Long rollout windows make reinforcement learning with verifiable rewards (RLVR) expensive, yet simply shortening them can remove the correctness signal: when…
机构:UIUC
来源:arXiv 2610.22887 | AI4Papers 论文推荐平台