Elamur Research⌕
☰
home / papers / 2026 / Concise Math Reasoning via Difficulty-Aware Distillation

Concise Math Reasoning via Difficulty-Aware Distillation

Paper · 2026 · Yifan Wu, Jingze Shi, Bingheng Wu et al.

Summary

Distilling reasoning from large models such as DeepSeek-R1 usually gives small students very long chain-of-thought traces. Students then learn to over-elaborate on every problem, whatever its difficulty, which raises both training and inference cost. The authors propose Difficulty-Aware Distillation (DAD), a two-stage data pipeline. First, the teacher writes a long solution. Then a specially designed prompt has the teacher judge the problem as easy, medium or hard and rewrite the solution to a matching length. Easy items get brief answers, while hard ones keep their necessary detail.

The resulting corpus, LiteCoT, has 100K examples. It was used to fine-tune Qwen2.5-based students called Liter, at 1.5B, 7B and 32B parameters. Across nine math and reasoning benchmarks, these models match or beat counterparts trained on roughly 800K verbose samples, using far fewer training and generation tokens. The work suggests that concise, difficulty-calibrated rationales matter more for math reasoning than sheer data volume.

Key points

  • LiteCoT traces average about 720 tokens. Comparable distillation sets are much longer: S1 about 5.1K, OpenThoughts about 6.8K, LIMO about 7K and ReasonFlux about 10.5K.
  • On AIME24, Liter-32B reaches 74.2% Pass@1 while generating only around 5K inference tokens. The 32B results table also reports 96.6% on MATH500 and 63.6% on GPQA.
  • DAD was compared with other ways of shortening CoT: Chain-of-Draft, LLMLingua-2 at two compression rates, and BudgetAware(2048). DAD gave the best accuracy–speed balance overall, for example 96.9% versus 50.0% for Chain-of-Draft on Sat Math.
  • The authors list three limitations: the method has been tested on text only, it helps less for code generation, and data quality is limited by the teacher model's ability.

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

🎓Université de MontréalMontreal · Canada87 works · 91,597 citations🔬Mila - Quebec Artificial Intelligence InstituteMontreal · Canada65 works · 14,425 citations🎓The Hong Kong University of Science and Technology (Guangzhou)Guangzhou · China67 works · 1,765 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputational Theory and Mathematics · 321 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,685 works

Cite

Cite as Yifan Wu, Jingze Shi, Bingheng Wu, Jiayi Zhang, Xiaotian Lin, Yizhang Zhu, Zhaoyang Yu, Bang Liu, Chenglin Wu, Nan Tang and Yuyu Luo. 2026. Concise Math Reasoning via Difficulty-Aware Distillation. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.2155

@inproceedings{wu2026concise,
  title = {{Concise Math Reasoning via Difficulty-Aware Distillation}},
  author = {Yifan Wu, Jingze Shi, Bingheng Wu, Jiayi Zhang, Xiaotian Lin, Yizhang Zhu, Zhaoyang Yu, Bang Liu, Chenglin Wu, Nan Tang and Yuyu Luo},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2155},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2155}
}

