MathSight: A Benchmark Exploring Have Vision-Language Models Really Seen in University-Level Mathematical Reasoning?
This addresses the need for better evaluation of genuine vision-grounded reasoning in AI models, though it is incremental as it focuses on benchmarking rather than new methods.
The paper tackles the problem of unclear visual contribution in vision-language models for mathematical reasoning by introducing MathSight, a benchmark that isolates visual input through variants like hand-drawn and photo-captured images, finding that visual contribution diminishes with problem difficulty and text-only models can outperform multimodal ones.
Recent advances in Vision-Language Models (VLMs) have achieved impressive progress in multimodal mathematical reasoning. Yet, how much visual information truly contributes to reasoning remains unclear. Existing benchmarks report strong overall performance but seldom isolate the role of the image modality, leaving open whether VLMs genuinely leverage visual understanding or merely depend on linguistic priors. To address this, we present MathSight, a university-level multimodal mathematical reasoning benchmark designed to disentangle and quantify the effect of visual input. Each problem includes multiple visual variants -- original, hand-drawn, photo-captured -- and a text-only condition for controlled comparison. Experiments on state-of-the-art VLMs reveal a consistent trend: the contribution of visual information diminishes with increasing problem difficulty. Remarkably, Qwen3-VL without any image input surpasses both its multimodal variants and GPT-5, underscoring the need for benchmarks like MathSight to advance genuine vision-grounded reasoning in future models.