When Language Shapes Thought: Cross-Lingual Transfer of Factual Knowledge in Question Answering

TL;DR

This work investigates cross-lingual transfer of factual knowledge in multilingual LLMs through Language-to-Thought prompting (L2T), which decouples input language from the model's internal thinking language. By evaluating L2T in Consistent, Transfer, and Align configurations across Chinese, Korean, and Arabic with four models, the study shows that aligning internal thought with the source language and maintaining input–output language consistency yields superior knowledge utilization, often outperforming traditional English-centric prompting. The results reveal that multilingual models are not universally improved by translating inputs to English; instead, the model's internal language dynamics and data distribution across languages critically shape performance. The findings advance our understanding of language-thought interactions in LLMs and offer practical guidance for designing prompts that optimize cross-lingual factual knowledge retrieval in real-world multilingual settings.

Abstract

Multilingual large language models (LLMs) offer promising opportunities for cross-lingual information access, yet their use of factual knowledge remains highly sensitive to the input language. Prior work has addressed this through English prompting and evaluation, assuming that English-based reasoning is universally beneficial. In this work, we challenge that assumption by exploring factual knowledge transfer from non-English to English through the lens of Language and Thought Theory. We introduce Language-to-Thought (L2T) prompting, which aligns the model's internal ''thinking'' language with the source of knowledge. Across three languages and four models, L2T consistently outperforms English-based reasoning, reversing the expected advantage of English prompts. Our code is available at https://github.com/GeomeunByeol/Language2Thought.

Figures (3)

• Figure 1: An example illustrating how language influences knowledge transfer. When a question requiring specific factual knowledge about Korea was in English, it led to incorrect responses. Employing our proposed L2T prompting strategy—which shifts the model’s language of thought—resulted in correct outcomes by aligning the knowledge required for the task with the internal thinking language.
• Figure 2: Average PPL on the Llama. (numbers) indicate the number of evaluation samples.
• Figure 3: L2T-Align results on the Gemini model, illustrating knowledge transfer attempts using input languages beyond English.