EfficientXLang: Towards Improving Token Efficiency Through Cross-Lingual Reasoning

2025-06-30

AI Summary

Current language reasoning model (LRM) research overemphasizes English, neglecting the multilingual pretraining foundation of modern LLMs and overlooking potential token-efficiency gains in non-English languages. Method: We systematically evaluate DeepSeek-R1, Qwen2.5, and Qwen3 across seven languages on four mathematical reasoning benchmarks to assess cross-lingual reasoning efficiency and accuracy. Contribution/Results: Non-English languages—including Chinese and Spanish—reduce input token consumption by 23.6% on average, without sacrificing final English-equivalent accuracy after back-translation. This efficiency gain is most pronounced in models with strong multilingual capabilities. Our study provides the first empirical evidence that multilingual competence serves as a key intrinsic mechanism for improving reasoning token efficiency—revealing a novel paradigm for cost-effective, high-efficiency language reasoning.

Technology Category

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

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Despite recent advances in Language Reasoning Models (LRMs), most research focuses solely on English, even though many models are pretrained on multilingual data. In this work, we investigate: Is English the most token-efficient language for reasoning? We evaluate three open-source RLMs: DeepSeek R1, Qwen 2.5 and Qwen 3, across four math datasets and seven typologically diverse languages. We find that reasoning in non-English languages not only reduces token usage, but also preserves accuracy. These gains persist even after translating the reasoning traces into English, suggesting genuine shifts in reasoning behavior rather than surface-level linguistic effects. The extent of improvement, however, depends on the models multilingual strength. Our findings motivate a broader view of reasoning in language models, highlighting the potential of multilingual reasoning and the importance of strong multilingual foundations. The code for our work can be found: https://github.com/microsoft/EfficientXLang.

Problem

• Investigates token efficiency in non-English reasoning tasks
• Compares reasoning performance across diverse languages and models
• Demonstrates multilingual reasoning reduces tokens while preserving accuracy

Innovation

• Cross-lingual reasoning reduces token usage
• Non-English reasoning preserves model accuracy
• Multilingual strength impacts efficiency gains