Comment by jdub

Reinforcement learning for "reasoning" perturbs the model to generate completions in a particular chain of thought / alternative selection structure. It's three next token predictors in a trench coat.