Distilling long-chain reasoning trajectories from large teacher models provides an effective way to improve the reasoning ability of small language models. H…
机构:上智院
来源:arXiv 2610.13285 | AI4Papers 论文推荐平台