I am pretty sure they had solutions in their training data. LLMs only predict the next token.