After DeepSeek-R1, a popular story took hold. Reinforcement learning with verifiable rewards, or RLVR, lets a model practice on math and code, get a reward only when the answer checks out, and keep improving. Many people expected that this would teach models genuinely new ways of reasoning, as reinforcement learning taught game-playing agents new strategies.
A Tsinghua and Shanghai Jiao Tong team asked whether that is actually true. Their answer, from a paper that received a Runner-Up award at NeurIPS 2025, is a careful no, at least for the training methods and models they tested.
RLVR-trained models beat their base models when each gets one try. But when both get many tries, the base model solves as many problems, or more. The correct reasoning paths of the RL model were already inside the base model. RL mostly made those paths likelier to be chosen. Distillation from a stronger model did add new reasoning.
A better question: what could the model ever solve?
The usual score is accuracy on one attempt. That tells you how often a model is right at first try. It does not tell you what it is capable of with more attempts. The paper uses a metric called pass@k: give the model k tries at a problem, and count the problem as solved if any one of them is correct.
When k is 1, pass@k is ordinary accuracy. When k is large, such as 256 or 1,024, it measures the boundary of what the model can do at all. Nobody would ship a product that needs 256 tries. The point is to ask whether RL pushed that boundary outward. To avoid crediting lucky guesses, the authors checked the reasoning chains by hand for samples of math problems, and used coding tasks with unit tests.
