Elamur Research⌕
☰
home / papers / 2026 / Optimizing Length Compression in Large Reasoning Models

Optimizing Length Compression in Large Reasoning Models

Paper · 2026 · Zhengxiang Cheng, Dongping Chen, Mingyang Fu et al.

Summary

Large Reasoning Models often keep re-verifying a solution after they have already reached the correct answer. The authors call this "invalid thinking" and measure it with a new Valid Thinking (VT) rate: the share of thinking tokens produced before the correct answer first appears. Across several strong LRMs, roughly a third or more of the reasoning comes after that point.

They split the usual efficiency goal into two principles. Brevity means stopping once the answer is found. Sufficiency means never dropping steps the solution actually needs. On this basis they build LC-R1, a GRPO-based post-training method with two rewards:

  • a Length Reward that pushes outputs to be shorter overall;
  • a Compress Reward that penalizes the unnecessary tail of reasoning a model keeps generating once it has initially arrived at the right answer. A lightweight parser, LC-Extractor, locates where that tail begins.

On two distilled Qwen models and seven benchmarks, LC-R1 cuts output length by about half while losing only about 2% accuracy. This places it at a strong point on the efficiency–accuracy Pareto frontier. The approach matters because it could make reasoning models substantially cheaper to run without losing their ability to explore solutions.

Key points

  • Diagnostic study: the authors measured VT rates for Qwen3-32B, QwQ-32B, DeepSeek-R1 and Llama-3.3-Nemotron-Super-49B-V1 on AIME25, MATH500, GSM8K, AMC and OlympiadBench. About 35–45% of thinking tokens turned out to be redundant.
  • Main result: LC-R1 was trained on DeepSeek-R1-Distill-Qwen-1.5B and -7B. It gave about 50% average length reduction for about a 2% accuracy drop, and it outperformed SFT, DPO, O1-Pruner and ThinkPrune on the trade-off.
  • Ablations: removing either the Length Reward or the Compress Reward makes results worse. The default weights (α = 1, γ = 1) are a good operating point.
  • Exploration is preserved: with 256 samples on AIME25, Pass@k curves for k = 1 to 128 nearly coincide with those of the original models. Compression is also fairly uniform across problem difficulty.
  • Baselines and limitations:
    • SFT reaches over 95% VT but breaks down on out-of-distribution benchmarks.
    • Experiments only go up to 7B models.
    • Compression may suppress verification that would have been useful on hard problems.

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

🎓University of Maryland, College ParkCollege Park · United States132 works · 25,380 citations🎓Mohamed bin Zayed University of Artificial IntelligenceAbu Dhabi · United Arab Emirates124 works · 5,676 citations🎓Huazhong University of Science and TechnologyWuhan · China45 works · 6,345 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 2Research subjects and languages the work is about.Reasoning · 1,426 worksEfficient models & compression · 242 works

Cite

Cite as Zhengxiang Cheng, Dongping Chen, Mingyang Fu and Tianyi Zhou. 2026. Optimizing Length Compression in Large Reasoning Models. 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.146

@inproceedings{cheng2026optimizing,
  title = {{Optimizing Length Compression in Large Reasoning Models}},
  author = {Zhengxiang Cheng, Dongping Chen, Mingyang Fu and Tianyi Zhou},
  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.146},
  url = {https://doi.org/10.18653/v1/2026.acl-long.146}
}