References

46 references, 21 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. rstar-math: Small llms can master math reasoning with self-evolved deep thinking. arXiv preprint arXiv:2501.04519.
  2. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  3. Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2024. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547.
  4. 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 arXiv:2103.03874.
  5. Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2504.01296. ↗
  6. Fanqi Kong, Xiaoyuan Zhang, Xinyu Chen, Yaodong Yang, Song-Chun Zhu, and Xue Feng. 2025. Enhancing llm-based social bot via an adversarial learning framework. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 23246–23271. find
  7. Boyan Li, Chong Chen, Zhujun Xue, Yinan Mei, and Yuyu Luo. 2025a. Deepeye-sql: A software-engineering-inspired text-to-sql framework. CoRR, abs/2510.17586. — [arXiv:2510.17586](https://arxiv.org/abs/2510.17586) find
  8. Simon A. Aytes, Jinheon Baek, and Sung Ju Hwang. 2025. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching. arXiv preprint arXiv:2503.05179.
  9. Christopher A Choquette-Choo, Krishnamurthy Dvijotham, Krishna Pillutla, Arun Ganesh, Thomas Steinke, and Abhradeep Thakurta. 2023. Correlated noise provably beats independent noise for differentially private learning. arXiv preprint arXiv:2310.06771. — [arXiv:2310.06771](https://arxiv.org/abs/2310.06771) find
  10. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  11. LLM Concise. Walk before you run! concise llm reasoning via reinforcement learning. find
  12. Hugging Face. 2025. Open r1: A fully open reproduction of deepseek-r1. find
  13. Boyan Li, Yuyu Luo, Chengliang Chai, Guoliang Li, and Nan Tang. 2024. The dawn of natural language to SQL: are we fully ready? [experiment, analysis & benchmark ]. Proc. VLDB Endow., 17(11):3318– 3331. find
  14. Boyan Li, Yiran Peng, Yupeng Xie, Sirong Lu, Yizhang Zhu, Xing Mu, Xinyu Liu, and Yuyu Luo. 2026. Deepeye: A steerable self-driving data agent system. arXiv preprint arXiv:2603.28889. — [arXiv:2603.28889](https://arxiv.org/abs/2603.28889) find
  15. Boyan Li, Jiayi Zhang, Ju Fan, Yanwei Xu, Chong Chen, Nan Tang, and Yuyu Luo. 2025b. Alpha-sql: Zero-shot text-to-sql using monte carlo tree search. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025. OpenReview.net. find
  16. Weizhe Lin, Xing Li, Zhiyuan Yang, Xiaojin Fu, Hui-Ling Zhen, Yaoyuan Wang, Xianzhi Yu, Wulong Liu, Xiaosong Li, and Mingxuan Yuan. 2025. Trimr: Verifier-based training-free thinking compression for efficient test-time scaling. arXiv preprint arXiv:2505.17155. — [arXiv:2505.17155](https://arxiv.org/abs/2505.17155) find
  17. Haotian Luo, Haiying He, Yibo Wang, Jinluan Yang, Rui Liu, Naiqiang Tan, Xiaochun Cao, Dacheng Tao, and Li Shen. 2025a. Adar1: From long-cot to hybrid-cot via bi-level adaptive reasoning optimization. Preprint, arXiv:2504.21659. — [arXiv:2504.21659](https://arxiv.org/abs/2504.21659) find
  18. Yijia Luo, Yulin Song, Xingyao Zhang, Jiaheng Liu, Weixun Wang, GengRu Chen, Wenbo Su, and Bo Zheng. 2025b. Deconstructing long chain-of-thought: A structured reasoning optimization framework for long cot distillation. arXiv preprint arXiv:2503.16385. math ai. 2025. Minervamath. find
  19. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. 2025. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393.
  20. OpenAI. 2025. Introducing openai o1. find
  21. Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, and 1 others. 2024. Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. arXiv preprint arXiv:2403.12968.
  22. Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, and 1 others. 2025. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614.
  23. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jia… — Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models
  24. NovaSky Team. 2025a. Sky-t1: Train your own o1 preview model within $450. https://novasky-ai.github.io/posts/sky-t1. Accessed: 2025-01-09. ↗
