Elamur Research⌕
☰
home / papers / 2026 / The Reasoning Lingua Franca: A Double-Edged Sword for Multilingual AI

The Reasoning Lingua Franca: A Double-Edged Sword for Multilingual AI

Paper · 2026 · Alan Saji, Raj Dabre, Anoop Kunchukuttan et al.

Summary

Large reasoning models (LRMs) often think in English even when a question is posed in another language. This makes their reasoning harder for non-English users to follow and may lose linguistic or cultural nuance. The authors test whether forcing reasoning into the question's language helps or hurts. They evaluate open-weight models, mainly Qwen3 32B and DeepSeek-R1-Distill-Llama-70B, on MGSM and on a translated GPQA Diamond. Native-language reasoning is induced through a system prompt plus prefix tokens inserted at the start of the reasoning trace.

English reasoning usually gives more correct answers. The advantage grows as languages become lower-resource and as tasks get harder: it is modest on MGSM and much larger on expert-level GPQA. English traces also show more cognitive behaviors, as judged by gpt-4o-mini.

English is not always better, though. Many wrong English-reasoned answers come from mistranslating the input, a failure the authors call "Lost in Translation." The work therefore argues for investing seriously in native-language reasoning and provides baselines for measuring progress.

Key points

  • Data: GPQA Diamond was translated into Danish and five Indic languages. Languages are grouped into high-, mid- and low-resource tiers.
  • Setup: Sampling used T = 0.6, p = 0.95 and k = 20. Each configuration was run four times (avg@4 with standard deviation) via the Deepinfra API, and answers were extracted from \boxed{}.
  • Cognitive behaviors: Sub-goal setting, verification, backtracking and backward chaining were counted for Qwen QwQ-32B on MGSM, with human–LLM agreement checks. All four appear more often in English traces.
  • LiT fraction (share of English-reasoning errors caused by translation):
    • MGSM: 0.77 for low-resource languages, falling to 0.30 for high-resource ones.
    • GPQA: between 0.44 and 0.33.
  • Danish has a relatively small English-vs-native gap on GPQA, consistent with its high-resource status.
  • Code: github.com/AI4Bharat/multilingual_reasoning_analysis.

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

🎓IT University of CopenhagenCopenhagen · Denmark27 works · 2,406 citations🏢Microsoft (India)Hyderabad · India10 works · 2,129 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 Alan Saji, Raj Dabre, Anoop Kunchukuttan and Ratish Puduppully. 2026. The Reasoning Lingua Franca: A Double-Edged Sword for Multilingual AI. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-short.25

@inproceedings{saji2026reasoning,
  title = {{The Reasoning Lingua Franca: A Double-Edged Sword for Multilingual AI}},
  author = {Alan Saji, Raj Dabre, Anoop Kunchukuttan and Ratish Puduppully},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-short.25},
  url = {https://doi.org/10.18653/v1/2026.eacl-short.25}
}

References

22 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Tushar Aggarwal, Kumar Tanmay, Ayush Agrawal, Kumar Ayush, Hamid Palangi, and Paul Pu Liang. 2025. Language models’ factuality depends on the language of inquiry. arXiv preprint arXiv:2502.17955. — [arXiv:2502.17955](https://arxiv.org/abs/2502.17955) find
  2. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  3. Deepinfra. 2025. Deepinfra api. https://deepinfra.com/docs/api-reference. Accessed 2024 - 09-21. ↗
  4. 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? abs/2505.17407.
  5. Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. 2025. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. arXiv preprint arXiv:2503.01307.
  6. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118. find
  7. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, and 5 others.… find
  8. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
  9. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  10. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.
  11. OpenAI. 2024. GPT-4o mini: advancing cost-efficient intelligence. Blog post. Accessed: 2025-8-06. find
  12. Jirui Qi, Shan Chen, Zidi Xiong, Raquel Fernández, Danielle S Bitterman, and Arianna Bisazza. 2025. When models reason in your language: Controlling thinking trace language comes at the cost of accuracy. arXiv preprint arXiv:2505.22888. — [arXiv:2505.22888](https://arxiv.org/abs/2505.22888) find
  13. QwenTeam. 2025. Qwen qwq-32b. https://qwen.ai/blog?id=6aed6aa257238a0b6c77a6753f18 0350c2fecc4a&from=research.research-list. Accessed: 2025-8-06. ↗
  14. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. 2024. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling.
  15. SarvamAI. 2025. Sarvam translate. https://www.sa rvam.ai/blogs/sarvam-translate. Accessed: 2025-7-06. ↗
  16. Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, and 1 others. 2022. Language models are multilingual chain-of-thought reasoners. arXiv preprint arXiv:2210.03057.
  17. Yiming Wang, Pei Zhang, Jialong Tang, Haoran Wei, Baosong Yang, Rui Wang, Chenshu Sun, Feitong Sun, Jiran Zhang, Junxuan Wu, and 1 others. 2025. Polymath: Evaluating mathematical reasoning in multilingual contexts. arXiv preprint arXiv:2504.18428. — [arXiv:2504.18428](https://arxiv.org/abs/2504.18428) find
  18. Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2024. Do llamas work in english? on the latent language of multilingual transformers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand,…
  19. Zhaofeng Wu, Xinyan Velocity Yu, Dani Yogatama, Jiasen Lu, and Yoon Kim. 2025. The semantic hub hypothesis: Language models share semantic representations across languages and modalities. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025.… find
  20. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  21. Zheng-Xin Yong, M Farid Adilazuarda, Jonibek Mansurov, Ruochen Zhang, Niklas Muennighoff, Carsten Eickhoff, Genta Indra Winata, Julia Kreutzer, Stephen H Bach, and Alham Fikri Aji. 2025. Crosslingual reasoning through test-time scaling. arXiv preprint arXiv:2505.05408. — [arXiv:2505.05408](https://arxiv.org/abs/2505.05408) find
  22. Yiran Zhao, Wenxuan Zhang, Guizhen Chen, Kenji Kawaguchi, and Lidong Bing. 2024. How do large language models handle multilingualism? In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, Decembe…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/saji-2026-reasoning-lingua-franca-double-edged/