Language-Specific Neurons: the Key to Multilingual Capabilities in Large Language Models

Large language models (LLMs) demonstrate remarkable multilingual capabilitieswithout being pre-trained on specially curated multilingual parallel corpora.It remains a challenging problem to explain the underlying mechanisms by whichLLMs process multilingual texts. In this paper, we delve into the compositionof Transformer architectures in LLMs to pinpoint language-specific regions.Specially, we propose a novel detection method, language activation probabilityentropy (LAPE), to identify language-specific neurons within LLMs. Based onLAPE, we conduct comprehensive experiments on two representative LLMs, namelyLLaMA-2 and BLOOM. Our findings indicate that LLMs' proficiency in processing aparticular language is predominantly due to a small subset of neurons,primarily situated in the models' top and bottom layers. Furthermore, weshowcase the feasibility to "steer" the output language of LLMs by selectivelyactivating or deactivating language-specific neurons. Our research providesimportant evidence to the understanding and exploration of the multilingualcapabilities of LLMs.