“Reinforcement learning (RL) does not generalize as effectively as pre-training; specializing a model on a specific domain using RL often degrades its performance in other areas.”