Quick reading on GenAI, tutors, and tutorials
In a nutshell: Tutorials with human tutors help learning (duh!), GenAI tutors nudge students away from the students’ own solutions towards other ones that they for some reason prefer, and humans are still better at grading math olympiad tasks! There, saved you some time! :-)
Cen & Sonkar (2026): “MUSE: Revealing Responsiveness Gaps of LLM Tutors on Multiple-Solution Programming Problems”
This is a really interesting article for anyone toying with the idea of using LLMs to “tutor” students: “LLM tutors exhibit systematic preferences among valid solutions, often favoring brute-force approaches when not providing general advice. More concerningly, tutors often impose these preferences, redirecting students toward their own preferred approach rather than supporting the student’s.” Of course, whether this is the case depends on what LLM is used, and also if it is trained to avoid this. Cen & Sonkar (2026) find that they can moderately improve LLM behaviour, and probably this can be improved over time. But right now, this is an interesting problem with GenAI “tutors” to be aware of!
Duchi et al. (2026): “What’s Old Is New: Improved Learning from Small-Group Tutorials”
Take-home work is now always under the suspicion of having created whatever output is returned with the help of (too much) GenAI, which potentially decouples performance from learning. So what if instead of submitting their homework, students met in small groups with a tutor in weekly tutorials and discussed problems? (Which is how pretty much my whole studies worked. As Duchi et al. (2026) write: what’s old is new…) Students still need to work through the homework and submit it (mostly to pace work), but now instead of being evaluated on the basis of written work, they need to explain their reasoning on a very similar problem to the one they did as homework — “homework becomes preparation for an assessment rather than the assessment itself“. Participation, engagement and effort in the tutorials was graded (mastery measured later on in-class exams). Tutors were primed to see the tutorial as an opportunity for learning, not a quiz, and the opportunity for learning also included engagement with what their peers were doing.
Duchi et al. (2026) find in a introductory math class and a randomised trial that the tutorials have a significant positive effect on grades: They did 14 of 100 points better on the midterm, and still 9 out of 100 on the final exam than students in the normal control condition. So this seems to be a good intervention, both because it makes students focus on learning and not on producing output, and because they get extra tutoring. But of course, as always, the question is how to pay for all the tutors…
Ospanov et al. (2026): “Can LLMs Reliably Grade Olympiad Proofs? A Study of Mathematical Verification with LLMs”
No, or at least not as well as experts. And that’s it for now!
Cen, S., & Sonkar, S. (2026). MUSE: Revealing Responsiveness Gaps of LLM Tutors on Multiple-Solution Programming Problems. arXiv preprint arXiv:2610.36367.
Duchi, J. C., Hartog, W., & Sun, D. L. (2026). What’s Old Is New: Improved Learning from Small-Group Tutorials. arXiv preprint arXiv:2610.39538.
Ospanov, A., Feng, Z., Ding, D., Liu, C., Bai, H., Sun, J., … & Farnia, F. (2026). Can LLMs Reliably Grade Olympiad Proofs? A Study of Mathematical Verification with LLMs. arXiv preprint arXiv:2610.36767.
It’s actually mean to post these pictures today because the weather is beautiful, but I won’t get to that spot on time to still see the sun, soooo….
