, revealing asymmetric specialization where neurons contribute selectively to specific languages while maintaining broader functionality.

• →CRANE framework redefines language specificity in neural networks by measuring functional necessity rather than activation magnitude.
• →Neuron-level interventions reveal asymmetric patterns where masking language-specific neurons degrades target language performance while preserving other languages.
• →The method identifies language-selective but non-exclusive neuron specializations in multilingual models.
• →Testing on English, Chinese, and Vietnamese shows CRANE isolates language components more precisely than activation-based methods.
• →The framework will be made publicly available for further research.