Two accounts of how reinforcement learning with verifiable rewards (RLVR) interacts with an imperfect verifier coexist: in one, reward hacking is inherited f…
机构:CMU
来源:arXiv 2611.13218 | AI4Papers 论文推荐平台