This paper teaches AI models not just how to solve problems but also how to tell when their own answers might be wrong.
Modern AI models can get very good at being correct, but in the process they often lose their ability to think in many different ways.