Comment by irthomasthomas

One of the things that came out of the decoded reasoning paper was that Claude models had memorized answers to tests but hid this memorization from the user output and pretended to derive the answer properly. It's only possible to cheat so blatantly in closed models where the reasoning is hidden.