MATH-Vision (MATH-V)

Standard multimodal math-reasoning benchmark (NeurIPS 2024): 3,040 competition-sourced mathematical problems with visual contexts, spanning 16 mathematical disciplines across 5 difficulty levels — built to fix the subject/diversity limits of MathVista. It exposed a wide gap between LMMs and humans at release and became a fixture in frontier VLM model cards (Gemini, Qwen-VL, InternVL releases all report MATH-V).

Pre-window arXiv date, listed for in-window adoption; from Hongsheng Li's MMLab-CUHK / MathLLM group.