Elamur Research⌕
☰
home / papers / 2026 / Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning

Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning

Paper · 2026 · Renliang Sun, Wei Cheng, Dawei Li et al.

Summary

Reasoning LLMs often keep producing chain-of-thought steps long after the answer is effectively settled. This "overthinking" wastes compute and can even push the model toward a wrong answer. The authors present REFRAIN, a framework that needs no training and decides at inference time when to cut reasoning short. It splits the trace into steps at blank lines and stops only when three conditions hold:

  • a provisional answer has already appeared;
  • the current step contains a reflection cue, drawn from vocabularies for self-checking, strategy shifts, uncertainty, and retrospective revision;
  • the step is semantically close to an earlier one, measured by cosine similarity of all-MiniLM-L6-v2 embeddings against a threshold τ.

Instead of fixing τ by hand, a sliding-window UCB bandit chooses it online from a set of candidates. Its reward is the answer's length-normalized likelihood minus a token-length penalty. On GSM8K, MATH-500, CSQA and GPQA-Diamond, using Qwen3 and gpt-oss models, REFRAIN cuts tokens by 20–55% while matching or beating vanilla CoT accuracy. This positions deciding when to stop as a cheap, unsupervised form of test-time scaling.

Key points

  • Two-stage stop discriminator: the stage-one check (reflection-trigger detection with a provisional-answer requirement) and the stage-two check (max cosine redundancy, threshold τ) must both fire before reasoning halts.
  • SW-UCB bandit for τ: this multi-armed bandit adapts the stopping threshold to problem difficulty without supervision or fine-tuning. It trades off exploration against exploitation.
  • Baselines and test-time scaling: compared with DEER, HALT-CoT, AlphaOne, No-thinking, DeepConf, the fine-tuned CoT-Valve, and fixed thinking budgets, REFRAIN sits above the budget-vs-accuracy curve for Qwen3-8B.
  • Wall-clock gains:
    • Qwen3-8B on CSQA: 54% fewer tokens and about 40% less inference time.
    • gpt-oss-20B on GSM8K: 45% fewer tokens and 28% less time.
    • The smaller time savings come from overhead when decoding is segmented.
  • Overhead and limitation: sentence-encoder scoring takes under 1% of runtime. The method requires access to streamed intermediate reasoning, so it cannot be applied to closed APIs that return only final answers.

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 California, Los AngelesLos Angeles · United States87 works · 54,961 citations🎓Arizona State UniversityTempe · United States50 works · 2,032 citations🏢NEC (United States)Irving · United States6 works · 539 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 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen and Wei Wang. 2026. Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning. 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.1256

@inproceedings{sun2026stop,
  title = {{Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning}},
  author = {Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen and Wei Wang},
  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.1256},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1256}
}