References

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

  1. Marah Abdin, Sahaj Agarwal, Ahmed Awadallah, Vidhisha Balachandran, Harkirat Behl, Lingjiao Chen, Gustavo de Rosa, Suriya Gunasekar, Mojan Javaheripi, Neel Joshi, Piero Kauffmann, Yash Lara, Caio César Teodoro Mendes, Arindam Mitra, Besmira Nushi, Dimitris Papailiopoulos, Olli Saarikivi, Shital Shah… find
  2. Pranjal Aggarwal and Sean Welleck. 2025. L1: Controlling how long a reasoning model thinks with reinforcement learning. Preprint, arXiv:2503.04697.
  3. Daman Arora and Andrea Zanette. 2025. Training language models to reason efficiently. Preprint, arXiv:2502.04463. find
  4. Simon A. Aytes, Jinheon Baek, and Sung Ju Hwang. 2025. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching. Preprint, arXiv:2503.05179.
  5. Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, Ido Shahaf, Oren Tropp, Ehud Karpas, Ran Zilberstein, Jiaqi Zeng, Soumye Singhal, Alexander Bukharin, ... Yian Zhang, and Chris Alexiuk. 2025. Llama-nemotron: Effici…
  6. Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. 2025. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models. Preprint, arXiv:2503.09567.
  7. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187.
  8. 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 arXiv:2110.14168.
  9. Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, and 1 others. 2025. The danger of overthinking: Examining the reasoning-action dilemma in agentic tasks. arXiv preprint arXiv:2502.08235. 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 Zhen Zhang. 2025. Deepseek-r1: Incentivizing reasoning capability i…
  11. Google. 2025a. Gemini 2.5 flash. https://developers.googleblog.com/en/ start-building-with-gemini-25-flash/. ↗
  12. Google. 2025b. Gemini 2.5 pro. https://cloud.google.com/vertex-ai/generative-ai/docs/ models/gemini/2-5-pro. ↗
  13. Alex Gu, Baptiste Rozière, Hugh Leather, Armando Solar-Lezama, Gabriel Synnaeve, and Sida I. Wang. 2024. Cruxeval: A benchmark for code reasoning, understanding and execution. arXiv preprint arXiv:2401.03065.
  14. Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2024. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547.
  15. Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. 2025. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning. arXiv preprint arXiv:2504.01296.
  16. Jian Hu, Jason Klein Liu, and Wei Shen. 2025. Reinforce++: An efficient rlhf algorithm with robustness to both prompt and reward models. Preprint, arXiv:2501.03262. find
  17. International Conference on Artificial Intelligence in Medicine. The 23rd international conference on artificial intelligence in medicine (aime 2025). https://aime25.aimedicine.info/. Accessed: 2025-06-10. ↗
  18. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.
  19. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. 2024. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974.
  20. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. In The Twelfth International Conference on Learning Representations.
  21. Jiawei Liu and Lingming Zhang. 2025. Code-r1: Reproducing r1 for code with reliable rewards. find
  22. Wei Liu, Ruochen Zhou, Yiyun Deng, Yuzhen Huang, Junteng Liu, Yuntian Deng, Yizhe Zhang, and Junxian He. 2025a. Learn to reason efficiently with adaptive length-based reward shaping. Preprint, arXiv:2505.15612. — [arXiv:2505.15612](https://arxiv.org/abs/2505.15612) find
  23. Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025b. Understanding r1-zero-like training: A critical perspective. Preprint, arXiv:2503.20783.
  24. 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. arXiv preprint arXiv:2504.21659. — [arXiv:2504.21659](https://arxiv.org/abs/2504.21659) find
  25. Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. 2025b. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning. Preprint, arXiv:2501.12570. find
  26. Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025a. Reasoning models can be effective without thinking. Preprint, arXiv:2504.09858. find
  27. Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025b. Cot-valve: Length-compressible chain-of-thought tuning. Preprint, arXiv:2502.09601.
  28. Mathematical Association of America. American mathematics competitions (amc). https://maa-amc.org/student-programs/amc/. Accessed: 2025-06-10. ↗
  29. OpenAI. 2024. Chatgpt. https://openai.com/o1/. ↗
  30. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728– 53741.
  31. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. 2024. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling.
  32. 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.
  33. Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. 2025. Dast: Difficulty-adaptive slow-thinking for large reasoning models. Preprint, arXiv:2503.04472. find
  34. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Hanjie Chen, Xia Hu, and 1 others. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419.
  35. Haoxiang Sun, Yingqian Min, Zhipeng Chen, Wayne Xin Zhao, Zheng Liu, Zhongyuan Wang, Lei Fang, and Ji-Rong Wen. 2025. Challenging the boundaries of reasoning: An olympiad-level math benchmark for large language models. Preprint, arXiv:2503.21380.
  36. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, and 77 others. 2025. Kimi k1.5: Scaling reinforcement l…
  37. Qwen Team. 2024. Qwen2.5: A party of foundation models. find
  38. Qwen Team. 2025b. Qwq-32b: Embracing the power of reinforcement learning. find
  39. 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. ↗
  40. Chenlong Wang, Yuanning Feng, Dongping Chen, Zhaoyang Chu, Ranjay Krishna, and Tianyi Zhou. 2025. Wait, we don’t need to "wait"! removing thinking tokens improves reasoning efficiency. Preprint, arXiv:2506.08343. find
  41. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models. Preprint, arXiv:2201.11903.
  42. Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Zheng Lin, Li Cao, and Weiping Wang. 2025. Dynamic early exit in reasoning models. Preprint, arXiv:2504.15895. find
  43. Jingyang Yi and Jiazheng Wang. 2025. Shorterbetter: Guiding reasoning models to find optimal inference length for efficient reasoning. Preprint, arXiv:2504.21370. — [arXiv:2504.21370](https://arxiv.org/abs/2504.21370) find
  44. Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. 2024. Distilling system 2 into system 1. Preprint, arXiv:2407.06023. find
  45. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, and 1 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476.
  46. Danlong Yuan, Tian Xie, Shaohan Huang, Zhuocheng Gong, Huishuai Zhang, Chong Luo, Furu Wei, and Dongyan Zhao. 2025. Efficient rl training for reasoning models via length-aware optimization. Preprint, arXiv:2505.12284. — [arXiv:2505.12284](https://arxiv.org/abs/2505.12284) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/cheng-2026-optimizing-length-compression-large-reasoning/