After an hour of GRE preparation, students taught by AI improved their test scores about as much as those taught by expert human tutors, researchers report in StudentBench, a study posted to arXiv. Participants took the final test independently, immediately after the session. The GRE is an exam used in graduate admissions.
The trial involved 2,383 adults, randomly assigned to AI tutoring, live video sessions with human tutors, or a control group watching educational videos unrelated to the GRE. Handshake AI funded the research and recruited participants through its platform.
Both kinds of tutor received students' initial test mistakes to guide their teaching, but neither saw the final test questions. The AI tutors planned lessons and generated practice problems.
The researchers judged the overall improvements comparable after accounting for differences in starting ability and test version. The same held for math alone, but they could not establish a match in verbal reasoning. Both tutoring groups improved more than participants without tutoring. The math result depended on those adjustments: without them, the analysis did not establish a match.
A different AI tutor led each of the seven subject areas. The best AI result exceeded the human average in three of four math areas, while human tutors remained close to the leading models in verbal reasoning.
The cheapest model, Gemma 4 31B, cost about seven cents per session in model usage. The authors compared AI costs with an assumed human tutoring rate of $75 an hour.
The paper remains a preprint, and the experiment did not test whether students retained the gains or improved their actual GRE scores.