Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?
This work addresses linguistic biases in LRMs to improve equity for users across diverse languages, though it is incremental as it identifies issues without proposing new solutions.
The study investigated how large reasoning models (LRMs) handle multilingual inputs and found that they default to reasoning in high-resource languages like English, regardless of input language, which preserves performance on reasoning tasks but causes declines when forced to reason in low-resource languages, with effects varying by task type.
Large reasoning models (LRMs) have demonstrated impressive performance across a range of reasoning tasks, yet little is known about their internal reasoning processes in multilingual settings. We begin with a critical question: {\it In which language do these models reason when solving problems presented in different languages?} Our findings reveal that, despite multilingual training, LRMs tend to default to reasoning in high-resource languages (e.g., English) at test time, regardless of the input language. When constrained to reason in the same language as the input, model performance declines, especially for low-resource languages. In contrast, reasoning in high-resource languages generally preserves performance. We conduct extensive evaluations across reasoning-intensive tasks (MMMLU, MATH-500) and non-reasoning benchmarks (CulturalBench, LMSYS-toxic), showing that the effect of language choice varies by task type: input-language reasoning degrades performance on reasoning tasks but benefits cultural tasks, while safety evaluations exhibit language-specific behavior. By exposing these linguistic biases in LRMs, our work highlights a critical step toward developing more equitable models that serve users across diverse linguistic backgrounds.