Elamur Research⌕
☰
home / papers / 2026 / Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe

Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe

Paper · 2026 · Yaxuan Li, Yuxin Zuo, Bingxiang He et al.

Summary

On-policy distillation (OPD) is now common in LLM post-training (Qwen3, MiMo, GLM-5). In OPD, a student generates its own rollouts and receives dense per-token feedback from the teacher's log-probabilities. Why OPD sometimes fails has been poorly understood. In one puzzling case, a stronger teacher leaves the student unchanged while a weaker one improves it.

The authors analyze OPD on mathematical reasoning using:

  • DeepSeek-Distill models (DS-1.5B, DS-7B)
  • their RL-tuned descendants (JustRL-1.5B, Skywork-OR1-Math-7B)
  • Qwen3 models

They find two conditions for success. First, teacher and student must reason in compatible ways. Second, the teacher must bring abilities the student has not already absorbed; a higher benchmark score alone is not enough. Weak-to-strong reverse distillation experiments support both conditions.

At the token level, successful runs show the two models' top-k candidate tokens converging on states the student visits. The paper proposes two fixes for failing setups. It also shows that the dense reward grows less reliable deeper into long trajectories. These results matter for teacher selection and for extending OPD to long-horizon or agentic tasks.

Key points

  • Same-family DS-1.5B and DS-7B teachers look nearly identical from the student's viewpoint. Higher accuracy therefore does not guarantee useful signal for OPD.
  • In effective runs, the top-k overlap ratio climbs from about 72% to 91%. The shared tokens hold 97–99% of probability mass, and supervising only these tokens matches full top-k OPD.
  • Two remedies for failing runs:
    • Off-policy cold start: SFT on teacher rollouts before OPD.
    • Teacher-aligned prompt selection: drawing prompts from the teacher's post-training data. This lowers student entropy, so these prompts must be mixed with out-of-distribution ones.
  • Reward quality drops with trajectory depth, and instability starts at late tokens. Even failing teachers give rewards that correlate with correctness, which points to flat local optimization geometry rather than a bad signal.
  • All experiments use math benchmarks only. Code is available at github.com/thunlp/OPD.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Tsinghua UniversityBeijing · China439 works · 94,783 citations🎓University of Illinois Urbana-ChampaignUrbana · United States228 works · 74,591 citations🎓Renmin University of ChinaBeijing · China140 works · 17,810 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 19,503 worksArtificial Intelligence · 18,805 worksArts and Humanities · 17,959 worksLanguage and Linguistics · 17,789 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 4,145 worksEfficient models & compression · 328 works

Cite

Cite as Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu and Ning Ding. 2026. Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe. arXiv:2604.13016.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2604.13016 · arXiv 2604.13016

@misc{li2026rethinking,
  title = {{Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe}},
  author = {Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu and Ning Ding},
  year = {2026},
  doi = {10.48550/arxiv.2604.13016},
  eprint = {2604.13016},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2604.13016}
}

