Automated Grading of Handwritten Mathematics Using Vision-Capable LLMs

This collection of recent arXiv papers covers security- and governance-adjacent research on AI systems, datasets, and architectures. Highlights include: an empirical evaluation of vision-capable LLM graders for handwritten mathematics showing high rubric-level accuracy but most errors (up to 87% in the best model) stem from image transcription failures, hallucinations, and equivalent-expression handling; a study of synthetic-media disclosure design that identifies tensions (normativity vs. neutrality, proactivity vs. precision) and the use of analogies (e.g., nutrition labels) in policy design