On-policy distillation (OPD) enables effective capability transfer between language models, yet the mechanisms underlying its failures are not fully understo…
机构:清华
来源:arXiv 2610.11247 | AI4Papers 论文推荐平台