References

47 references, 32 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. In The twelfth international conference on learning representations, 2024.
  2. Mislav Balunović, Jasper Dekoninck, Ivo Petrov, Nikola Jovanović, and Martin Vechev. Matharena: Evaluating llms on uncontaminated math competitions. arXiv preprint arXiv:2505.23281, 2025.
  3. Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. find
  4. Dan Busbridge, Amitis Shidani, Floris Weers, Jason Ramapuram, Etai Littwin, and Russ Webb. Distillation scaling laws. arXiv preprint arXiv:2502.08606, 2025.
  5. Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4794–4802, 2019. find
  6. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan… find
  7. Ken Ding. Hdpo: Hybrid distillation policy optimization via privileged self-distillation. arXiv preprint arXiv:2603.23871, 2026.
  8. Yuqian Fu, Haohuan Huang, Kaiwen Jiang, Yuanheng Zhu, and Dongbin Zhao. Revisiting on-policy distillation: Empirical failure modes and simple fixes. arXiv preprint arXiv:2603.25562, 2026.
  9. Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. arXiv preprint arXiv:2306.08543, 2023.
  10. Etash Guha, Ryan Marten, Sedrick Keh, Negin Raoof, Georgios Smyrnis, Hritik Bansal, Marianna Nezhurina, Jean Mercat, Trung Vu, Zayne Sprague, et al. Openthoughts: Data recipes for reasoning models. arXiv preprint arXiv:2506.04178, 2025.
  11. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633–638, 2025. find
  12. Bingxiang He, Zekai Qu, Zeyuan Liu, Yinghao Chen, Yuxin Zuo, Cheng Qian, Kaiyan Zhang, Weize Chen, Chaojun Xiao, Ganqu Cui, et al. Justrl: Scaling a 1.5 b llm with a simple rl recipe. arXiv preprint arXiv:2512.16649, 2025a. — [arXiv:2512.16649](https://arxiv.org/abs/2512.16649) find
  13. Bingxiang He, Yuxin Zuo, Zeyuan Liu, Shangziqi Zhao, Zixuan Fu, Junlin Yang, Cheng Qian, Kaiyan Zhang, Yuchen Fan, Ganqu Cui, et al. How far can unsupervised rlvr scale llm training? arXiv preprint arXiv:2603.08660, 2026. — [arXiv:2603.08660](https://arxiv.org/abs/2603.08660) find
  14. Jujie He, Jiacai Liu, Chris Yuhao Liu, Rui Yan, Chaojie Wang, Peng Cheng, Xiaoyu Zhang, Fuxiang Zhang, Jiacheng Xu, Wei Shen, et al. Skywork open reasoner 1 technical report. arXiv preprint arXiv:2505.22312, 2025b.
  15. Zhiwei He, Tian Liang, Jiahao Xu, Qiuzhi Liu, Xingyu Chen, Yue Wang, Linfeng Song, Dian Yu, Zhenwen Liang, Wenxuan Wang, et al. Deepmath-103k: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning. arXiv preprint arXiv:2504.11456, 2025c.
  16. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
  17. Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, et al. Reinforcement learning via self-distillation. arXiv preprint arXiv:2601.20802, 2026.
  18. Ijun Jang, Jewon Yeom, Juan Yeo, Hyunggu Lim, and Taesup Kim. Stable on-policy distillation through adaptive target reformulation. arXiv preprint arXiv:2601.07155, 2026. find
  19. Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In Findings of the association for computational linguistics: EMNLP 2020, pages 4163–4174, 2020.
  20. Woogyeol Jin, Taywon Min, Yongjin Yang, Swanand Ravindra Kadhe, Yi Zhou, Dennis Wei, Nathalie Baracaldo, and Kimin Lee. Entropy-aware on-policy distillation of language models. arXiv preprint arXiv:2603.07079, 2026.
  21. Jeonghye Kim, Xufang Luo, Minbeom Kim, Sangmook Lee, Dohyung Kim, Jiwon Jeon, Dongsheng Li, and Yuqing Yang. Why does self-distillation (sometimes) degrade the reasoning capability of llms? arXiv preprint arXiv:2603.24472, 2026. — [arXiv:2603.24472](https://arxiv.org/abs/2603.24472) find
  22. Yoon Kim and Alexander M Rush. Sequence-level knowledge distillation. In Proceedings of the 2016 conference on empirical methods in natural language processing, pages 1317–1327, 2016.
  23. Jongwoo Ko, Sara Abdali, Young Jin Kim, Tianyi Chen, and Pashmina Cameron. Scaling reasoning efficiently via relaxed on-policy distillation. arXiv preprint arXiv:2603.11137, 2026.
