My experience is that LLMs can't actually do 3 at all. The intersection of knowledge has to already be in the training data. It hallucinates if the intersection of knowledge is original. That is exactly what should expect though given the architecture.
