MathVision-Wild: Photographic Math Problems for Real-World Visual Reasoning

MathVision-Wild provides 1,000 to 10,000 photographic versions of the MathVision test dataset captured in diverse physical environments. Created by MathLLMs and updated in late 2025, it transitions digital math problems into real-world visual contexts to evaluate Vision Language Model (VLM) performance.

Use Cases

• Benchmarking Vision Language Models on mathematical reasoning using photographic inputs
• Training image-to-text models to interpret math problems in physical environments
• Evaluating robustness of VQA systems against environmental noise and lighting variations

Strengths

• 1,000 to 10,000 records
• Real-world physical environment variations including lighting and backgrounds
• Apache 2.0 open license
• Direct alignment with the established MathVision-testmini benchmark

Limitations

• Small sample size relative to general-purpose VQA datasets
• Limited to the testmini subset scope
• Potential for label noise inherent in photographic captures of digital content

Provenance

Source

MathLLMs / MathVision benchmark

Collection Method

Photographing existing MathVision-testmini data in various physical environments

Freshness

Last updated December 2025.