Daryna Dementieva, Nikolay Babakov, Kathy Hämmerl, Ilseyar Alimova, Jindřich Libovický, Shu Okabe, Miras Baisbay, Lukas Edman, Abrorkhon Inomkhujaev, Antonia Karamolegkou, Mateusz Lango, Volkan Özer, Nikola Selic, Subhankar Swain, Tsedeniya Kinfe Temesgen, Galit Bary Weisberg, Alexander Fraser
5 min
Mathematical reasoning is a cornerstone for evaluating Large Language Models (LLMs), yet current benchmarks are heavily skewed toward high-resource languages like English and Chinese. This linguistic bias limits our understanding of how reasoning capabilities transfer to underrepresented languages. While recent datasets like PolyMath have expanded coverage, they still focus on high-resource settings. PluraMath addresses this by extending the PolyMath benchmark to 18 underrepresented languages across six language families, ranging from mid-resource to extreme low-resource settings.
To ensure high-quality evaluation, the researchers employed a human-curated pipeline. They first generated initial translations using automated systems and then recruited native speakers—all holding advanced degrees in mathematics or computer science—to perform rigorous manual verification. This process ensured that the natural language, mathematical terminology, and LaTeX formatting were accurate and culturally appropriate. The resulting dataset was used to benchmark 27 reasoning LLMs, spanning various scales from sub-billion parameter models to closed-source frontier systems, using three different prompting strategies: native-language reasoning, English-instructed reasoning (EN-CoT), and backtranslation.
The study confirms a persistent performance gap between high-resource and underrepresented languages. A fine-grained analysis shows that mathematical reasoning ability is not easily transferred; performance gains are only weakly associated with translation capability and are not substantially improved by prompting in high-resource languages. Instead, stronger reasoning performance correlates primarily with general instruction-following ability. Larger models, particularly proprietary systems, demonstrate greater stability across linguistic conditions, whereas smaller models often struggle with reasoning-answer misalignment and incomplete traces in underrepresented languages.
PluraMath provides a critical resource for stress-testing LLMs in diverse linguistic environments. By open-sourcing the dataset, the data acquisition pipeline, and the evaluation framework, the authors aim to lower the barrier for developing multilingual benchmarks for underrepresented communities. This work highlights that simply scaling models or using translation-based prompting is insufficient to bridge the reasoning gap, suggesting that future research must focus on better instruction-following and more robust multilingual training strategies.
Mathematical reasoning has become a central task for evaluating and tuning reasoning Large Language Models (LLMs), yet existing benchmarks remain heavily biased toward high-resource languages, with English and Chinese dominating both pre-training corpora and evaluation suites. The recently released PolyMath (Wang et al., 2025) dataset represents a significant step forward, yet its coverage is still limited to 18 only high-resource languages. To address this gap, we introduce PluraMath, an extension of PolyMath to 18 additional {underrepresented languages spanning 6 language families -- ranging from mid-resource to extreme low-resource settings. We constructed the dataset through a human-curated pipeline, where native speakers thoroughly validated pre-computed translations. Using PluraMath, we then benchmark 27 reasoning LLMs across four model scales -- small, mid-size, large, and closed-source ensembles -- probing the multilingual mathematical reasoning capabilities of state-of-the-art models under diverse linguistic conditions. Our fine-grained analysis confirms a persistent gap in mathematical reasoning performance between high-resource and underrepresented languages, with stronger results largely associated with better instruction-following ability. We fully open-source our dataset, data acquisition pipeline, and evaluation framework, with the goal of lowering the barrier to multilingual benchmark development for underrepresented communities.
Sam: That's the interesting part to me. If the translations were verified and accurate, why were the models still struggling?
Alex: That's exactly what the study digs into. You might assume the issue is translation quality—that the AI is tripping over unfamiliar words or awkward phrasing. But the evidence points somewhere else. The researchers found that how well a model performs in a given language is only weakly connected to how well it can translate text in that language. What it's much more strongly connected to is something called general instruction-following ability—how well the model understands and responds to directions overall.
Sam: So it's less like the model is confused by the foreign words, and more like... it loses its footing when the instructions aren't in the language it was primarily trained on?
Alex: That's a good way to put it. Think of it like this: imagine you're a strong chess player, but you learned all the rules and strategies from a coach who only ever spoke to you in English. If someone suddenly gives you instructions mid-game in a language you barely know, you might still remember how the pieces move—but you'd struggle to follow the game plan. The underlying skill is there, but accessing it becomes harder when the instructions don't match your primary training.
Sam: That's a meaningful distinction. It suggests that just translating more training data isn't necessarily the whole solution—the models need to learn to follow instructions fluently across languages, not just recognize words in them.
Alex: That's what the study implies, yes. And it points to a real challenge for making AI tools genuinely useful on a global scale. The researchers also note that for languages with extremely limited written resources online, even this approach has limits—fully manual translation would be required, which is significantly more costly and time-consuming.
Sam: And I imagine there's also a question of whether the math problems themselves are equally relevant across different educational systems and cultures.
Alex: The study raises exactly that concern. Mathematical curricula aren't identical around the world, and a problem that feels standard in one context might be unfamiliar or framed differently in another. The researchers acknowledge this as an open question.
Sam: So PLURAMATH is less a final answer and more a careful first step—a way of making the problem visible so it can actually be addressed.
Alex: That's a fair characterisation. By building a verified, multilingual benchmark, the study gives researchers a clearer picture of where these gaps exist and what's actually driving them. That kind of diagnostic work is necessary before meaningful improvements can be made. Thanks for listening to ResearchPod.