Reinforcement learning with verifiable rewards (RLVR) is known to substantially improve mathematical reasoning in language models, but what it actually adds to the base model has been unclear. A new preprint on arXiv sets out to answer that question directly.
The authors investigate through a sequence of interventions applied at generation time—that is, by changing how the model produces its output rather than by modifying its weights. This approach lets them probe which behaviors RLVR introduces while the model is already running.
The preprint's framing, reflected in its title, is that RL's contribution can be pinned to a few decisions the base model would not make on its own. As of the abstract, full results are available in the paper (arXiv:2610.24663).