  24. Gengsheng Li, Tianyu Yang, Junfeng Fang, Mingyang Song, Mao Zheng, Haiyun Guo, Dan Zhang, Jinqiao Wang, and Tat-Seng Chua. Unifying group-relative and self-distillation policy optimization via sample routing. arXiv preprint arXiv:2604.02288, 2026. — [arXiv:2604.02288](https://arxiv.org/abs/2604.02288) find
  25. Jia Li, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Huang, Kashif Rasul, Longhui Yu, Albert Q Jiang, Ziju Shen, et al. Numinamath: The largest public dataset in ai4maths with 860k pairs of competition math problems and solutions. Hugging Face repository, 13:9, 2024. find
  26. Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubramanian, and Radha Poovendran. Small models struggle to learn from strong reasoners. In Findings of the Association for Computational Linguistics: ACL 2025, pages 25366–25394, 2025.
  27. Kevin Lu and Thinking Machines Lab. On-policy distillation. Thinking Machines Lab: Connectionism, 2025. doi: 10.64434/tml.20251026. https://thinkingmachines.ai/blog/on-policy-distillation. — [doi:10.64434/tml.20251026](https://doi.org/10.64434/tml.20251026) ↗
  28. Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 5191–5198, 2020.
  29. Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019. URL http://arxiv.org/abs/1908.10084 . ↗
  30. Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He, Zhipeng Wang, and Jiachen Sun. Crisp: Compressed reasoning via iterative self-policy distillation, 2026. URL https://arxiv.org/abs/2603.054 33 . ↗
  31. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.
  32. Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2021.
  33. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.
  34. Idan Shenfeld, Mehul Damani, Jonas Hübotter, and Pulkit Agrawal. Self-distillation enables continual learning. arXiv preprint arXiv:2601.19897, 2026.
  35. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in neural information processing systems, 33:5776–5788, 2020. find
  36. Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International Conference on Learning Representations, 2021.
  37. Bangjun Xiao, Bingquan Xia, Bo Yang, Bofei Gao, Bowen Shen, Chen Zhang, Chenhong He, Chiheng Lou, Fuli Luo, Gang Wang, et al. Mimo-v2-flash technical report. arXiv preprint arXiv:2601.02780, 2026.
  38. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.
  39. Chenxu Yang, Chuanyu Qin, Qingyi Si, Minghui Chen, Naibin Gu, Dingyu Yao, Zheng Lin, Weiping Wang, Jiaqi Wang, and Nan Duan. Self-distilled rlvr. arXiv preprint arXiv:2604.03128, 2026a.
  40. Wenkai Yang, Weijie Liu, Ruobing Xie, Kai Yang, Saiyong Yang, and Yankai Lin. Learning beyond teacher: Generalized on-policy distillation with reward extrapolation. arXiv preprint arXiv:2602.12125, 2026b. — [arXiv:2602.12125](https://arxiv.org/abs/2602.12125) find
  41. Tianzhu Ye, Li Dong, Qingxiu Dong, Xun Wu, Shaohan Huang, and Furu Wei. Online experiential learning for language models. arXiv preprint arXiv:2603.16856, 2026a. — [arXiv:2603.16856](https://arxiv.org/abs/2603.16856) find
  42. Tianzhu Ye, Li Dong, Xun Wu, Shaohan Huang, and Furu Wei. On-policy context distillation for language models. arXiv preprint arXiv:2602.12275, 2026b.
  43. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476, 2025.
  44. Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chengxing Xie, Cunxiang Wang, et al. Glm-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763, 2026.
  45. Guoliang Zhao, Ruobing Xie, An Wang, Shuaipeng Li, Huaibing Xie, and Xingwu Sun. Self-distillation for multi-token prediction, 2026a. URL https://arxiv.org/abs/2603.23911 . — [arXiv:2603.23911](https://arxiv.org/abs/2603.23911) ↗
  46. Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models. arXiv preprint arXiv:2601.18734, 2026b.
  47. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangko…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/li-2026-rethinking-policy-distillation-large-language/