Elamur Research⌕
☰
home / papers / 2026 / Long-Chain Reasoning Distillation via Adaptive Prefix Alignment

Long-Chain Reasoning Distillation via Adaptive Prefix Alignment

Paper · 2026 · Zhenghao Liu, Zhuoyang Wu, Xinze Li et al.

Summary

Small student models often learn poorly from the long, reflective chain-of-thought traces that reasoning models such as DeepSeek-R1 produce. These traces can be longer and more convoluted than the student can absorb. Earlier prefix-based distillation cuts every trace at a fixed fraction, which can leave too little for hard problems and too much for easy ones. The authors propose P-ALIGN, a framework for distilling math reasoning in two stages. First, the student itself judges whether a truncated teacher prefix is enough for it to solve the problem, and a binary search finds the shortest such prefix. Second, the student treats that retained prefix as starting context and writes the rest of the reasoning on its own. These completed trajectories then serve as SFT targets. Across several mathematical reasoning benchmarks, P-ALIGN beats all baselines by more than 3%. The work shows that matching supervision to the student's capacity, rather than copying full teacher traces, gives more effective and more concise reasoning distillation.

Key points

  • Adaptive truncation relies on student self-judgment and binary search to find a minimal sufficient prefix. This replaces the fixed truncation ratio λ used in earlier prefix-SFT work such as Ji et al. 2025 and Chen et al. 2025b.
  • Fixed-ratio experiments show a split by difficulty. On AIME24&25, longer prefixes tend to help, while on MATH500 they hurt, likely through overthinking. P-ALIGN outperforms every fixed ratio tested.
  • The authors split teacher traces into ten chunks and measured student token-level entropy. Entropy rises toward the end of a trace, so later steps are more uncertain. Conditioning on the prefix works much better than conditioning on the middle or suffix segments.
  • P-ALIGN's student outputs are shorter than those from full Long-CoT SFT. UPFT outputs are shorter still, which the authors read as overfitting to surface patterns. With GLM-4.5 as judge, P-ALIGN has the highest preference win rate.
  • Limitations include dependence on strong teacher reasoning models, which is computationally expensive, and on the student's own judging ability. Code is at github.com/NEUIR/P-ALIGN.

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 · China333 works · 87,019 citations🏢Alibaba GroupHangzhou · China168 works · 15,386 citations🎓Northeastern UniversityShenyang · China42 works · 1,940 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Zhenghao Liu, Zhuoyang Wu, Xinze Li, Yukun Yan, Shuo Wang, Zulong Chen, Yu Gu, Ge Yu and Maosong Sun. 2026. Long-Chain Reasoning Distillation via Adaptive Prefix Alignment. 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.822

@inproceedings{liu2026long,
  title = {{Long-Chain Reasoning Distillation via Adaptive Prefix Alignment}},
  author = {Zhenghao Liu, Zhuoyang Wu, Xinze Li, Yukun Yan, Shuo Wang, Zulong Chen, Yu Gu, Ge Yu and Maosong Sun},
  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.822},
  url = {https://doi.org/10.18653/v1/2026.acl-long.822}
}

