arXiv:2607. 10805v1 Announce Type: cross Abstract: On-Policy Self-Distillation (OPSD) has emerged as a crucial paradigm for enhancing and aligning Large Language Models (LLMs).
By Keqin Peng, Chen Li, Yuanxin Ouyang, Yancheng Yuan, Liang Ding
arXiv:2605.10889v2 Announce Type: replace-cross
Abstract: On-policy distillation offers dense, per-token supervision for training reasoning models; however, it remains unclear under which conditions...
By Mohammadreza Armandpour, Fatih Ilhan, David Harrison, Ajay Jaiswal, Duc N. M Hoang, Fartash Faghri, Yizhe Zhang, Minsik Cho, Mehrdad Farajtabar
The paper introduces INSPIRE, an Internalize-Then-Improve framework designed to enhance example-driven mathematical reasoning in large language models. It combines Reference-Guided Student Internalization (RGSI) to generate high-quality preference pairs with a staged rubric preference training that separates method learning from correctness. Experiments across various model sizes show consistent gains, even outperforming larger open-source models, and maintain performance on out-of-distribution mathematical tasks.
By Shuai Wang, Jiayi Kuang, Yinghui Li, Haojing Huang, Xinnian Liang, Ying Shen, Liang Lin
EOPSA (Efficient On-Policy Self-Distilled Safety Alignment) addresses inefficiencies in On-Policy Self-Distillation (OPSD) for safety alignment by focusing training on safety-critical tokens. It introduces Adaptive Rollout Scheduling, which limits generation length based on a Teacher Rescue Rate metric, and Selective Distillation, which filters out safety-neutral tokens to concentrate gradient updates on safety-pivotal transitions. Experiments on models up to 32B parameters show that EOPSA reduces rollout computation by about 50% and backpropagates through only roughly 2% of tokens, outperforming full-token distillation baselines in safety compliance and reasoning retention.
The paper investigates data efficiency and selection in On‑Policy Distillation (OPD) for large language models. It shows that 1‑shot OPD—training on a single example—consistently improves performance, especially when the example is hard, and that longer chain‑of‑thought (CoT) paths drive the gains rather than token entropy. Based on these findings, the authors propose a simple hard‑example selection strategy that, using only eight carefully chosen hard examples, matches the performance of a 17,000‑example baseline across models from 1.5B to 7B parameters.
By Zhinan Hou, Jiaqi Zhang, Xunliang Cai, Keyou You
arXiv:2604. 10688v2 Announce Type: replace-cross Abstract: On-policy reinforcement learning has become the dominant paradigm for reasoning alignment in large language models, yet its sparse, outcome-level rewards make token-level credit assignment notoriously difficult.
By Binbin Zheng, Xing Ma, Yiheng Liang, Jingqing Ruan, Xiaoliang Fu, Kepeng Lin, Benchang Zhu, Ke Zeng, Xunliang Cai