Why Do Multilingual Reasoning Gaps Emerge in Reasoning Language Models?

Researchers identify that reasoning language models exhibit worse performance in low-resource languages due to failures in language understanding rather than reasoning capability itself. The study proposes Selective Translation, which strategically adds English translations only when understanding failures are detected, achieving near full-translation performance while translating just 20% of inputs.

The multilingual reasoning gap in large language models represents a critical equity and accessibility challenge in AI development. While reasoning language models excel at complex problem-solving tasks, this capability degrades significantly when processing non-English inputs, creating disparities that affect billions of non-English speakers globally. This research advances the field by moving beyond documenting the problem to identifying its root cause: the models struggle to convert multilingual inputs into the English-dominant reasoning traces that power their internal reasoning processes, rather than failing at reasoning itself.

The gap has emerged as language models have grown more capable, with researchers increasingly documenting performance disparities across languages. However, pinpointing whether the issue stems from understanding, reasoning, or integration has remained ambiguous. This work provides crucial clarity by showing understanding failures are the primary culprit, enabling targeted interventions rather than wholesale model retraining.

The Selective Translation approach offers practical significance for developers and organizations serving multilingual user bases. By detecting when a model struggles to understand input and only then providing English translation, the method maintains computational efficiency while improving performance. The ability to achieve near full-translation results with 80% fewer translations has direct implications for deployment costs and latency in production systems.

Looking forward, this research opens pathways for more sophisticated detection methods and mitigation strategies. The public release of code and data enables broader community research into understanding-centric approaches across language models. As AI systems increasingly serve global audiences, solving multilingual reasoning gaps becomes essential for responsible AI deployment.

• →Multilingual reasoning gaps stem primarily from language understanding failures, not reasoning deficits.
• →Selective Translation bridges performance gaps while translating only ~20% of inputs through targeted detection.
• →Supervised detection methods effectively identify when models fail to understand multilingual inputs.
• →Understanding-centric solutions offer more efficient alternatives to wholesale model retraining across languages.
• →Publicly available code enables broader research into equitable multilingual reasoning systems.