Contrary to many recent claims, we show that RL actually learns new skills and generalizes surprisingly well in LLMs. Check out our new preprint.