Elamur Research⌕
☰
home / papers / 2026 / Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models

Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models

Paper · 2026 · Yongjiang Liu, Haoxi Li, Xiaosong Ma et al.

Summary

Large Reasoning Models (LRMs) tend to overthink: they produce long, repetitive chains of thought even for trivial questions. The authors argue that this happens because the models cannot judge a problem's difficulty before solving it, so they apply the same heavy strategy to everything. They test this by prompting models to rate the difficulty of GSM8K questions versus harder sets such as MATH, AIME2024 and OmniMath. The ratings largely blur easy and hard tasks together.

To fix this, they introduce Think-How-to-Think (TH2T), a two-stage fine-tuning scheme on a hybrid dataset that mixes short and long reasoning traces:

  • Stage 1 (Difficulty Hypnosis): difficulty cues are placed at the start of the model's output, steering its overall reasoning depth.
  • Stage 2 (Redundancy Hypnosis): cues are inserted inside ongoing reasoning so the model learns to cut unneeded reflections, double-checks and loops.

On R1-Distill-Qwen models at 7B, 14B and 32B scale, TH2T shortens outputs by more than 70% on easy tasks and about 40% on hard ones while keeping accuracy. This points toward cheaper reasoning that adapts to difficulty.

Key points

  • TH2T is compared with D-Prompt, NoThinking, TokenSkip and CoT-Value on GSM8K and MATH. It reaches strong accuracy at much lower relative token length.
  • On MATH, looping endings appear in 59.5% of the original model's wrong answers but only 21.4% after TH2T. No loops were observed on GSM8K.
  • The method generalizes beyond its training domain. It works on AIME2024 and OmniMath, and on four STEM tasks (QA, coding, science, multi-task) it cuts tokens by over 30% without losing accuracy.
  • The ablation supports the two-stage design (7B model):
    • Difficulty Hypnosis mainly helps easy tasks, cutting tokens by 26.6% on GSM8K versus 4.8% on MATH.
    • Redundancy Hypnosis matters more on hard tasks, cutting tokens by 17.3% on MATH versus 11.6% on GSM8K.
  • After TH2T, the model is more confident on its first token and rates difficulty more accurately. The authors read this as evidence that difficulty awareness is reflected in the model's internal states.

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

🎓Hong Kong University of Science and TechnologyHong Kong · China142 works · 25,136 citations🎓Hong Kong Polytechnic UniversityHong Kong · China84 works · 4,956 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 Yongjiang Liu, Haoxi Li, Xiaosong Ma, Jie Zhang and Song Guo. 2026. Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition 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.1766

@inproceedings{liu2026think,
  title = {{Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models}},
  author = {Yongjiang Liu, Haoxi Li, Xiaosong Ma, Jie Zhang and Song Guo},
  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.1766},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1766}
}

