This is key. LLMs will cheat and just look up training data whenever it can. Needs to be actively countered if one is to make it "reason".