The paper introduces CROP, a method for selective on‑policy distillation that prioritizes token‑level supervision based on task relevance. CROP uses paraphrase‑calibrated counterfactual sensitivity to measure how much each response token depends on the semantic content of the input, constructing validated original‑paraphrase‑counterfactual triplets for each prompt. Experiments in two teacher‑student settings show that CROP outperforms other selectors, improving aggregate performance by 1.92 and 2.96 points.
By Enhan Li, Junhao He, Hongyang Du
arXiv:2605. 03677v2 Announce Type: replace Abstract: On-policy distillation (OPD) has recently emerged as an effective post-training paradigm for consolidating the capabilities of specialized expert models into a single student model.
By Wenjin Hou, Shangpin Peng, Weinong Wang, Zheng Ruan, Yue Zhang, Zhenglin Zhou, Mingqi Gao, Yifei Chen, Kaiqi Wang, Hongming Yang, Chengquan Zhang, Zhuotao Tian, Han Hu, Yi Yang, Fei Wu, Hehe Fan
On-policy distillation (OPD) supervises a student language model on trajectories sampled from its current policy, but assigns equal credit to response tokens with unequal supervision value. Selective OPD addresses this limitation by allocating supervision non-uniformly across response tokens according to their estimated training value.
arXiv:2605. 07804v3 Announce Type: replace-cross Abstract: On-policy distillation (OPD) leverages dense teacher rewards to enhance reasoning models.
By Zhicheng Yang, Zhijiang Guo, Yifan Song, Minrui Xu, Yongxin Wang, Yiwei Wang, Xiaodan Liang, Jing Tang
arXiv:2606. 00147v1 Announce Type: cross Abstract: Domain-specific supervised fine-tuning (SFT) often improves in-domain performance at the cost of degrading a model's general capabilities.
By Yuduo Li, Xiaofeng Shi, Qian Kou, Longbin Yu, Hua Zhou
OmniOPSD is an on‑policy self‑distillation framework that enhances multimodal affective reasoning by converting label‑level supervision into evidence‑rich teacher rationales. It couples student rollouts with token‑level supervision from a local self‑teacher, enabling training without extra human annotation or online model calls. Experiments on 3B and 7B models show OmniOPSD surpasses supervised fine‑tuning and GRPO, achieving top scores on MER‑UniBench and MME‑Emotion recognition.
By Zebang Cheng, Shuimu Chen, Boxue Yang, Yuanshen Guan, Jingyi Chen, Zheng Lian, Xiaojiang Peng, Fei Ma, LaiZhong Cui, Qi Tian