Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
RL with verifiable rewards can make correct reasoning easier to sample without expanding the set of problems a model can solve. Follow the pass@k crossover, base-only coverage, training checkpoints, and the limits of that negative result.
Reading focus: Why pass@1 measures early success while high-k pass@k probes whether any sampled correct route remains. How matched base and RLVR models can reverse order as the sample budget grows, including base-only solved problems. What one training trajectory and the entropy controls establish, and why math benchmarks, finite samples, and paired-model availability bound the claim.
NeurIPS 2025 Runner-Up. Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song, and Gao Huang. 13 min read, medium difficulty.