I think they are overfitting more, I'm seeing it perform worse on esoteric logic puzzles.