MathCoder-VL: Bridging Vision and Code for Enhanced Multimodal Mathematical Reasoning
This work solves the problem of improving multimodal mathematical reasoning for AI systems, particularly in handling mathematical figures, and is incremental as it builds on existing methods with novel datasets and alignment techniques.
The paper tackles the problem of multimodal mathematical reasoning by addressing the lack of detailed mathematical figure data in existing datasets, proposing a method that uses code as supervision for cross-modal alignment and creating large-scale datasets. The resulting model, MathCoder-VL, achieves state-of-the-art performance, surpassing GPT-4o and Claude 3.5 Sonnet by 8.9% and 9.2% in geometry problem-solving on MathVista.
Natural language image-caption datasets, widely used for training Large Multimodal Models, mainly focus on natural scenarios and overlook the intricate details of mathematical figures that are critical for problem-solving, hindering the advancement of current LMMs in multimodal mathematical reasoning. To this end, we propose leveraging code as supervision for cross-modal alignment, since code inherently encodes all information needed to generate corresponding figures, establishing a precise connection between the two modalities. Specifically, we co-develop our image-to-code model and dataset with model-in-the-loop approach, resulting in an image-to-code model, FigCodifier and ImgCode-8.6M dataset, the largest image-code dataset to date. Furthermore, we utilize FigCodifier to synthesize novel mathematical figures and then construct MM-MathInstruct-3M, a high-quality multimodal math instruction fine-tuning dataset. Finally, we present MathCoder-VL, trained with ImgCode-8.6M for cross-modal alignment and subsequently fine-tuned on MM-MathInstruct-3M for multimodal math problem solving. Our model achieves a new open-source SOTA across all six metrics. Notably, it surpasses GPT-4o and Claude 3.5 Sonnet in the geometry problem-solving subset of MathVista, achieving improvements of 8.9% and 9.2%. The dataset and models will be released at https://github.com/mathllm/MathCoder.