If we're lucky, LLMs force people to put more effort into assignments and grading and then that would help kids learn to learn as well.
There's also the fact that kids are being taught the very basics, the sorts of things increasingly intelligent models are most likely to be able to solve first. I don't think there's any level of effort that can be put into designing assignments to get around this.
Similar to how teachers haven't really been able to do anything to stop kids from sticking their algebra problems into wolframalpha or other tools besides just making them do the work in class (which then cuts into teaching time).
Beyond a certain point, all that can be done is for teachers to try to instill the importance of practice into students, and for parents to be more proactive in monitoring how their kid is doing their homework.
Maybe we'll see an increase in after-school classes for kids to do homework in, under teacher supervision.
I'm afraid it might be exactly opposite. Having all the knowledge at hand. all the time will lead to knowledge atrophy. Just like it already happens with ability to travel without navigation.
I hope somebody figures this out but I don't know what the solution looks like.