Elamur Research⌕
☰
home / papers / 2026 / Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning

Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning

Paper · 2026 · Xue Zhang, Yunlong Liang, Fandong Meng et al.

Summary

Reasoning-oriented LLMs that think before answering often handle non-English prompts poorly. They tend to drift into English or a mix of languages during reasoning, and their accuracy drops compared with English. The authors introduce M-Thinker, a family of models (1.5B, 4B, 7B) trained with GRPO using two new reward signals alongside the usual format and correctness rewards:

  • Language Consistency (LC) reward: a strict penalty applied whenever langdetect finds anything other than the input language in the reasoning or the final answer.
  • Cross-lingual Thinking Alignment (CTA) reward: an LLM judge (DeepSeek-V3-0324) scores how closely the non-English reasoning matches the model's own correct English solution. This lets the model's stronger English reasoning act as a teacher.

Training proceeds in stages: cold-start SFT, then rejection sampling, then iterative RL. On MMATH and PolyMath, M-Thinker reaches almost complete language consistency while improving accuracy. It also transfers to languages it was not trained on, most effectively to languages related to the training ones. This matters for languages such as Ukrainian, where users need reasoning that is both readable and correct in their own language.

Key points

  • The LC reward is binary: 0 only if both the reasoning and the answer are detected as purely the input language, and −1 otherwise. The CTA score ranges from 0 to 1 and reflects overlap in intermediate results.
  • Training covered five languages with 3K samples each and two RL iterations. Code is available at github.com/XZhang00/M-Thinker.
  • Out-of-domain transfer follows language families. Training on fr helped pt, es and en, while training on ja helped zh and vi. Generalization to ko was the lowest.
  • The CTA judge was checked against human annotators for ja, ko, fr, pt and th, giving an average Pearson correlation of 0.88 (range 0.83–0.93).
  • In a human audit of code-switching, 29% of responses from the 1.5B model mixed languages, compared with 7% for the 7B model. The authors note that CTA may be unreliable for extremely low-resource languages.

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

🏢TencentShenzhen · China134 works · 19,743 citations🎓Beijing Jiaotong UniversityBeijing · China22 works · 1,754 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 Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Kaiyu Huang, Yufeng Chen, Xu Jinan and Jie Zhou. 2026. Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.951

@inproceedings{zhang2026think,
  title = {{Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning}},
  author = {Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Kaiyu Huang, Yufeng Chen, Xu Jinan and Jie Zhou},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.951},
  url = {https://doi.org/10.18653/v1/2026.acl-long.951}
}

