Does RL Really Incentivize Reasoning Beyond the Base Model?

LeapLab's defining "limits of RLVR" result (NeurIPS 2025 best-paper runner-up, ~900 citations): pass@k analysis shows RLVR mostly sharpens distributions over capabilities already present in the base model rather than creating new reasoning capacity — the paper every subsequent RL-for-reasoning study has to answer. With Shanghai AI Lab co-authors; Tsinghua-led.