Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Notes

TLDR; RL actually worsen Pass@K metrics. Base model already has the reasoning path that RL-ed models have.