Distribution-aware Identification of Multilingual Language Neurons

This method identifies language-specific neurons in mLLMs by analyzing pairwise activation distribution overlaps across the full activation range. It improves upon entropy-based methods by capturing the relational nature of multilingual representations.