References

47 references, 26 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, and 1 others. 2025. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925.
  2. 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.
  3. Cheng-Han Chiang and Hung-yi Lee. 2024. Over-reasoning and redundant calculation of large language models. arXiv preprint arXiv:2401.11467.
  4. 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.
  5. 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
  6. Yingqian Cui, Pengfei He, Jingying Zeng, Hui Liu, Xianfeng Tang, Zhenwei Dai, Yan Han, Chen Luo, Jing Huang, Zhen Li, and 1 others. 2025. Stepwise perplexity-guided refinement for efficient chain-of-thought reasoning in large language models. arXiv preprint arXiv:2502.13260. find
  7. Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. 2025. Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514. — [arXiv:2504.06514](https://arxiv.org/abs/2504.06514) find
  8. Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. 2025. Deep think with confidence. arXiv preprint arXiv:2508.15260. find
  9. Aurélien Garivier and Eric Moulines. 2008. On upper-confidence bound policies for non-stationary bandit problems. arXiv preprint arXiv:0805.3415. — [arXiv:0805.3415](https://arxiv.org/abs/0805.3415) find
  10. Aryo Pradipta Gema, Alexander Hägele, Runjin Chen, Andy Arditi, Jacob Goldman-Wetzler, Kit Fraser-Taliente, Henry Sleight, Linda Petrini, Julian Michael, Beatrice Alex, and 1 others. 2025. Inverse scaling in test-time compute. arXiv preprint arXiv:2507.14417. — [arXiv:2507.14417](https://arxiv.org/abs/2507.14417) find
  11. 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.
  12. Xingyang He, Xiao Ling, and Jie Liu. 2025. Smartthinker: Learning to compress and preserve reasoning by step-level length control. arXiv preprint arXiv:2507.04348. — [arXiv:2507.04348](https://arxiv.org/abs/2507.04348) find
  13. 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. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).
  14. Jiameng Huang, Baijiong Lin, Guhao Feng, Jierun Chen, Di He, and Lu Hou. 2025. Efficient reasoning for large reasoning language models via certainty-guided reflection suppression. arXiv preprint arXiv:2508.05337. — [arXiv:2508.05337](https://arxiv.org/abs/2508.05337) find
  15. Yuxuan Jiang, Dawei Li, and Frank Ferraro. 2025. Drp: Distilled reasoning pruning with skill-aware step decomposition for efficient large reasoning models. arXiv preprint arXiv:2505.13975.
  16. Yassir Laaouach. 2025. Halt-cot: Model-agnostic early stopping for chain-of-thought reasoning via answer entropy. In 4th Muslims in ML Workshop co-located with ICML 2025. find
  17. Ruosen Li, Ziming Luo, Quan Zhang, Ruochen Li, Ben Zhou, Ali Payani, and Xinya Du. 2025a. Aalc: Large language model efficient reasoning via adaptive accuracy-length control. arXiv preprint arXiv:2506.20160. — [arXiv:2506.20160](https://arxiv.org/abs/2506.20160) find
  18. 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. arXiv preprint arXiv:2401.10480. find
  19. Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, and 1 others. 2025b. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419.
  20. 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. 2025. Skywork-reward-v2: Scaling preference data curation via human-ai synergy. arXiv preprint arXiv:2507.01352.
  21. Kang Liu, Yongkang Liu, Xiaocui Yang, Peidong Wang, Wen Zhang, Shi Feng, Yifei Zhang, and Daling Wang. 2026. Neat: Neuron-based early exit for large reasoning models. arXiv preprint arXiv:2602.02010.
  22. Xin Liu and Lu Wang. 2025. Answer convergence as a signal for early stopping in reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 17907–17918.
  23. Zihan Liu, Yang Chen, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2024. Acemath: Advancing frontier math reasoning with post-training and reward modeling. arXiv preprint. find
  24. Jingyuan Ma, Damai Dai, Zihang Yuan, Weilin Luo, Bin Wang, Qun Liu, Lei Sha, Zhifang Sui, and 1 others. 2024. Large language models struggle with unreasonability in math problems. arXiv preprint arXiv:2403.19346. — [arXiv:2403.19346](https://arxiv.org/abs/2403.19346) find
  25. Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025a. Reasoning models can be effective without thinking. arXiv preprint arXiv:2504.09858. find
  26. Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025b. Cot-valve: Length-compressible chain-of-thought tuning. arXiv preprint arXiv:2502.09601.
  27. 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.
  28. Yongqi Tong, Dawei Li, Sizhe Wang, Yujia Wang, Fei Teng, and Jingbo Shang. 2024. Can llms learn from previous mistakes? investigating llms’ errors to boost for reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3065–30…
  29. Xiaoxuan Wang, Han Zhang, Haixin Wang, Yidan Shi, Ruoyan Li, Kaiqiao Han, Chenyi Tong, Haoran Deng, Renliang Sun, Alexander Taylor, and 1 others. 2026. Arlarena: A unified framework for stable agentic reinforcement learning. arXiv preprint arXiv:2602.21534. — [arXiv:2602.21534](https://arxiv.org/abs/2602.21534) find
  30. 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.
  31. Zihao Wei, Liang Pang, Jiahao Liu, Wenjie Shi, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Fei Sun, Huawei Shen, and Xueqi Cheng. 2026. The evolution of thought: Tracking llm overthinking via reasoning dynamics analysis. Preprint, arXiv:2508.17627.
  32. Ziqing Qiao, Yongheng Deng, Jiali Zeng, Dong Wang, Lai Wei, Guanbo Wang, Fandong Meng, Jie Zhou, Ju Ren, and Yaoxue Zhang. 2025. Concise: Confidence-guided compression in step-by-step efficient reasoning. arXiv preprint arXiv:2505.04881.
  33. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and 1 others. 2019. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.
  34. 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.
  35. Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. Softcot++: Test-time scaling with soft chain-of-thought reasoning. arXiv preprint arXiv:2505.11484. — [arXiv:2505.11484](https://arxiv.org/abs/2505.11484) find
  36. Aman Sharma and Paras Chopra. 2025. Think just enough: Sequence-level entropy as a confidence signal for llm reasoning. arXiv preprint arXiv:2510.08146. find
  37. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  38. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, and 1 others. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419.
  39. Richard S Sutton, Andrew G Barto, and 1 others. 1998. Reinforcement learning: An introduction, volume 1. MIT press Cambridge. find
  40. 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,…
  41. Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Zheng Lin, Li Cao, and Weiping Wang. 2025b. Dynamic early exit in reasoning models. arXiv preprint arXiv:2504.15895. find
  42. Wang Yang, Xiang Yue, Vipin Chaudhary, and Xiaotian Han. 2025c. Speculative thinking: Enhancing small-model reasoning with large model guidance at inference time. arXiv preprint arXiv:2504.12329. — [arXiv:2504.12329](https://arxiv.org/abs/2504.12329) find
  43. Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. 2025. Demystifying long chain-of-thought reasoning in llms. arXiv preprint arXiv:2502.03373.
  44. Longxuan Yu, Delin Chen, Siheng Xiong, Qingyang Wu, Dawei Li, Zhikai Chen, Xiaoze Liu, and Liangming Pan. 2025. Causaleval: Towards better causal reasoning in language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistic… find
  45. Junyu Zhang, Runpei Dong, Han Wang, Xuying Ning, Haoran Geng, Peihao Li, Xialin He, Yutong Bai, Jitendra Malik, Saurabh Gupta, and 1 others. 2025a. Alphaone: Reasoning models thinking slow and fast at test time. arXiv preprint arXiv:2505.24863.
  46. Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2025b. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301.
  47. Chengshuai Zhao, Zhen Tan, Pingchuan Ma, Dawei Li, Bohan Jiang, Yancheng Wang, Yingzhen Yang, and Huan Liu. 2025. Is chain-of-thought reasoning of llms a mirage? a data distribution lens. arXiv preprint arXiv:2508.01191.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sun-2026-stop-when-enough-adaptive-early/