References

38 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. Preprint, arXiv:1911.02116. find
  2. DeepSeek-AI. 2024. Deepseek-v3 technical report. Preprint, arXiv:2412.19437. find
  3. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.
  4. Fahim Faisal, Kaiqiang Song, Song Wang, Simin Ma, Shujian Liu, Haoyun Deng, and Sathish Reddy Indurthi. 2025. Aligning multilingual reasoning with verifiable semantics from a high-resource expert model. Preprint, arXiv:2509.25543. — [arXiv:2509.25543](https://arxiv.org/abs/2509.25543) find
  5. Yuchun Fan, Yongyu Mu, YiLin Wang, Lei Huang, Junhao Ruan, Bei Li, Tong Xiao, Shujian Huang, Xiaocheng Feng, and Jingbo Zhu. 2025a. SLAM: Towards efficient multilingual reasoning via selective language alignment. In Proceedings of the 31st International Conference on Computational Linguistics, pages…
  6. Yuchun Fan, Yilin Wang, Yongyu Mu, Lei Huang, Bei Li, Xiaocheng Feng, Tong Xiao, and Jingbo Zhu. 2025b. Language-specific layer matters: Efficient multilingual enhancement for large vision-language models. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 12473– 12500. find
  7. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Yuanzhuo Wang, Wen Gao, Lionel Ni, and Jian Guo. 2025. A survey on llm-as-a-judge. Preprint, arXiv:2411.15594. find
  8. Kaiyu Huang, Fengran Mo, Xinyu Zhang, Hongliang Li, You Li, Yuanchi Zhang, Weijian Yi, Yulong Mao, Jinchen Liu, Yuzhuang Xu, Jinan Xu, Jian-Yun Nie, and Yang Liu. 2025. A survey on large language models with multilingualism: Recent advances and new frontiers. Preprint, arXiv:2405.10936. — [arXiv:2405.10936](https://arxiv.org/abs/2405.10936) find
  9. Jaedong Hwang, Kumar Tanmay, Seok-Jin Lee, Ayush Agrawal, Hamid Palangi, Kumar Ayush, Ila Fiete, and Paul Pu Liang. 2025. Learn globally, speak locally: Bridging the gaps in multilingual reasoning. Preprint, arXiv:2507.05418. — [arXiv:2507.05418](https://arxiv.org/abs/2507.05418) find
  10. Jungyup Lee, Jemin Kim, Sang Park, and SeungJae Lee. 2025. Making qwen3 think in korean with reinforcement learning. Preprint, arXiv:2508.10355. — [arXiv:2508.10355](https://arxiv.org/abs/2508.10355) find
  11. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. Preprint, arXiv:2305.20050.
  12. Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. 2024. Statistical rejection sampling improves preference optimization. Preprint, arXiv:2309.06657.
  13. Wenyang Luo, Wayne Xin Zhao, Jing Sha, Shijin Wang, and Ji-Rong Wen. 2025. Mmath: A multilingual benchmark for mathematical reasoning. Preprint, arXiv:2505.19126. find
  14. Mistral-AI. 2025. Magistral. arXiv:2506.10910. arXiv e-prints, OpenAI. 2025. Competitive programming with large reasoning models. Preprint, arXiv:2502.06807. find
  15. Cheonbok Park, Jeonghoon Kim, Joosung Lee, Sanghwan Bae, Jaegul Choo, and Kang Min Yoo. 2025. Cross-lingual collapse: How language-centric foundation models shape reasoning in large language models. Preprint, arXiv:2506.05850.
  16. Rui Qi, Zhibo Man, Yufeng Chen, Fengran Mo, Jinan Xu, and Kaiyu Huang. 2025. SoT: Structured-of-thought prompting guides multilingual reasoning in large language models. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 11024–11039, Suzhou, China. Association for Comput… find
  17. Rui Qi, Fengran Mo, Yufeng Chen, Xue Zhang, Shuo Wang, Hongliang Li, Jinan Xu, Meng Jiang, Jian-Yun Nie, and Kaiyu Huang. 2026. Language-coupled reinforcement learning for multilingual retrieval-augmented generation. Preprint, arXiv:2601.14896. 20775 10
  18. Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’20, page 3505–3506,… find
  19. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  20. Shuaijie She, Wei Zou, Shujian Huang, Wenhao Zhu, Xiang Liu, Xiang Geng, and Jiajun Chen. 2024. Mapo: Advancing multilingual reasoning through multilingual alignment-as-preference optimization. Preprint, arXiv:2401.06838.
  21. Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256.
  22. 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? Preprint, arXiv:2505.17407.
