Introducing new capabilities to frontier models has long been the goal of posttraining, which predominantly employs supervised finetuning (SFT) and reinforce…
机构:哈佛
来源:arXiv 2610.02140 | AI4Papers 论文推荐平台