StudentBench compares 2,383 participants across the categories of model-based tutoring, human tutoring and no tutoring, over the course of a one-hour session.
Filed on September 23, 2026 and funded by Handshake AI Research, the StudentBench study is based on a public platform that has collected more than 175,000 messages between students and models. The protocol compares the progress of 2,383 participants on quantitative and verbal GRE questions, depending on whether they receive language model tutoring, an expert human tutor, or no coaching.
The learning gains obtained with AI are statistically equivalent to those of expert human tutoring, and in five of the seven GRE domains the best artificial tutor outperforms humans on average. The cost gap is of a different order of magnitude: $0.0052 versus $4.81 per percentage point earned, or a ratio of 918. A second study compared model generated lesson plans and exercises to experienced tutors, over the course of 2,028 pairwise assessments.
The scope of the result deserves to be delineated. It is a one-hour session of preparation for a standardized exam, not a long-term accompaniment, and the human tutors retain the advantage over verbal reasoning as well as over the quality marks awarded in the second study.
Read the full article on the original source
Go to the source →