  23. Bryan Chen Zhengyu Tan, Weihua Zheng, Zhengyuan Liu, Nancy F. Chen, Hwaran Lee, Kenny Tsu Wei Choo, and Roy Ka-Wei Lee. 2026. BLEnD-vis: Benchmarking multimodal cultural understanding in vision language models. In Proceedings of the 19th Conference of the European Chapter of the Association for Comp…
  24. Haoyu Wang, Shuo Wang, Yukun Yan, Xujia Wang, Zhiyu Yang, Yuzhuang Xu, Zhenghao Liu, Ning Ding, Xu Han, Zhiyuan Liu, and Maosong Sun. 2024. Ultralink: An open-source knowledge-enhanced multilingual supervised fine-tuning dataset. Preprint, arXiv:2402.04588. find
  25. Jiaan Wang, Fandong Meng, and Jie Zhou. 2025a. Deeptrans: Deep reasoning translation via reinforcement learning. Preprint, arXiv:2504.10187.
  26. Yiming Wang, Pei Zhang, Jialong Tang, Haoran Wei, Baosong Yang, Rui Wang, Chenshu Sun, Feitong Sun, Jiran Zhang, Junxuan Wu, Qiqian Cang, Yichang Zhang, Fei Huang, Junyang Lin, Fei Huang, and Jingren Zhou. 2025d. Polymath: Evaluating mathematical reasoning in multilingual contexts. arXiv preprint ar… — [arXiv:2504.18428](https://arxiv.org/abs/2504.18428) find
  27. Zheng Weihua, Xin Huang, Zhengyuan Liu, Tarun Kumar Vangani, Bowei Zou, Xiyan Tao, Yuhao Wu, AiTi Aw, Nancy F Chen, and Roy Ka-Wei Lee. 2026. Adamcot: Rethinking cross-lingual factual reasoning through adaptive multilingual chain-of-thought. In Proceedings of the AAAI Conference on Artificial Intell… find
  28. Zheng Weihua, Roy Ka-Wei Lee, Zhengyuan Liu, Wu Kui, AiTi Aw, and Bowei Zou. 2025. CCL-XCoT: An efficient cross-lingual knowledge transfer method for mitigating hallucination generation. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 1768–1788, Suzhou, China. Associa… find
  29. Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, and Xiangzheng Zhang. 2025. Lightr1: Curriculum sft, dpo and rl for long cot from scratch and beyond. Preprint, arXiv:2503.10460. find
  30. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025a. Qwen3 technical report. Preprint, arXiv:2505.09388.
  31. Wen Yang, Junhong Wu, Chen Wang, Chengqing Zong, and Jiajun Zhang. 2025b. Language imbalance driven rewarding for multilingual self-improving. Preprint, arXiv:2410.08964. — [arXiv:2410.08964](https://arxiv.org/abs/2410.08964) find
  32. Songming Zhang, Xue Zhang, Tong Zhang, Bojie Hu, Yufeng Chen, and Jinan Xu. 2025a. Aligndistil: Token-level language model alignment as adaptive policy distillation. Preprint, arXiv:2503.02832. find
  33. Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Yufeng Chen, Jinan Xu, and Jie Zhou. 2024. Multilingual knowledge editing with language-agnostic factual neurons. Preprint, arXiv:2406.16416.
  34. Jiaan Wang, Fandong Meng, and Jie Zhou. 2025b. Extrans: Multilingual deep reasoning translation via exemplar-enhanced reinforcement learning. Preprint, arXiv:2505.12996. — [arXiv:2505.12996](https://arxiv.org/abs/2505.12996) find
  35. Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Yufeng Chen, Jinan Xu, and Jie Zhou. 2025b. Cm-align: Consistency-based multilingual alignment for large language models. Preprint, arXiv:2509.08541. find
  36. Mingyang Wang, Lukas Lange, Heike Adel, Yunpu Ma, Jannik Strötgen, and Hinrich Schütze. 2025c. Language mixing in reasoning language models: Patterns, impact, and internal causes. Preprint, arXiv:2505.14815.
  37. Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Yufeng Chen, Jinan Xu, and Jie Zhou. 2025c. Less, but better: Efficient multilingual expansion for LLMs via layer-wise mixture-of-experts. In Proceedings of the 63rd Annual Meeting of the Association 20776 11 for Computational Linguistics (Volu…
  38. Xue Zhang, Songming Zhang, Yunlong Liang, Fandong Meng, Yufeng Chen, Jinan Xu, and Jie Zhou. 2025d. A dual-space framework for general knowledge distillation of large language models. Preprint, arXiv:2504.11426. — [arXiv:2504.11426](https://arxiv.org/abs/2504.11426) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-think-natively-unlocking-multilingual-reasoning/