Comment by Veedrac

The distinction is that it's not 'predicting the next token'. Instead it's _determining_ the next token based on a prediction of its reward signal.