Every recurrence of a looped language model adds cost in training, decoding, prefill, and reinforcement learning (RL). The closer recurrent states get to fix…
机构:CMU
来源:arXiv 2610.06833 | AI4Papers 论文推荐平台