References

69 references, 27 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Daman Arora and Andrea Zanette. 2025. Training language models to reason efficiently. Preprint, arXiv:2502.04463. find
  2. 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.
  3. Qiguang Chen, Dengyun Peng, Jinhao Liu, HuiKang Su, Jiannan Guan, Libo Qin, and Wanxiang Che. 2025a. Aware first, think less: Dynamic boundary self-awareness drives extreme reasoning efficiency in large language models. arXiv preprint arXiv:2508.11582. — [arXiv:2508.11582](https://arxiv.org/abs/2508.11582) find
  4. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025b. Do not think that much for 2+3=? on the overthinking of o1-like llms. Preprint, arXiv:2412.21187.
  5. Jeffrey Cheng and Benjamin Van Durme. 2024. Compressed chain of thought: Efficient reasoning through dense representations. Preprint, arXiv:2412.13171. 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. Preprint, arXiv:2110.14168.
  7. Yuntian Deng, Yejin Choi, and Stuart Shieber. 2024. From explicit cot to implicit cot: Learning to internalize cot step by step. Preprint, arXiv:2405.14838. find
  8. Mengru Ding, Hanmeng Liu, Zhizhang Fu, Jian Song, Wenbo Xie, and Yue Zhang. 2024. Break the chain: Large language models can be shortcut reasoners. Preprint, arXiv:2406.06580. — [arXiv:2406.06580](https://arxiv.org/abs/2406.06580) find
  9. Anastasia Efklides. 2011. Interactions of metacognition with motivation and affect in self-regulated learning: The masrl model. Educational psychologist, 46(1):6– 25. find
  10. Jonathan St B T Evans and Keith E Stanovich. 2013. Dual-process theories of higher cognition: Advancing the debate. Perspectives on Psychological Science, 8(3):223–241. find
  11. Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2025. Thinkless: LLM learns when to think. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  12. John H Flavell. 1979. Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry. American Psychologist, 34(10):906–911. find
  13. Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. 2025. Deep think with confidence. arXiv preprint arXiv:2508.15260. find
  14. Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, Zhengyang Tang, Benyou Wang, Daoguang Zan, Shanghaoran Quan, Ge Zhang, Lei Sha, Yichang Zhang, Xuancheng Ren, Tianyu Liu, and Baobao Chang. 2024. Omni-math: A universal olympiad level ma…
  15. Yuyao Ge, Shenghua Liu, Yiwei Wang, Lingrui Mei, Lizhe Chen, Baolong Bi, and Xueqi Cheng. 2025. Innate reasoning is not enough: In-context learning enhances reasoning large language models with less overthinking. Preprint, arXiv:2503.19602. — [arXiv:2503.19602](https://arxiv.org/abs/2503.19602) find
  16. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and et al. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  17. 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.
  18. Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2025. Token-budget-aware LLM reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 24842–24855, Vienna, Austria. Association for Computational Linguistics.
  19. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space. Preprint, arXiv:2412.06769. find
  20. Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, and 1 others. 2024. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. In Proceedings of the 62nd Annual Meeting…
  21. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021a. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR).
  22. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, and Dawn Song. 2021b. Measuring mathematical problem solving with the math dataset. NeurIPS.
  23. Shijue Huang, Hongru Wang, Wanjun Zhong, Zhaochen Su, Jiazhan Feng, Bowen Cao, and Yi R Fung. 2025. Adactrl: Towards adaptive and controllable reasoning via difficulty-aware budgeting. arXiv preprint arXiv:2505.18822. — [arXiv:2505.18822](https://arxiv.org/abs/2505.18822) find
  24. 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.
  25. Daniel Kahneman. 2011. Thinking, Fast and Slow. Farrar, Straus and Giroux. find
  26. Yu Kang, Xianghui Sun, Liangyu Chen, and Wei Zou. 2025. C3ot: Generating shorter chain-of-thought without compromising effectiveness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24312–24320. find
  27. Ayeong Lee, Ethan Che, and Tianyi Peng. 2025. How well do llms compress their own chain-of-thought? a token complexity approach. Preprint, arXiv:2503.01141. — [arXiv:2503.01141](https://arxiv.org/abs/2503.01141) find
  28. Dacheng Li, Shiyi Cao, Tyler Griggs, Shu Liu, Xiangxi Mo, Eric Tang, Sumanth Hegde, Kourosh Hakhamaneshi, Shishir G. Patil, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica. 2025a. Llms can easily learn to reason from demonstrations structure, not content, is what matters! Preprint, arXiv:2502.0737… find
  29. Yiwei Li, Peiwen Yuan, Shaoxiong Feng, Boyuan Pan, Xinglin Wang, Bin Sun, Heda Wang, and Kan Li. 2024. Escape sky-high cost: Early-stopping self-consistency for multi-step reasoning. In The Twelfth International Conference on Learning Representations. find
  30. Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, Yingying Zhang, Fei Yin, Jiahua Dong, Zhiwei Li, Bao-Long Bi, Ling-Rui Mei, Junfeng Fang, Xiao Liang, Zhijiang Guo, and 2 others. 2025b. From system 1 to system 2:…
  31. Baohao Liao, Yuhui Xu, Hanze Dong, Junnan Li, Christof Monz, Silvio Savarese, Doyen Sahoo, and Caiming Xiong. 2025. Reward-guided speculative decoding for efficient LLM reasoning. In Forty-second International Conference on Machine Learning. find
  32. Alec Helyar, Aleksander Madry, Alex Beutel, and et al. 2024a. Openai o1 system card. Preprint, arXiv:2412.16720.
  33. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, and et al. 2024b. Gpt-4 technical report. Preprint, arXiv:2303.08774.
  34. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, and et al. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  35. 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.
