LLMs Are Globally Multilingual Yet Locally Monolingual: Exploring Knowledge Transfer via Language and Thought Theory

2025-05-30

AI Summary

Multilingual large language models (LLMs) exhibit unstable factual knowledge transfer from non-English to English, primarily due to strong coupling between linguistic input and internal cognitive representations, which impairs cross-lingual knowledge extraction. To address this, we propose the Language-to-Thought (L2T) prompting strategy, grounded in the Language and Thought Theory—first explicitly modeling the “language–thought–knowledge” triadic relationship. L2T mitigates language dependency by aligning latent semantic representations (“thought”) across languages rather than performing surface-level translation, and introduces a non-translation-based training augmentation. Experiments demonstrate substantial gains in cross-lingual factual reasoning, challenging the English-centrism assumption. Crucially, L2T achieves more consistent and robust multilingual knowledge integration without requiring explicit translation. This work establishes a novel paradigm for investigating cognitive mechanisms and engineering optimizations in multilingual LLMs.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Knowledge Representation Languages

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Multilingual large language models (LLMs) open up new possibilities for leveraging information across languages, but their factual knowledge recall remains inconsistent depending on the input language. While previous studies have attempted to address this issue through English-based prompting and evaluation, we explore non-English to English transfer via Language and Thought Theory. This perspective allows us to examine language-thought binding in LLMs and uncover why factual knowledge often fails to transfer effectively. We propose the Language-to-Thought (L2T) prompting strategy, which analyzes the relationship between input language, internal cognitive processes, and knowledge. Experimental results challenge the assumption that English-based approaches consistently outperform other languages and offer a novel insight that aligning the model's internal thought with the knowledge required for the task is critical for successful cross-lingual transfer. Furthermore, we show that applying L2T during training can alleviate LLMs' reliance on the input language and facilitate cross-linguistic knowledge integration without translation-based learning. Code and datasets will be available.

Problem

• Multilingual LLMs have inconsistent factual knowledge recall across languages
• English-based prompting does not always outperform other languages
• Aligning model's internal thought with task knowledge improves cross-lingual transfer

Innovation

• Proposes L2T prompting for cross-lingual knowledge transfer
• Analyzes language-thought binding in multilingual LLMs
• Reduces input language reliance via thought alignment