References

42 references, 23 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 2025. AIME. aime problems and solutions. Accessed: September 23, 2025. 2025. AMC12. amc 12 problems and solutions. Accessed: September 23, 2025. find
  2. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter,… — Language Models are Few-Shot Learners
  3. Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. 2025a. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models. ArXiv preprint, abs/2503.09567.
  4. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al. 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. ArXiv preprint, abs/2412.21187.
  5. Xinjie Chen, Minpeng Liao, Guoxin Chen, Chengxi Li, Biao Fu, Kai Fan, and Xinggao Liu. 2025b. From data-centric to sample-centric: Enhancing llm reasoning via progressive optimization. ArXiv preprint, abs/2507.06573. find
  6. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. ArXiv preprint, abs/2110.14168.
  7. DeepSeek-AI, Daya Guo, Dejian Yang, and Haowei Zhang. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. ArXiv preprint, abs/2501.12948.
  8. Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I. Webb, Rob J. Hyndman, and Pablo Montero-Manso. 2021. Monash time series forecasting archive. ArXiv preprint, abs/2105.06643. — [arXiv:2105.06643](https://arxiv.org/abs/2105.06643) find
  9. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, and Abhishek Kadian. 2024. The llama 3 herd of models. ArXiv preprint, abs/2407.21783.
  10. Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. The false promise of imitating proprietary llms. ArXiv preprint, abs/2305.15717.
  11. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. ArXiv preprint, abs/2103.03874.
  12. Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. In Proceedings of ACL Findings, pages 8003–…
  13. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In Proceedings of ICLR.
  14. Ke Ji, Jiahao Xu, Tian Liang, Qiuzhi Liu, Zhiwei He, Xingyu Chen, Xiaoyuan Liu, Zhijie Wang, Junying Chen, Benyou Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. The first few tokens are all you need: An efficient and effective unsupervised prefix fine-tuning method for reasoning models. ArXiv prep… — [arXiv:2503.02875](https://arxiv.org/abs/2503.02875) find
  15. Zhensheng Jin, Xinze Li, Yifan Ji, Chunyi Peng, Zhenghao Liu, Qi Shi, Yukun Yan, Shuo Wang, Furong Peng, and Ge Yu. 2025. ReCUT: Balancing reasoning length and accuracy in LLMs via stepwise trails and preference optimization. In Proceedings of EMNLP Findings, pages 14269–14282. find
  16. Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? In Proceedings of NeurIPS, pages 5574– 5584.
  17. Chengpeng Li, Zheng Yuan, Guanting Dong, Keming Lu, Jiancan Wu, Chuanqi Tan, Xiang Wang, and Chang Zhou. 2023. Query and response augmentation cannot help out-of-domain math reasoning generalization. ArXiv preprint, abs/2310.05506. find
  18. Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubramanian, and Radha Poovendran. 2025. Small models struggle to learn from strong reasoners. ArXiv preprint, abs/2502.12143. find
  19. Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jian-guang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. ArXiv preprint, abs/2308.09583.
  20. Renjie Luo, Jiaxi Li, Chen Huang, and Wei Lu. 2025a. Through the valley: Path to effective long CoT training for small language models. In Proceedings of EMNLP, pages 4972–4992.
  21. Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2025b. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. IEEE Transactions on Audio, Speech and Language Processing. find
  22. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori B Hashimoto. 2025. s1: Simple test-time scaling. In Proceedings of EMNLP, pages 20286– 20332.
  23. Libo Qin, Qiguang Chen, Fuxuan Wei, Shijue Huang, and Wanxiang Che. 2023. Cross-lingual prompting: Improving zero-shot chain-of-thought reasoning across languages. In Proceedings of EMNLP, pages 2695–2709.
  24. Yiliu Sun, Zicheng Zhao, Yang Wei, Yanfang Zhang, and Chen Gong. 2025. Well begun, half done: Reinforcement learning with prefix optimization for llm reasoning. ArXiv preprint, abs/2512.15274. — [arXiv:2512.15274](https://arxiv.org/abs/2512.15274) find
  25. Zihan Wang, Zihan Liang, Zhou Shao, Yufei Ma, Huangyu Dai, Ben Chen, Lingtao Mao, Chenyi Lei, Yuqing Ding, and Han Li. 2025. Infogain-rag: Boosting retrieval-augmented generation through document information gain-based reranking and filtering. In Proceedings of EMNLP, pages 4972–4992. find
  26. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of NeurIPS, pages 24824–24837.
  27. Alexander Wettig, Aatmik Gupta, Saumya Malik, and Danqi Chen. 2024. Qurating: Selecting high-quality data for training language models. ArXiv preprint, abs/2402.09739. find
  28. Zhuoyang Wu, Xinze Li, Zhenghao Liu, Yukun Yan, Zhiyuan Liu, Minghe Yu, Cheng Yang, Yu Gu, Ge Yu, and Maosong Sun. 2025. Enhancing long-chain reasoning distillation through error-aware self-reflection. ArXiv preprint, abs/2505.22131.
  29. Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. ArXiv preprint, abs/2502.18600. find
  30. An Yang, Anfeng Li, Baosong Yang, and et al. Beichen Zhang. 2025. Qwen3 technical report. ArXiv preprint, abs/2505.09388.
  31. An Yang, Baosong Yang, Beichen Zhang, and Binyuan Hui. 2024. Qwen2.5 technical report. ArXiv preprint, abs/2412.15115. find
  32. Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. 2025. Limo: Less is more for reasoning. ArXiv preprint, abs/2502.03387.
  33. Huifeng Yin, Yu Zhao, Minghao Wu, Xuanfan Ni, Bo Zeng, Hao Wang, Tianqi Shi, Liangying Shao, Chenyang Lyu, Longyue Wang, Weihua Luo, and Kaifu Zhang. 2025. Marco-o1 v2: Towards widening the distillation bottleneck for reasoning models. ArXiv preprint, abs/2503.01461.
  34. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. Star: Bootstrapping reasoning with reasoning. In Proceedings of NeurIPS, pages 15476– 15488.
  35. Peiyi Wang, Lei Li, Liang Chen, Feifan Song, Binghuai Lin, Yunbo Cao, Tianyu Liu, and Zhifang Sui. 2023. Making large language models better reasoners with alignment. ArXiv preprint, abs/2309.02144. — [arXiv:2309.02144](https://arxiv.org/abs/2309.02144) find
  36. Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, et al. 2025a. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. ArXiv preprint, abs/2508.06471.
  37. Renxi Wang, Haonan Li, Xudong Han, Yixuan Zhang, and Timothy Baldwin. 2024. Learning from failure: Integrating negative examples when fine-tuning large language models as agents. ArXiv preprint, abs/2402.11651. — [arXiv:2402.11651](https://arxiv.org/abs/2402.11651) find
  38. Wenhao Zeng, Yaoning Wang, Chao Hu, Yuling Shi, Chengcheng Wan, Hongyu Zhang, and Xiaodong Gu. 2025b. Pruning the unsurprising: Efficient code reasoning via first-token surprisal. ArXiv preprint, abs/2508.05988. — [arXiv:2508.05988](https://arxiv.org/abs/2508.05988) find
  39. Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, and Xi Victoria Lin. 2022. Opt: Open pre-trained transformer language models. ArXiv preprint, abs/2205.01068.
  40. Yongheng Zhang, Qiguang Chen, Jingxuan Zhou, Peng Wang, Jiasheng Si, Jin Wang, Wenpeng Lu, and Libo Qin. 2024. Wrong-of-thought: An integrated reasoning framework with multi-perspective verification and wrong information. In Proceedings of EMNLP Findings, pages 6644–6653. find
  41. Dataset s1K-1.1 (2025) AIME25 (2025) AIME24 (2025) AMC12 (2025) MATH500 (2021) find
  42. Train 1,000 — — Test -30 30 83 500 Backbone Model Qwen2.5-7B Qwen3-8B Llama3.2-3B Long-CoT 9,291 9,291 8,755 Prefix 4,531 4,402 4,780 P-ALIGN 5,453 5,732 6,035 *Table 4: Average Token Length Statistics of Training Dataset Across Different Student Models. We evaluate the average token lengths of th… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-long-chain-reasoning-distillation-adaptive/