Researchers introduce ScratchMath benchmark to help MLLMs diagnose student errors in handwritten mathematics rather than just finding correct answers.

New benchmark dataset contains 1,720 handwritten math samples from Chinese primary and middle school students

Addresses gap where current multimodal large language models focus on generating correct answers rather than explaining student errors

Introduces two key tasks: Error Cause Explanation (ECE) and Error Cause Classification (ECC) with seven error categories

Designed to handle challenges of authentic handwritten work including diverse handwriting styles, complex layouts, and varied problem-solving approaches

Aims to enable personalized educational feedback by helping AI systems understand the reasoning behind student mistakes