Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved mathematical reasoning, yet progress on hard prompts often stalls: a prompt …
机构:上交
来源:arXiv 2611.02747 | AI4Papers 论文推荐平台