Policy optimization is central to language model post-training, yet principled guidance remains limited. We revisit its theoretical foundations to identify a…
机构:Microsoft
来源:arXiv 2611.14467 | AI4Papers 论文推荐平台