Reinforcement learning for large language models often uses separate systems for inference and training. Even with identical weights and BF16 tensors, implem…
机构：字节
来源：arXiv 2610.32951 | AI4Papers 论文推荐平台