M3Kang: Evaluating Multilingual Multimodal Mathematical Reasoning in Vision-Language Models

TL;DR

The paper addresses the gap in evaluating multilingual, multimodal mathematical reasoning by introducing M3Kang, a massively multilingual benchmark derived from Kangaroo Math Competition problems translated into 108 languages and aligned with diagrams. It presents a translation-based pipeline (including backtranslation quality metrics) to create a rich multilingual multimodal dataset and provides English-only (M2Kang) as a baseline subset. A comprehensive benchmarking study across open and closed vision-language models reveals that diagram-based math remains challenging, performance scales with language presence and model size, and multilingual techniques can generalize to multimodal settings, with MTR delivering strong gains. The work includes a thorough comparison with human performance and open-sourcing of the dataset and reproducibility framework, offering a scalable resource for developing inclusive, multilingual multimodal reasoning systems.

Abstract

Despite state-of-the-art vision-language models (VLMs) have demonstrated strong reasoning capabilities, their performance in multilingual mathematical reasoning remains underexplored, particularly when compared to human performance. To bridge this gap, we introduce M3Kang, the first massively multilingual, multimodal mathematical reasoning dataset for VLMs. It is derived from the Kangaroo Math Competition, the world's largest mathematics contest, which annually engages over six million participants under the age of 18 across more than 90 countries. M3Kang includes 1,747 unique multiple-choice problems organized by grade-level difficulty, with translations into 108 culturally diverse languages, some of them including diagrams essential for solving them. Using this dataset, we conduct extensive benchmarking on both closed- and open-source SOTA models. We observe that, despite recent advances, models still struggle with basic math and diagram-based reasoning, with performance scaling with language presence and model size, but not with grade level. We also find that multilingual techniques can be effectively extended to the multimodal setting, resulting in significant improvements over baseline approaches. Our analysis also incorporates performance data from over 68,000 students, enabling direct comparison with human performance. We are open-sourcing M3Kang, including the English-only subset M2Kang, along with the framework and codebase used to construct the dataset.

Figures (19)

• Figure 1: Pipeline to create the base dataset in Catalan
• Figure 2: Pipeline to create the dataset in English
• Figure 3: Pipeline to translate dataset from language $l_1$ (source) to $l_2$ (target)
• Figure 4: Pipeline to compute backtranslation-based translation quality metrics
• Figure 5: In green, average accuracy of VLMs across culturally diverse languages sorted by their Internet presence. In blue, for each VLM, Spearman's correlation coefficient between model accuracy and Internet presence.