The Invisible Leash: Why RLVR May or May Not Escape Its Origin
Reinforcement Learning with Verifiable Rewards (RLVR) is investigated to see if it expands a model's reasoning or just amplifies known high-reward outputs. The study finds RLVR primarily preserves existing solutions rather than expanding coverage, creating an 'invisible leash' that restricts discovery to patterns the base model already knows. This suggests current RLVR optimizes for precision at the expense of finding new reasoning paths.