“LLMs are currently poor at reframing a user's question to address an underlying misconception, a key skill of effective human teachers.”