Elamur Research⌕
☰
home / papers / 2026 / Do Thoughts Depth Affect Multilingual Reasoning?

Do Thoughts Depth Affect Multilingual Reasoning?

Paper · 2026 · Linjian Yang, Xinyan Wang and Kunpeng Liu

Summary

Chain-of-Thought prompting is usually assumed to help LLM reasoning, and longer chains are often treated as better. This assumption has rarely been tested across languages with uneven training resources. The authors examine whether explicitly limiting the number of reasoning steps changes accuracy, and whether the effect depends on language resource level.

They prompt two instruction-tuned models, DeepSeek and Moonshot, on MGSM (math) and X-CSQA (commonsense). Seven typologically diverse languages are grouped into high-, medium- and low-resource tiers. Step-constrained CoT with a fixed number of steps is compared against free CoT and direct zero-shot answering, using manually translated prompts.

Adding more steps does not reliably raise accuracy. High- and mid-resource languages stay fairly stable and tend to gain most at moderate depth. Low-resource languages benefit little and often lose accuracy as chains grow. The authors attribute this to error accumulation, overthinking and spurious justifications, which weak representations make worse. They conclude that CoT depth should be tuned to the language and task, not maximised by default.

Key points

  • Six prompting settings: CoT = −1 is free "think step by step", CoT = 0 is a direct answer, and k ∈ {2, 4, 6, 8} requires exactly k concise steps. All settings share one prompt template.
  • Data and decoding:
    • MGSM contains 250 grade-school math problems.
    • X-CSQA is based on CommonsenseQA and has a 1,074-item test set.
    • Models were queried via API with temperature 0 and a cap of 1,000 output tokens.
  • Language diversity: The languages are en, zh, de, ja, sw, th and ur. They cover Latin, CJK, Thai and Perso-Arabic scripts, as well as analytic, inflectional and agglutinative morphology.
  • Illustrative results:
    • German accuracy on DeepSeek drops steadily as depth increases, on both benchmarks.
    • Swahili holds up at moderate depth but falls at k = 6–8.
    • Urdu improves slightly at moderate depth, then plateaus.
  • Interpretation: The authors argue that low-resource failures reflect representation gaps and an inability to filter reasoning noise, not missing reasoning ability. They propose adaptive per-language CoT depth as future work.

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

🎓Portland State UniversityPortland · United States7 works · 53 citations🎓Clemson UniversityClemson · United States3 works · 28 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 2Research subjects and languages the work is about.Reasoning · 1,685 worksMultilingual & cross-lingual NLP · 882 works

Cite

Cite as Linjian Yang, Xinyan Wang and Kunpeng Liu. 2026. Do Thoughts Depth Affect Multilingual Reasoning?. Proceedings of the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.mellm-1.14

@inproceedings{yang2026do,
  title = {{Do Thoughts Depth Affect Multilingual Reasoning?}},
  author = {Linjian Yang, Xinyan Wang and Kunpeng Liu},
  booktitle = {Proceedings of the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.mellm-1.14},
  url = {https://doi.org/10.18653/v1/2026.mellm-1.14}
}

