Crosslingual Capabilities and Knowledge Barriers in Multilingual Large Language Models

TL;DR

This paper evaluates multilingual large language models' ability to transfer knowledge across languages, revealing limitations in deep crosslingual understanding and proposing fine-tuning as a solution.

Contribution

It identifies the crosslingual knowledge barriers in LLMs and demonstrates that fine-tuning on mixed-language data effectively reduces these gaps.

Findings

Models show surface-level crosslingual abilities

Deep crosslingual transfer remains limited

Fine-tuning on mixed-language data improves crosslingual performance

Abstract

Large language models (LLMs) are typically multilingual due to pretraining on diverse multilingual corpora. But can these models relate corresponding concepts across languages, i.e., be crosslingual? This study evaluates state-of-the-art LLMs on inherently crosslingual tasks. We observe that while these models show promising surface-level crosslingual abilities on machine translation and embedding space analyses, they struggle with deeper crosslingual knowledge transfer, revealing a crosslingual knowledge barrier in both general (MMLU benchmark) and domain-specific (Harry Potter quiz and TOFU benchmark) contexts. Since simple inference-time mitigation methods offer only limited improvement, we propose fine-tuning of LLMs on mixed-language data, which effectively reduces these gaps, even when using out-of-domain datasets like WikiText. Our findings suggest the need for explicit optimization to unlock the full crosslingual potential of LLMs. Our code is publicly available at https://github.com/google-research/crosslingual-knowledge-barriers.