dataset_info:
features:
- name: id
dtype: int64
- name: problem
dtype: string
- name: solution
dtype: string
- name: answer
dtype: string
- name: url
dtype: string
- name: year
dtype: string
splits:
- name: en
num_bytes: 139586
num_examples: 30
- name: ja
num_bytes: 141630
num_examples: 30
- name: zh
num_bytes: 138216
num_examples: 30
- name: ru
num_bytes: 145896
num_examples: 30
- name: es
num_bytes: 139938
num_examples: 30
- name: fr
num_bytes: 140144
num_examples: 30
- name: de
num_bytes: 140273
num_examples: 30
- name: sw
num_bytes: 139589
num_examples: 30
- name: bn
num_bytes: 151343
num_examples: 30
- name: te
num_bytes: 152964
num_examples: 30
- name: th
num_bytes: 149720
num_examples: 30
download_size: 902117
dataset_size: 1579299
configs:
- config_name: default
data_files:
- split: en
path: data/en-*
- split: ja
path: data/ja-*
- split: zh
path: data/zh-*
- split: ru
path: data/ru-*
- split: es
path: data/es-*
- split: fr
path: data/fr-*
- split: de
path: data/de-*
- split: sw
path: data/sw-*
- split: bn
path: data/bn-*
- split: te
path: data/te-*
- split: th
path: data/th-*
When Models Reason in Your Language: Controlling Thinking Trace Language Comes at the Cost of Accuracy
https://arxiv.org/abs/2505.22888
Jirui Qi, Shan Chen, Zidi Xiong, Raquel Fernández, Danielle S. Bitterman, Arianna Bisazza Recent Large Reasoning Models (LRMs) with thinking traces have shown strong performance on English reasoning tasks. However, their ability to think in other languages is less studied. This capability is as important as answer accuracy for real world applications because users may find the reasoning trace useful for oversight only when it is expressed in their own language. We comprehensively evaluate two leading families of LRMs on our XReasoning benchmark and find that even the most advanced models often revert to English or produce fragmented reasoning in other languages, revealing a substantial gap in multilingual reasoning. Prompt based interventions that force models to reason in the users language improve readability and oversight but reduce answer accuracy, exposing an important trade off. We further show that targeted post training on just 100 examples mitigates this mismatch, though some accuracy loss remains. Our results highlight the limited multilingual reasoning capabilities of current LRMs and outline directions for future work. Code and data are available at this https URL.
Please cite if you find the data helpful:
@misc{qi2025modelsreasonlanguagecontrolling,
title={When Models Reason in Your Language: Controlling Thinking Trace Language Comes at the Cost of Accuracy},
author={Jirui Qi and Shan Chen and Zidi Xiong and Raquel Fernández and Danielle S. Bitterman and Arianna Bisazza},
year={2025},
eprint={2505.22888},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2505.22888},
}