Core Concepts
Large language models handle multilingualism by translating queries into English, processing them using English with multilingual knowledge, and then translating responses back into the original language.
Abstract
Large language models (LLMs) demonstrate remarkable performance across multiple languages. A framework is proposed to understand how LLMs handle multilingual inputs, involving understanding, problem-solving, and response generation. Language-specific neurons are identified and their significance measured through a novel method called PLND. Deactivating these neurons impacts the multilingual capabilities of LLMs significantly. Fine-tuning language-specific neurons enhances multilingual abilities with minimal training effort.
Stats
Large language models demonstrate remarkable performance across multiple languages.
Deactivating language-specific neurons impacts the multilingual capabilities of LLMs significantly.
Fine-tuning language-specific neurons enhances multilingual abilities with minimal training effort.
Quotes
"LLMs understand the user input and convert diverse linguistic features into a unified representation."
"We propose a new framework that conceptualizes the operational stages of LLMs when processing multilingual inputs."