In verifier-style RLVR, group-relative optimization often treats advantage scale as an implementation detail. This paper separates two low-variance cases: su…
机构:阿里
来源:arXiv 2609.19164 | AI4Papers 论文推荐平台