From many research paper i skimmed ( without llm, from youtube videos though :p) i understand its the friction of not understanding and then getting +/- feedback faster with lots of iteration helps us learn fast. But for knowledge work it seems very hard as there isnt any verifier that can give me the mental mode that will evolve/extend my current mental model.
One feedback mechanism can be testing myself on the topic but software engineering related tasks are pretty hard to do without expermiemnt and experience.