  25. Open Thoughts Team. 2025b. Open Thoughts. find
  26. Fengwei Teng, Zhaoyang Yu, Quan Shi, Jiayi Zhang, Chenglin Wu, and Yuyu Luo. 2025. Atom of thoughts for markov llm test-time scaling. arXiv preprint arXiv:2502.12018. — [arXiv:2502.12018](https://arxiv.org/abs/2502.12018) find
  27. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  28. Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, and 1 others. 2025. Light-r1: Curriculum sft, dpo and rl for long cot from scratch and beyond. arXiv preprint arXiv:2503.10460.
  29. Yifan Wu, Lutao Yan, Leixian Shen, Yunhai Wang, Nan Tang, and Yuyu Luo. 2024. Chartinsights: Evaluating multimodal large language models for low-level chart question answering. Preprint, arXiv:2405.07001. find
  30. Yuyang Wu, Yifei Wang, Tianqi Du, Stefanie Jegelka, and Yisen Wang. 2025. When more is less: Understanding chain-of-thought length in llms. arXiv preprint arXiv:2502.07266.
  31. Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600.
  32. Cheng Yang, Jiaxuan Lu, Haiyuan Wan, Junchi Yu, and Feiwei Qin. 2025a. From what to why: A multi-agent system for evidence-based chemical reaction condition reasoning. arXiv preprint arXiv:2509.23768. — [arXiv:2509.23768](https://arxiv.org/abs/2509.23768) find
  33. Ling Yang, Zhaochen Yu, Bin Cui, and Mengdi Wang. 2025b. Reasonflux: Hierarchical llm reasoning via scaling thought templates. arXiv preprint arXiv:2502.06772. find
  34. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Preprint, arXiv:2305.10601. find
  35. Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. 2025. Limo: Less is more for reasoning. arXiv preprint arXiv:2502.03387.
  36. Huifeng Yin, Yu Zhao, Minghao Wu, Xuanfan Ni, Bo Zeng, Hao Wang, Tianqi Shi, Liangying Shao, Chenyang Lyu, Longyue Wang, and 1 others. 2025. Towards widening the distillation bottleneck for reasoning models. arXiv e-prints, pages arXiv–2503. find
  37. TIGER-Lab. 2024. Mmlu-stem. find
  38. Bin Yu, Hang Yuan, Haotian Li, Xueyin Xu, Yuliang Wei, Bailing Wang, Weizhen Qi, and Kai Chen. 2025a. Long-short chain-of-thought mixture supervised fine-tuning eliciting efficient reasoning in large language models. arXiv preprint arXiv:2505.03469.
  39. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. Preprint, arXiv:2203.11171.
  40. Bin Yu, Hang Yuan, Yuliang Wei, Bailing Wang, Weizhen Qi, and Kai Chen. 2025b. Long-short chain-of-thought mixture supervised fine-tuning eliciting efficient reasoning in large language models. arXiv preprint arXiv:2505.03469.
  41. Qianjin Yu, Keyu Wu, Zihan Chen, Chushu Zhang, Manlin Mei, Lingjun Huang, Fang Tan, Yongsheng Du, Kunlin Liu, and Yurui Zhu. 2025c. Rethinking the generation of high-quality cot data from the perspective of llm-adaptive question difficulty grading. arXiv preprint arXiv:2504.11919. — [arXiv:2504.11919](https://arxiv.org/abs/2504.11919) find
  42. Jintian Zhang, Yuqi Zhu, Mengshu Sun, Yujie Luo, Shuofei Qiao, Lun Du, Da Zheng, Huajun Chen, and Ningyu Zhang. 2025. Lightthinker: Thinking step-by-step compression. CoRR, abs/2502.15589.
  43. Xiaotian Zhang, Chunyang Li, Yi Zong, Zhengyu Ying, Liang He, and Xipeng Qiu. 2023. Evaluating the performance of large language models on gaokao benchmark. arXiv preprint arXiv:2305.12474.
  44. Han Zhao, Haotian Wang, Yiping Peng, Sitong Zhao, Xiaoyu Tian, Shuaiting Chen, Yunjie Ji, and Xiangang Li. 2025a. 1.4 million open-source distilled reasoning dataset to empower large language model training. arXiv preprint arXiv:2503.19633. — [arXiv:2503.19633](https://arxiv.org/abs/2503.19633) find
  45. Shangziqi Zhao, Jiahao Yuan, Guisong Yang, and Usman Naseem. 2025b. Can pruning improve reasoning? revisiting long-cot compression with capability in mind for better reasoning. arXiv preprint arXiv:2505.14582. — [arXiv:2505.14582](https://arxiv.org/abs/2505.14582) find
  46. (1) Our DAD method possesses excellent generalization capabilities, as its effectiveness is not dependent on a specific teacher model. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-concise-math-reasoning-difficulty-aware/