  36. Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025. Understanding r1-zero-like training: A critical perspective. Preprint, arXiv:2503.20783.
  37. 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.
  38. Haotian Luo, Haiying He, Yibo Wang, Jinluan Yang, Rui Liu, Naiqiang Tan, Xiaochun Cao, Dacheng Tao, and Li Shen. 2025a. Ada-r1: Hybrid-cot via bi-level adaptive reasoning optimization. Preprint, arXiv:2504.21659. — [arXiv:2504.21659](https://arxiv.org/abs/2504.21659) find
  39. Matthew Renze and Erhan Guven. 2024. The benefits of a concise chain of thought on problem-solving in large language models. In 2024 2nd International Conference on Foundation and Large Language Models (FLLM), pages 476–483. IEEE. find
  40. 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
  41. Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J. Reddi. 2025. Reasoning with latent thoughts: On the power of looped transformers. In The Thirteenth International Conference on Learning Representations. find
  42. 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
  43. Xuan Shen, Yizhou Wang, Xiangxi Shi, Yanzhi Wang, Pu Zhao, and Jiuxiang Gu. 2025a. Efficient reasoning with hidden thinking. Preprint, arXiv:2501.19201. — [arXiv:2501.19201](https://arxiv.org/abs/2501.19201) find
  44. Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025b. CoT-valve: Length-compressible chain-of-thought tuning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6025–6035, Vienna, Austria. Association for C…
  45. Yu Meng and Mengzhou Xia. 2024. Simpo: Simple preference optimization with a reference-free reward. Advances in Neural Information Processing Systems, 37:124198–124235. find
  46. Xuefei Ning, Zinan Lin, Zixuan Zhou, Zifu Wang, Huazhong Yang, and Yu Wang. 2024. Skeleton-of-thought: Prompting LLMs for efficient parallel generation. In The Twelfth International Conference on Learning Representations. find
  47. OpenAI, :, Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, Zhaoxiang Liu, and Shiguo Lian. 2025b. Dast: Difficulty-adaptive slow-thinking for large reasoning models. Pre… find
  48. Leheng Sheng, An Zhang, Zijian Wu, Weixiang Zhao, Changshuo Shen, Yi Zhang, Xiang Wang, and Tat-Seng Chua. 2025. On reasoning strength planning in large reasoning models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  49. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations. find
  50. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, and Xia Hu. 2025. Stop overthinking: A survey on efficient reasoning for large language models. Preprint, arXiv:2503.16419. 2023. Self-evaluation guided beam search for…
  51. Hanshi Sun, Momin Haider, Ruiqi Zhang, Huitao Yang, Jiahao Qiu, Ming Yin, Mengdi Wang, Peter Bartlett, and Andrea Zanette. 2024a. Fast best-of-n decoding via speculative rejection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. find
  52. Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. Preprint, arXiv:2502.18600. find
  53. Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, and et al. 2024b. A survey of reasoning with foundation models. Preprint, arXiv:2312.11562. — [arXiv:2312.11562](https://arxiv.org/abs/2312.11562) find
  54. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, and 1 others. 2025. Kimi k1.5: Scaling reinforcement learning with llms. Preprint, arXiv:2501.12599.
  55. Kaiwen Yan, Xuanqing Shi, Hongcheng Guo, Wenxuan Wang, Zhuosheng Zhang, and Chengwei Qin. 2025. Drqa: Dynamic reasoning quota allocation for controlling overthinking in reasoning large language models. arXiv preprint arXiv:2508.17803. — [arXiv:2508.17803](https://arxiv.org/abs/2508.17803) find
  56. Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Zheng Lin, Li Cao, and Weiping Wang. 2025a. Dynamic early exit in reasoning models. Preprint, arXiv:2504.15895. find
  57. Wenkai Yang, Shuming Ma, Yankai Lin, and Furu Wei. 2025b. Towards thinking-optimal scaling of test-time compute for llm reasoning. Preprint, arXiv:2502.18080.
  58. Qwen Team. 2025. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github.io/blog/qwq-32b/. ↗
  59. Fei Yu, Hongbo Zhang, Prayag Tiwari, and Benyou Wang. 2024. Natural language reasoning, a survey. ACM Computing Surveys, 56(12):1–39. find
  60. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations.
  61. Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. 2025a. Reasoning models know when they’re right: Probing hidden states for self-verification. Preprint, arXiv:2504.05419. find
  62. Yiming Wang, Pei Zhang, Siyuan Huang, Baosong Yang, Zhuosheng Zhang, Fei Huang, and Rui Wang. 2025. Sampling-efficient test-time scaling: Self-estimating the best-of-n sampling in early decoding. Preprint, arXiv:2503.01422. find
  63. 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.
  64. Haoyue Zhang, Hualei Zhang, Xiaosong Ma, Jie Zhang, and Song Guo. 2025b. Lazyeviction: Lagged kv eviction with attention pattern observation for efficient long reasoning. arXiv preprint arXiv:2506.15969.
  65. Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. 2025c. Adaptthink: Reasoning models can learn when to think. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 3716–3730.
  66. Tong Wu, Chong Xiang, Jiachen T. Wang, G. Edward Suh, and Prateek Mittal. 2025. Effectively controlling reasoning models through thinking intervention. Preprint, arXiv:2503.24370. — [arXiv:2503.24370](https://arxiv.org/abs/2503.24370) find
  67. Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. Preprint, arXiv:2502.12067.
  68. Tian Xie, Zitian Gao, Qingnan Ren, Haoming Luo, Yuqian Hong, Bryan Dai, Joey Zhou, Kai Qiu, Zhirong Wu, and Chong Luo. 2025. Logic-rl: Unleashing llm reasoning with rule-based reinforcement learning. Preprint, arXiv:2502.14768.
  69. Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-think-how-think-mitigating-overthinking/