References

27 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, and 1 others. 2024. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954. find
  2. Kyle Cox, Jiawei Xu, Yikun Han, Rong Xu, Tianhao Li, Chi-Yang Hsu, Tianlong Chen, Walter Gerych, and Ying Ding. 2025. Mapping from meaning: Addressing the miscalibration of prompt-sensitive language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23696– 237… find
  3. Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2024. Chain-of-verification reduces hallucination in large language models. In Findings of the association for computational linguistics: ACL 2024, pages 3563–3578.
  4. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199– 22213.
  5. Hongliang Li, Jinan Xu, Gengping Cui, Changhao Guan, Fengran Mo, and Kaiyu Huang. 2025. Multilingual collaborative defense for large language models. arXiv preprint arXiv:2505.11835. find
  6. Bill Yuchen Lin, Seyeon Lee, Xiaoyang Qiao, and Xiang Ren. 2021. Common sense beyond english: Evaluating and improving multilingual language models for commonsense reasoning. arXiv preprint arXiv:2106.06937.
  7. Fengran Mo, Zhan Su, Yuchen Hui, Jinghan Zhang, Jia Ao Sun, Zheyuan Liu, Chao Zhang, Tetsuya Sakai, and Jian-Yun Nie. 2026. Opendecoder: Open large language model decoding to incorporate document quality in rag. arXiv preprint arXiv:2601.09028. — [arXiv:2601.09028](https://arxiv.org/abs/2601.09028) find
  8. Benedikt Ebing and Goran Glavaš. 2024. To translate or not to translate: A systematic investigation of translation-based cross-lingual transfer to low-resource languages. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu… find
  9. Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan Firat, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In International conference on machine learning, pages 4411–4421. PMLR. find
  10. Chao Huang, Fengran Mo, Yufeng Chen, Changhao Guan, Zhenrui Yue, Xinyu Wang, Jinan Xu, and Kaiyu Huang. 2025. Boosting data utilization for multilingual dense retrieval. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 12362–12378, Suzhou, China. Assoc… find
  11. Kaiyu Huang, Fengran Mo, Xinyu Zhang, Hongliang Li, You Li, Yuanchi Zhang, Weijian Yi, Yulong Mao, Jinchen Liu, Yuzhuang Xu, and 1 others. 2026. A survey on large language models with multilingualism: Recent advances and new frontiers. Artificial Intelligence Review. find
  12. Kaiyu Huang, Keli Xiao, Fengran Mo, Bo Jin, Zhuang Liu, and Degen Huang. 2021. Domain-aware word segmentation for chinese language: A document-level context-aware model. Transactions on Asian and Low-Resource Language Information Processing, 21(2):1–16. find
  13. Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. 2022. Decomposed prompting: A modular approach for solving complex tasks. arXiv preprint arXiv:2210.02406.
  14. 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. find
  15. 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. arXiv preprint arXiv:2601.14896.
  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. Kimi Team, Yifan Bai, Yiping Bao, Y Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, Ningxin Chen, and 1 others. 2025. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534.
  18. Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. 2023. Towards understanding chain-of-thought prompting: An empirical study of what matters. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pag…
  19. Hao Wang, Pinzhi Huang, Jihan Yang, Saining Xie, and Daisuke Kawahara. 2025. Traveling across languages: Benchmarking cross-lingual consistency in multimodal llms. arXiv preprint arXiv:2505.15075. — [arXiv:2505.15075](https://arxiv.org/abs/2505.15075) find
  20. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  21. Zihan Wang, Stephen Mayhew, Dan Roth, and 1 others. 2019. Cross-lingual ability of multilingual bert: An empirical study. arXiv preprint arXiv:1912.07840. find
  22. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  23. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mt5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 conference of the North American chapter of the association for computational ling…
  24. Jinghan Zhang, Fengran Mo, Tharindu Cyril Weerasooriya, Ruimin Dai, Xiaoyan Han, Yanjie Fu, Dakuo Wang, and Kunpeng Liu. 2026. Starpo: Stability-augmented reinforcement policy optimization. arXiv preprint arXiv:2604.08905. — [arXiv:2604.08905](https://arxiv.org/abs/2604.08905) find
  25. Jinghan Zhang, Xiting Wang, Fengran Mo, Yeyang Zhou, Wanfu Gao, and Kunpeng Liu. 2025a. Entropy-based exploration conduction for multi-step reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 3895–3906. find
  26. Jinghan Zhang, Xiting Wang, Weijieying Ren, Lu Jiang, Dongjie Wang, and Kunpeng Liu. 2025b. Ratt: A thought structure for coherent and correct llm reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 26733–26741. find
  27. Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, and 1 others. 2022. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-do-thoughts-depth-affect-multilingual/