It's not clear to me that the generalization beyond languages is valid. In extensive reading the unknown words are connected to more known words; that seems to me likely the largest factor in why it works better. But the other domains like math problems don't have this context effect, because the problems are independent. Why would it still work?