Autoregressive (AR) image models make pictures by choosing tokens one-by-one, but they were judged only on picking likely tokens, not on how good the final picture looks in pixels.
This paper teaches AI models to reason better by first copying only good examples and later learning from mistakes too.