The paper introduces a method for offline on‑policy distillation that addresses the problem of imperfect teacher supervision. By training on teacher‑successful problems and measuring changes in token likelihoods on teacher‑failed trajectories, the authors derive a learnability signal that weights the distillation loss. This approach improves performance on mathematical reasoning and code generation tasks while reducing computational cost compared to online distillation.
By Yihao Ai, Weilong Yan
Offline on-policy distillation gains efficiency by collecting student trajectories and teacher supervision once and reusing them throughout optimization. The same reuse makes imperfect supervision per...
The paper introduces Selective Supervision for Direct-OPD (S$^2$D-OPD), a refinement of Direct On-Policy Distillation that filters out states where the teacher’s policy change is minimal, as measured by the teacher‑reference Jensen‑Shannon divergence. By masking low‑divergence states and keeping only the top 10% of states per response, S$^2$D-OPD improves held‑out accuracy on AIME and HMMT benchmarks across multiple teacher‑student pairs without additional forward passes.
By Yibo Zhao, Zixuan Yang, Yunshi Lan, Xiang Li
arXiv:2609.14636v1 Announce Type: new
Abstract: On-policy distillation (OPD) has become a standard approach for transferring capabilities from large teachers to compact students. Its cost, however, i...
By Zhiyu Gui, Kexin Huang, Jia Guo, Junkang Wu, Zihao Wang, Zhiqiang Zhang, Jun Zhou, Jiancan Wu, Xiang Wang
The paper introduces Teacher-Gated On-Policy Distillation (TGOPD), a method that verifies teacher reliability at the prompt level before applying dense supervision in on-policy distillation. TGOPD uses verifier-scored teacher probes to decide whether to route a prompt to dense OPD or to a verifier-grounded alternative. Experiments on 4B and 35B models across mathematics, code, and instruction tasks show TGOPD outperforms vanilla OPD and improves teacher GPU utilization from 9.8% to 78.9% in a 4B single-domain run.
By Zhiwei Zhang, Zechen Sun, Fei Zhao, Kang Peng, Bin Liang, Huayu Deng, Yao Hu, Kam-Fai Wong, Mu Chuan
SCOPE-OPSD introduces a Fisher‑conditioned privileged subspace for on‑policy self‑distillation (OPSD) that projects the teacher‑student residual onto a frozen rank‑64 factor derived from residual covariance and language‑model‑head Fisher sensitivity. The method adds no extra rollouts or inference modules and, across multiple Qwen3 model checkpoints and trajectory lengths, consistently matches or surpasses pure OPSD and a matched random baseline, achieving significant gains in most settings. A cross‑fitted diagnostic shows a 4.40‑fold increase in captured privileged‑gap compared to the random orientation.
By Yunmeng Chen (Chongqing Ant Consumer Finance Co., Ltd), Kunyu Wang (Alibaba Cloud Computing Co., Ltd), Peihan Li (Chongqing Ant Consumer Finance Co., Ltd), Yi Wang (Chongqing Ant Consumer Finance Co., Ltd), Shuyin Xia (Chongqing University of Posts and Telecommunications), Yi Liu (Chongqing Ant Consumer Finance Co., Ltd), Xinyong Cheng (Alibaba Cloud Computing Co., Ltd), Dehui Wang (Alibaba Cloud Computing Co., Ltd), Xiangyong Zhai (Alibaba Cloud Computing Co., Ltd), Yanxing Liu (Chongqing Ant Consumer Finance Co., Ltd), Song Liu (Chongqing Ant Consumer Finance Co., Ltd)