Elamur Research⌕
☰
home / papers / 2026 / Distillation Traps and Guards: A Calibration Knob for LLM Distillability

Distillation Traps and Guards: A Calibration Knob for LLM Distillability

Paper · 2026 · Weixiao Zhan, Yongcheng Jing, Leszek Rutkowski et al.

Summary

Knowledge distillation (KD) from large LLMs into smaller students sometimes fails without warning, and the same mechanism lets others copy proprietary models. The authors ran a pilot study of KL estimators: token- vs sequence-level, forward vs reverse, and on- vs off-policy. They used Gemma-3 and Qwen-3 teacher–student pairs on math, knowledge QA and open-ended generation. The study uncovered a number of what the authors call "distillation traps": noise in the distribution's tail, instability under off-policy sampling, and, above all, the gap between teacher and student. In students these traps show up as overconfident hallucination, collapse of self-correction, and local decoding degradation.

Building on this, they propose a post-hoc calibration of the teacher via reinforcement fine-tuning (RFT). The reward combines task utility, a KL anchor and a cross-tokenizer calibration term. A single coefficient η pushes the teacher toward being either easier or harder to distill. Across seven tasks, students of "distillable" teachers beat SFT and KD baselines. "Undistillable" teachers keep their own accuracy, but students distilled from them collapse. This gives one knob for both better KD and protection of model IP.

Key points

  • Pilot study setup:
    • Model pairs were a Gemma-3-4B student with 12B/27B teachers, and a Qwen3-1.7B student with 8B/14B/32B teachers.
    • Each prompt got 16 rollouts at temperature 1.0 with no top-k/top-p truncation.
    • Benchmarks included CSQA, MMLU-Pro, superGPQA, Dolly and Vicuna.
  • Teachers often prefer wrong answers:
    • On some tasks, teachers assign lower sequence RKL to wrong-answer traces more than half the time (up to 65.6% on superGPQA).
    • Off-policy sequence RKL estimates swing across many orders of magnitude and can even be negative.
    • The K3 trick only partly stabilizes these estimates.
  • Undistillable teachers keep accuracy but break students:
    • The Qwen3-8B undistillable teacher scores 0.831 on out-of-distribution CSQA.
    • Its misled Qwen3-1.7B student drops to 0.065.
  • Calibration directly moves the trap metrics:
    • For Qwen3-8B on CSQA, distillable calibration lowers token RKL (0.762→0.619) and wrong-trace preference (36.7%→27.8%).
    • Undistillable calibration raises wrong-trace preference to 48.7%.
  • Failure modes in misled students: these were classified with help from Claude Sonnet 4.5 as an LLM judge.
    • Extrinsic hallucination.
    • Intrinsic hallucination: self-correction collapse, where students flip from correct intermediate answers to wrong conclusions. Teachers given the same prefix stay correct 89–95% of the time.
    • Local degradation: repetition and drift in answer formatting.

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

🎓Nanyang Technological UniversitySingapore · Singapore156 works · 25,644 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksEfficient models & compression · 242 works

Cite

Cite as Weixiao Zhan, Yongcheng Jing, Leszek Rutkowski and Dacheng Tao. 2026. Distillation Traps and Guards: A Calibration Knob for LLM Distillability. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.908

@inproceedings{zhan2026distillation,
  title = {{Distillation Traps and Guards: A Calibration Knob for LLM Distillability}},
  author = {Weixiao Zhan, Yongcheng Jing, Leszek Rutkowski and Dacheng Tao},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.908},
  url = {https://doi.org/10.18653/v1/2026.acl-long.908}
}

