Reinforcement learning (RL) post-training improves the reasoning capabilities of large language models, but its cost is dominated by autoregressive rollout g…
机构：UCSD
来源：arXiv 2610.32881 | AI4Papers 论文推荐平台