New preprint argues that making models better at controlling their reasoning can make chain-of-thought monitors less reliable · Polylog