# Crosslingual Reasoning through Test-Time Scaling

**Authors:** Zheng-Xin Yong, M. Farid Adilazuarda, Jonibek Mansurov, Ruochen Zhang, Niklas Muennighoff, Carsten Eickhoff, Genta Indra Winata, Julia Kreutzer, Stephen H. Bach, Alham Fikri Aji  
**arXiv:** [2505.05408](https://arxiv.org/abs/2505.05408)  
**Date:** 2025-05-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.05408v1/

## Key points

- Study conducted by Zheng-Xin Yong et al. on reasoning capabilities of large language models
- Focus on English-centric models used for multilingual tasks
- Exploration of generalization of English reasoning finetuning with long chain-of-thoughts (CoTs) across different languages
- Scaling up inference compute for English-centric reasoning language models leads to improved multilingual mathematical reasoning
- Scaled-up RLMs outperform larger models in certain scenarios
- Observation of quote-and-think pattern in non-English inputs by English-centric RLMs
- Effective strategy discovered to control language in long CoT reasoning processes
- Models reason better and more efficiently in high-resource languages compared to low-resource ones
- Challenges observed in out-of-domain reasoning generalization, particularly from STEM-related topics to cultural commonsense knowledge within the English language context
- Conclusion highlights potentials and limitations of Test-Time Scaling for English-centric models
- Recommendation to allow English-centric RLMs to reason in high-resource languages while enhancing capabilities in low-resource languages and out-of-domain contexts

## Summary

In the study "Crosslingual Reasoning through Test-Time Scaling," conducted by 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, the authors delve into the reasoning capabilities of large language models with a focus on English-centric models that are often used for multilingual tasks. The research aims to explore the extent to which English reasoning finetuning with long chain-of-thoughts (CoTs) can be generalized across different languages. The findings of the study indicate that scaling up inference compute for English-centric reasoning language models (RLMs) leads to improved multilingual mathematical reasoning across various languages, including low-resource languages. Surprisingly, these scaled-up RLMs even outperform models that are twice their size in certain scenarios. Additionally, it was observed that while English-centric RLMs predominantly use English in their CoTs, they consistently adopt a quote-and-think pattern when reasoning about non-English inputs. Furthermore, the researchers discovered an effective strategy to control the language used in long CoT reasoning processes. They noted that models tend to reason better and more efficiently in high-resource languages compared to low-resource ones. However, there were challenges observed in terms of out-of-domain reasoning generalization; particularly from STEM-related topics to cultural commonsense knowledge even within the English language context. In conclusion <kwd>, Crosslingual Reasoning </kwd> highlights both the potentials and limitations of <kwd> Test-Time Scaling </kwd> for English-centric models. The authors recommend allowing English-centric RLMs to reason in high-resource languages while acknowledging the need for further research to enhance reasoning capabilities in low-resource languages and out-of-domain contexts. This comprehensive analysis sheds light on important considerations for practitioners working with large language models in multilingual settings.