Today's best AI can create test versions that test the same skill differently.

Schools want to stop cheating without spying on students through webcams. To do this, they want to give every student a different test question. But this only works if the questions look completely different on the surface while still testing the exact same underlying skill at the same difficulty. The researchers created a benchmark called VARIA to test whether modern large language models can actually make these equivalent test versions. They tested top artificial intelligence models across different prompting methods. They found that top models performed well together, achieving composite integrity scores between 0.81 and 0.88. Older, smaller models scored much lower, between 0.50 and 0.55. The researchers also found clear trade-offs. Asking models to think step by step or giving them no examples best preserved the exact skill being tested. Providing example questions or using strict rules created the most visual variety, but these two styles traded off on readability matching.

Why it matters

If we replace surveillance-based proctoring with personalized tests, each student must get a unique version that still measures the same skill. This paper shows that today's best AI models can do this reliably, while weaker models fail. That gap determines whether AI-generated assessments can be trusted in real classrooms and certifications.

Questions answered