Why reasoning models will generalize