Comment by adgjlsfhk1

the presumption is that these results are a side-effect of reinforcement learning to make the models smarter.