reinforcement learning

Reinforcement learning is a way of training a model by trial, error and reward. Instead of showing it thousands of correct answers to copy, you let it attempt something, score the attempt, and nudge it toward whatever earned a higher score. It is roughly how you would teach a dog a trick, except the dog is a few hundred billion numbers and the treat is a mathematical signal. Most of the recent jump in reasoning model quality came from this, not from bigger models.

The catch is that it needs a reliable way to tell good from bad. Code either compiles or it does not, a maths proof either checks out or it does not, so those areas improved fast. For work where nobody can automatically score the result, like writing a tactful email or judging a business decision, the method has much less to grip on. That is one reason models feel uneven: they were trained hard where scoring was easy.