Elamur Research⌕
☰
home / papers / 2025 / Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?

Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?

Paper · 2025 · Zhi Rui Tam, Cheng-Kuang Wu, Yu Ying Chiu et al.

Summary

Large reasoning models (LRMs) write an explicit thinking trace before they answer. It is unclear which language that trace uses when the prompt is not in English, and whether the choice matters. Tam et al. show that open-weight LRMs such as QwQ-32B and Qwen3-30B-A3B mostly think in English, and sometimes in Chinese. The authors call these the "reasoning hubs." The models still reply in the user's language.

To test the consequences, the authors steer the thinking language by inserting a short language-specific seed phrase right after the <think> tag. They evaluate eight languages (English, Chinese, Spanish, Russian, Japanese, Korean, Telugu, Swahili) on MMMLU, MATH-500, CulturalBench-Hard and LMSYS-Toxic. Forcing models to think in the input language lowers accuracy on math and knowledge tasks, most sharply for low-resource languages. Switching to a hub language keeps or improves scores. On cultural and safety tasks the pattern reverses: native-language thinking helps cultural answers and can reduce toxic output.

For teams building non-English applications, including Ukrainian ones (not covered in the study), the language of hidden reasoning is a tunable factor. It involves an accuracy-versus-safety trade-off.

Key points

  • Hub phenomenon: For the larger models, the reasoning language differs from the answer language in over 90% of cases. Mid-resource inputs (Japanese, Korean) are reasoned about either in the input language or in Chinese. Low-resource inputs (Swahili, Telugu) default to English.
  • Steering method: Seed phrases were chosen as the most frequent openings of native-language traces. Examples include "Okay" (en), "Хорошо" (ru), "まず" (ja), "嗯" (zh) and "Primero" (es).
  • Effect size: Forcing reasoning in a non-hub language costs up to about 30 percentage points of accuracy for low-resource languages. Reasoning in a hub language improves math and knowledge tasks by up to 26.8%.
  • Benchmarks used:
    • MMMLU: human-translated MMLU in 14 languages; 32 of the 57 subjects were used.
    • MATH-500: machine-translated into 7 languages with the Google Translate API.
    • CulturalBench-Hard: 1,200 questions covering 30 countries/regions.
    • LMSYS-Toxic: 2,000 prompts from LMSYS-1M.
  • Reasoning-style analysis: LLM annotators segment traces and classify them into four reasoning habits.
    • Chinese seed phrases correlate with subgoal setting (r ≈ 0.50) and verification (r = 0.41).
    • English seed phrases correlate with backward chaining (r = 0.30).
    • These findings are correlational, not causal.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of WashingtonSeattle · United States353 works · 200,043 citations🎓National Taiwan UniversityTaipei · Taiwan29 works · 2,116 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 2Research subjects and languages the work is about.Reasoning · 1,426 worksMultilingual & cross-lingual NLP · 828 works

Cite

Cite as Zhi Rui Tam, Cheng-Kuang Wu, Yu Ying Chiu, Chieh-Yen Lin, Yun-Nung Chen and Hung-yi Lee. 2025. Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?. arXiv:2505.17407.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.17407 · arXiv 2505.17407

@misc{tam2025language,
  title = {{Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?}},
  author = {Zhi Rui Tam, Cheng-Kuang Wu, Yu Ying Chiu, Chieh-Yen Lin, Yun-Nung Chen and Hung-yi Lee},
  year = {2025},
  doi = {10.48550/arxiv.2505.17407},
  eprint = {2505.17407},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.17407}
}

References

24 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Allan BI Bernardo and Marissa O Calleja. The effects of stating problems in bilingual students’ first and second languages on solving mathematical word problems. The Journal of Genetic Psychology, 166(1):117–129, 2005. find
  2. Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, et al. Reasoning models don’t always say what they think. arXiv preprint arXiv:2505.05410, 2025. find
  3. Yu Ying Chiu, Liwei Jiang, Bill Yuchen Lin, Chan Young Park, Shuyue Stella Li, Sahithya Ravi, Mehar Bhatia, Maria Antoniak, Yulia Tsvetkov, Vered Shwartz, et al. Culturalbench: a robust, diverse and challenging benchmark on measuring the (lack of) cultural knowledge of llms. arXiv preprint arXiv:241… — [arXiv:2410.02677](https://arxiv.org/abs/2410.02677) find
  4. Julen Etxaniz, Gorka Azkune, Aitor Soroa, Oier Lacalle, and Mikel Artetxe. Do multilingual language models think better in english? In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Paper…
  5. Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. arXiv preprint arXiv:2503.01307, 2025.
  6. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.
  7. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020.
  8. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021.
  9. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021.
  10. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024.
  11. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024.
  12. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025.
  13. Gustaw Opiełka, Hannes Rosenbusch, and Claire E Stevenson. Analogical reasoning inside large language models: Concept vectors and the limits of abstraction. arXiv preprint arXiv:2503.03666, 2025. — [arXiv:2503.03666](https://arxiv.org/abs/2503.03666) find
  14. Simon Ott, Konstantin Hebenstreit, Valentin Liévin, Christoffer Egeberg Hother, Milad Moradi, Maximilian Mayrhauser, Robert Praas, Ole Winther, and Matthias Samwald. Thoughtsource: A central hub for large language model reasoning data. Scientific data, 10(1):528, 2023.
  15. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022, 2023.
  16. Jan Rupnik, Andrej Muhic, and P Skraba. Cross-lingual document retrieval through hub languages. In Neural Information Processing Systems Workshop, 2012. find
  17. Lisa Schut, Yarin Gal, and Sebastian Farquhar. Do multilingual llms think in english? arXiv preprint arXiv:2502.15603, 2025. — [arXiv:2502.15603](https://arxiv.org/abs/2502.15603) find
  18. Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, et al. Language models are multilingual chain-of-thought reasoners. In The Eleventh International Conference on Learning Representations, 2023.
  19. Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, et al. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference.…
  20. Zhaofeng Wu, Xinyan Velocity Yu, Dani Yogatama, Jiasen Lu, and Yoon Kim. The semantic hub hypothesis: Language models share semantic representations across languages and modalities. arXiv preprint arXiv:2411.04986, 2024. find
  21. Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning. arXiv preprint arXiv:2502.03387, 2025.
  22. Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284, 2023.
  23. Zhexin Zhang, Leqi Lei, Lindong Wu, Rui Sun, Yongkang Huang, Chong Long, Xiao Liu, Xuanyu Lei, Jie Tang, and Minlie Huang. Safetybench: Evaluating the safety of large language models with multiple choice questions. arXiv preprint arXiv:2309.07045, 2023.
  24. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric P Xing, et al. Lmsys-chat-1m: A large-scale real-world llm conversation dataset. arXiv preprint arXiv:2309.11998, 2023. Table 4 ∣ Decoding parameters used for each model during…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tam-2025-language-matters-multilingual-input-reasoning/