Reinforcement Learning with Verifiable Rewards Makes Models Faster, Not Smarter

RLVR trains reasoning models with programmatic verifiers instead of human labels. Recent research suggests most gains come from search compression rather than new capabilities. What actually works.