References

55 references, 21 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, Matthieu Geist, and Olivier Bachem. 2024. On-policy distillation of language models: Learning from self-generated mistakes. Preprint, arXiv:2306.13649.
  2. AI@Meta. 2024. Llama 3 model card. find
  3. Our calibration method requires fine-tuning the teacher model via reinforcement learning, which is typically more computationally intensive than standard student distillation training. \#\# Ethical Considerations While our undistillable teacher method is designed to protect intellectual property and s… find
  4. Alon Albalak, Duy Phung, Nathan Lile, Rafael Rafailov, Kanishk Gandhi, Louis Castricato, Anikait Singh, Chase Blagden, Violet Xiang, Dakota Mahan, and Nick Haber. 2025. Big-math: A large-scale, high-quality math dataset for reinforcement learning in language models. Preprint, arXiv:2502.17387. find
  5. Anshumann, Mohd Abbas Zaidi, Akhil Kedia, Jinwoo Ahn, Taehwak Kwon, Kangwook Lee, Haejun Lee, and Joohyung Lee. 2025. Sparse logit sampling: Accelerating knowledge distillation in llms. Preprint, arXiv:2503.16870.
  6. Hongzhan Chen, Ruijun Chen, Yuqi Yi, Xiaojun Quan, Chenliang Li, Ming Yan, and Ji Zhang. 2024. Knowledge distillation of black-box large language models. Preprint, arXiv:2401.07013.
  7. Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu, and Xinchao Wang. 2026. dparallel: Learnable parallel decoding for dllms. In International Conference on Learning Representations. find
  8. Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. *Figure 4: Example of self-correction collapse. The… find
  9. DeepMind. 2025. Gemma 3 technical report. Preprint, arXiv:2503.19786. find
  10. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. Deepseek-r1: Incentivizing reasoning capability in ll…
  11. Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. 2018. Born again neural networks. In International conference on machine learning, pages 1607–1616. PMLR. find
  12. Jianping Gou, Baosheng Yu, Stephen J. Maybank, and Dacheng Tao. 2021. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789–1819. find
  13. Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. 2024. Minillm: Knowledge distillation of large language models. Preprint, arXiv:2306.08543.
  14. Yuxian Gu, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2025. Miniplm: Knowledge distillation for pre-training language models. Preprint, arXiv:2410.17215.
  15. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. Preprint, arXiv:1503.02531.
  16. Zheng Yi Ho, Siyuan Liang, Sen Zhang, Yibing Zhan, and Dacheng Tao. 2024. Novo: Norm voting off hallucinations with attention heads in large language models. arXiv preprint arXiv:2410.08970. — [arXiv:2410.08970](https://arxiv.org/abs/2410.08970) find
  17. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. Preprint, arXiv:2106.09685.
  18. Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. 2025. Why language models hallucinate. Preprint, arXiv:2509.04664. — [arXiv:2509.04664](https://arxiv.org/abs/2509.04664) find
  19. Yoon Kim and Alexander M. Rush. 2016. Sequence-level knowledge distillation. Preprint, arXiv:1606.07947.
  20. John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. 2024. A watermark for large language models. Preprint, arXiv:2301.10226.
  21. Siyuan Liang, Tianmeng Fang, Zhe Liu, Aishan Liu, Yan Xiao, Jinyuan He, Ee-Chien Chang, and Xiaochun Cao. 2025a. Safemobile: Chain-level jailbreak detection and automated evaluation for multimodal mobile agents. arXiv preprint arXiv:2507.00841. — [arXiv:2507.00841](https://arxiv.org/abs/2507.00841) find
  22. Siyuan Liang, Jiajun Gong, Tianmeng Fang, Aishan Liu, Tao Wang, Xiaochun Cao, Dacheng Tao, and Chang Ee-Chien. 2026a. Trapflow: Controllable website fingerprinting defense via dynamic backdoor learning. IEEE Transactions on Information Forensics and Security. find
  23. Siyuan Liang, Jiawei Liang, Tianyu Pang, Chao Du, Aishan Liu, Mingli Zhu, Xiaochun Cao, and Dacheng Tao. 2025b. Revisiting backdoor attacks against large vision-language models from domain shift. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9477–9486. find
  24. Siyuan Liang, Jiayang Liu, Jiecheng Zhai, Tianmeng Fang, Rongcheng Tu, Aishan Liu, Xiaochun Cao, and Dacheng Tao. 2026b. T2vshield: Model-agnostic jailbreak defense for text-to-video models. International Journal of Computer Vision, 134(4):144. find
  25. Siyuan Liang, Kuanrong Liu, Jiajun Gong, Jiawei Liang, Yuan Xun, Ee-Chien Chang, and Xiaochun Cao. 2024a. Unlearning backdoor threats: Enhancing backdoor defense in multimodal contrastive learning via local token unlearning. arXiv preprint arXiv:2403.16257. — [arXiv:2403.16257](https://arxiv.org/abs/2403.16257) find
  26. Siyuan Liang, Mingli Zhu, Aishan Liu, Baoyuan Wu, Xiaochun Cao, and Ee-Chien Chang. 2024b. Bad-clip: Dual-embedding guided backdoor attack on multimodal contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24645–24654. find
  27. Chris Yuhao Liu, Liang Zeng, Yuzhen Xiao, Jujie He, Jiacai Liu, Chaojie Wang, Rui Yan, Wei Shen, Fuxiang Zhang, Jiacheng Xu, Yang Liu, and Yahui Zhou. 2025a. Skywork-reward-v2: Scaling preference data curation via human-ai synergy. Preprint, arXiv:2507.01352.
