Beyond English-Centric LLMs: What Language Do Multilingual Language Models Think in?

TL;DR

This paper investigates the internal language representations of multilingual LLMs, revealing how models activate different latent languages internally and how these shift across layers during multilingual tasks.

Contribution

It introduces the concept of latent internal languages in LLMs and analyzes their activation patterns across different models and languages, especially Japanese and English.

Findings

Llama2 relies solely on English as its internal latent language.

Swallow and LLM-jp utilize both Japanese and English as latent languages.

Models activate the latent language most related to the target language.

Abstract

In this study, we investigate whether non-English-centric LLMs, despite their strong performance, `think' in their respective dominant language: more precisely, `think' refers to how the representations of intermediate layers, when un-embedded into the vocabulary space, exhibit higher probabilities for certain dominant languages during generation. We term such languages as internal $\textbf{latent languages}$. We examine the latent language of three typical categories of models for Japanese processing: Llama2, an English-centric model; Swallow, an English-centric model with continued pre-training in Japanese; and LLM-jp, a model pre-trained on balanced English and Japanese corpora. Our empirical findings reveal that, unlike Llama2 which relies exclusively on English as the internal latent language, Japanese-specific Swallow and LLM-jp employ both Japanese and English, exhibiting dual internal latent languages. For any given target language, the model preferentially activates the latent language most closely related to it. In addition, we explore how intermediate layers respond to questions involving cultural conflicts between latent internal and target output languages. We further explore how the language identity shifts across layers while keeping consistent semantic meaning reflected in the intermediate layer representations. This study deepens the understanding of non-English-centric large language models, highlighting the intricate dynamics of language representation within their intermediate layers.