Can you teach an AI tutor to know the subject, follow the curriculum, spot mistakes, and teach well, all at once?

This explores whether training a tutoring model on four capabilities at once helps or forces tradeoffs. The corpus has one direct source on this, and it reports gains, not costs. Can educational models do more than just answer questions correctly? describes OmniEdu, which organizes K–12 model training around subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action instead of around data source or task type. It reports consistent improvements across model sizes and competitive results on tutoring benchmarks. Its central claim is that answer accuracy alone doesn't make a model educationally useful. The summary I have doesn't mention any capability getting worse, but it doesn't rule that out either.

The tradeoff question gets sharper when you look at what happens elsewhere when one model learns different kinds of skills. Does training order reshape how models handle different task types? finds that structured domains push a model's outputs toward certainty, while creative, open-ended domains push toward variety. Training on them in the wrong order lets the first damage the second. Ordering the tasks (structured first) beat mixing everything together by 6.2%. Subject competence and curriculum grounding look like the structured kind, and knowing what move to make with a confused student looks open-ended. That mapping is my inference, not a finding from either paper. It suggests the four capabilities can pull against each other, and that how you sequence them may matter as much as which ones you include.

Two other notes suggest what the tradeoffs would look like and how to get around them. Can student simulators match both behavior and learn from teaching? is about simulating students, not tutoring. But it shows the same problem in a smaller form. Models that track student state faithfully ignore tutor corrections, and prompted role-play follows guidance but loses the individual student. Only a two-stage design (pooled training, then per-student specialization) delivered both. So balancing two goals is possible but usually needs deliberate architecture. Can models dynamically activate expert skills at inference time? offers another route. Skills stored as separate expert vectors can be mixed at inference without interfering with each other, which sidesteps the training-time balancing problem.

One warning comes from a neighboring result. Does richer teacher context hurt student generalization? shows that teacher signals which make a student more confident and concise also make it worse on unfamiliar problems, because it stops expressing uncertainty. A tutor tuned to sound sure of itself on curriculum material could lose the caution that diagnosing an unfamiliar misconception requires. Whether OmniEdu's four-way split avoids this isn't something the corpus can say.

The corpus supports this much: organizing training around pedagogical capabilities improves tutoring benchmarks. It does not directly measure whether one capability trades against another. The neighboring work suggests where to look, namely training order, confidence style, and whether skills are trained jointly or composed.