Reinforcement learning perturbs the model such that the token prediction process (inference) tends towards the desired result.