Reproducing specific chunks of long form text from distilled (inherently lossy) model data is not something that I would expect LLMs to be good at.

And of course, there's no actual reasoning or logic going on, so they cannot compete in this context with a curious 12 year old, either.