Do multilingual reasoning patterns extend beyond math and science?

Ascertain whether the multilingual reasoning patterns observed when Large Reasoning Models reason in English versus the question’s language on the MGSM and GPQA Diamond benchmarks (mathematical and scientific tasks) extend to open-ended questions and other task types.

Background

The paper compares Large Reasoning Models’ reasoning in English versus the language of the question across MGSM and GPQA Diamond, documenting higher final-answer accuracy and richer cognitive behaviors when reasoning in English, alongside a translation-induced failure mode (“Lost in Translation”).

The authors explicitly note that their evaluation is confined to mathematical and scientific tasks and raise the question of whether the same observed patterns would generalize to open-ended questions or other task types, identifying this as an open question.