What Pretraining and Midtraining Make Learnable from Rewards?

A reward can identify a correct answer while leaving the computation needed for new inputs undetermined. We study how pretraining and midtraining supply the information and computation that make reward adaptation effective.