Edhivo: At Stony Brook, two LLMs share the writing of assessments and the correction of open-ended responses for a semester

A model generates multiple choice questions and open questions from the course materials, a second corrects according to a grid, and the teacher keeps control of the validation.

Evaluating students often is costly in terms of time: questions have to be written and then open-ended answers must be corrected, which delays feedback. Edhivo, a platform designed by Moraish Kapoor and Jiawei Zhou, tackles the two bottlenecks with a two-model language architecture. The first creates multiple-choice questions and open-ended questions from the documents provided by the teacher; the second scores students' answers according to a multi-criteria grid and produces almost immediate qualitative feedback, at several levels of detail.

The teacher keeps the last word

All interactions with AI are anchored in the course materials to limit hallucinations, and a human proofreading step preserves the teacher's authority. The platform also integrates content distribution and dashboards.

It was rolled out in the spring semester of 2026 in a real course at Stony Brook University, with 36 students. According to the authors, it made it possible to increase the number of assessments while significantly reducing the effort of writing and the time to return, paving the way for pedagogical formats centered on active learning. The study remains a small-scale pilot, without a control group.