GRPO shares a response-level advantage across tokens, leaving unresolved how strongly each token should contribute to learning. Updates induced by the shared…
机构:南大
来源:arXiv 2610.32772 | AI4Papers 论文推荐平台