A new paper on arXiv examines how well a large language model, GPT-5.1, can generate feedback on handwritten university mathematics work. The researchers fed the model students' handwritten answers along with their grades, then had it produce both written feedback and practice exercises. They then evaluated the quality of that feedback and tested whether students who received it performed better on later exams.
The study sits at the intersection of AI-assisted education and assessment, asking whether automated feedback can be both accurate and pedagogically useful. Because the input is handwritten, the work also tests the model's ability to process non-digital student work at scale.
Since this is a single preprint, the results should be treated as early evidence. The abstract does not include effect sizes or comparison groups, so the practical impact remains unclear until full details and peer review are available.