They won't, and efficiency will be maximised by not needing to have real answers at all.
LLMs excel at outputting answer-shaped feelgood.