  28. Zihe Liu, Jiashun Liu, Yancheng He, Weixun Wang, Jiaheng Liu, Ling Pan, Xinyu Hu, Shaopan Xiong, Ju Huang, Jian Hu, Shengyi Huang, Siran Yang, Jiamang Wang, Wenbo Su, and Bo Zheng. 2025b. Part i: Tricks or traps? a deep dive into rl for llm reasoning. Preprint, arXiv:2508.08221. find
  29. Hongyu Li, Liang Ding, Meng Fang, and Dacheng Tao. 2024. Revisiting catastrophic forgetting in large language model tuning. Preprint, arXiv:2406.04836. find
  30. Kevin Lu and Thinking Machines Lab. 2025. On-policy distillation. Thinking Machines Lab: Connectionism. Https://thinkingmachines.ai/blog/on-policy-distillation. find
  31. Pingzhi Li, Zhen Tan, Huaizhi Qu, Huan Liu, and Tianlong Chen. 2025. Doge: Defensive output generation for llm protection against knowledge distillation. Preprint, arXiv:2505.19504. find
  32. Haoyu Ma, Tianlong Chen, Ting-Kuei Hu, Chenyu You, Xiaohui Xie, and Zhangyang Wang. 2021. Undistillable: Making a nasty teacher that cannot teach students. Preprint, arXiv:2105.07381. — [arXiv:2105.07381](https://arxiv.org/abs/2105.07381) find
  33. Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. 2020. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, pages 5191– 5198.
  34. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Train…
  35. Charilaos Pipis, Shivam Garg, Vasilis Kontonis, Vaishnavi Shrivastava, Akshay Krishnamurthy, and Dimitris Papailiopoulos. 2025. Wait, wait, wait... why do reasoning models loop? Preprint, arXiv:2512.12895. — [arXiv:2512.12895](https://arxiv.org/abs/2512.12895) find
  36. Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Preprint, arXiv:2305.18290.
  37. Zhiyao Ren, Siyuan Liang, Aishan Liu, and Dacheng Tao. 2025. Iclshield: Exploring and mitigating in-context learning backdoor attacks. arXiv preprint arXiv:2507.01321. — [arXiv:2507.01321](https://arxiv.org/abs/2507.01321) find
  38. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2020. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. Preprint, arXiv:1910.01108. find
  39. John Schulman. 2020. Approximating KL divergence. Blog post. find
  40. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. Preprint, arXiv:1707.06347. find
  41. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  42. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. CommonsenseQA: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  43. M-A-P Team, Xinrun Du, Yifan Yao, Kaijing Ma, Bingli Wang, Tianyu Zheng, Kang Zhu, Minghao Liu, Yiming Liang, Xiaolong Jin, Zhenlin Wei, Chujie Zheng, Kaixing Deng, Shuyue Guo, Shian Jia, Sichao Jiang, Yiyan Liao, Rui Li, Qinrui Li, and 76 others. 2025. Supergpqa: Scaling llm evaluation across 285 g… find
  44. Inar Timiryasov and Jean-Loup Tastet. 2023. Baby llama: knowledge distillation from an ensemble of teachers trained on a small dataset with no performance penalty. Preprint, arXiv:2308.02019. find
  45. Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. 2020. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl. ↗
  46. Qi Wang and Jinjia Zhou. 2025. Topkd: Top-scaled knowledge distillation. Preprint, arXiv:2508.04539. — [arXiv:2508.04539](https://arxiv.org/abs/2508.04539) find
  47. Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, Yuqiong Liu, An Yang, Andrew Zhao, Yang Yue, Shiji Song, Bowen Yu, Gao Huang, and Junyang Lin. 2025. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforceme…
  48. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. 2024. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark…
  49. Yuqiao Wen, Zichao Li, Wenyu Du, and Lili Mou. 2023. f-divergence minimization for sequence-level knowledge distillation. Preprint, arXiv:2307.15190.
  50. Yisong Xiao, Aishan Liu, Siyuan Liang, Zonghao Ying, Xianglong Liu, and Dacheng Tao. 2025. Detoxifying large language models via autoregressive reward guided representation editing. arXiv preprint arXiv:2510.01243. — [arXiv:2510.01243](https://arxiv.org/abs/2510.01243) find
  51. Jiashu Xu, Fei Wang, Mingyu Derek Ma, Pang Wei Koh, Chaowei Xiao, and Muhao Chen. 2024a. Instructional fingerprinting of large language models. Preprint, arXiv:2401.12255. find
  52. Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. 2024b. A survey on knowledge distillation of large language models. Preprint, arXiv:2402.13116. — [arXiv:2402.13116](https://arxiv.org/abs/2402.13116) find
  53. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report.
  54. Zonghao Ying, Aishan Liu, Siyuan Liang, Lei Huang, Jinyang Guo, Wenbo Zhou, Xianglong Liu, and Dacheng Tao. 2026. Safebench: A safety evaluation framework for multimodal large language models. International Journal of Computer Vision, 134(1):18. find
  55. Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Keming Lu, Chuanqi Tan, Chang Zhou, and Jingren Zhou. 2023. Scaling relationship on learning mathematical reasoning with large language models. Preprint, arXiv:2308.01825.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhan-2026-distillation-traps-guards-calibration-knob/