Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models
This research provides insights into the internal mechanisms of LLMs for multilingual processing, which is incremental but important for understanding and controlling model behavior.
The paper tackled the problem of explaining how large language models (LLMs) process multilingual texts by identifying language-specific neurons using a novel detection method called LAPE, and found that a small subset of neurons in top and bottom layers is key to language proficiency, enabling language steering.
Large language models (LLMs) demonstrate remarkable multilingual capabilities without being pre-trained on specially curated multilingual parallel corpora. It remains a challenging problem to explain the underlying mechanisms by which LLMs process multilingual texts. In this paper, we delve into the composition of Transformer architectures in LLMs to pinpoint language-specific regions. Specially, we propose a novel detection method, language activation probability entropy (LAPE), to identify language-specific neurons within LLMs. Based on LAPE, we conduct comprehensive experiments on several representative LLMs, such as LLaMA-2, BLOOM, and Mistral. Our findings indicate that LLMs' proficiency in processing a particular language is predominantly due to a small subset of neurons, primarily situated in the models' top and bottom layers. Furthermore, we showcase the feasibility to "steer" the output language of LLMs by selectively activating or deactivating language-specific neurons. Our research provides important evidence to the understanding and exploration of the multilingual capabilities of LLMs.