Elamur Research⌕
☰
home / papers / 2026 / ThinkBrake: Efficient Reasoning via Log-Probability Margin Guided Decoding

ThinkBrake: Efficient Reasoning via Log-Probability Margin Guided Decoding

Paper · 2026 · Sangjun Song, Minjae Oh, Seungkyu Lee et al.

Summary

Large Reasoning Models (LRMs) often overthink. They find a correct intermediate answer in their chain of thought, keep deliberating, and then replace it with a wrong one. The authors first measure this with Qwen3-4B-Thinking. They force the </think> delimiter at each sentence boundary and pick the best cut-off in hindsight. This oracle raises accuracy by about 8% on average and shrinks reasoning length by 72% on math and BFCL tool-calling tasks.

Building on that finding, they propose ThinkBrake, a decoding rule that needs no training. At every sentence boundary, it compares the log-probability of the likeliest non-stop token with that of </think>. Reasoning ends once this gap falls below a threshold τ. The authors prove this rule is a temperature-scaled logit margin test, which also amounts to KL-regularized test-time realignment with a reward bonus for </think>.

Across six LRMs and eight benchmarks, ThinkBrake cuts thinking tokens by up to 30% while keeping accuracy competitive. Its traces can also be used with DPO to train models that reason concisely on their own.

Key points

  • Oracle analysis on Qwen3-4B-Thinking:
    • Error rate drops to about 6% on math (GSM8K, MATH500, AIME2024, AIME2025) and 5% on BFCL-v1.
    • Token savings are 54% on math and 89% on tool use.
    • On AIME, 61% of failures are recovered with 44% fewer tokens.
  • ThinkBrake stops when log(p(top token) / p(</think>)) ≤ τ. A linear probability-gap variant (ThinkBrake-p) performed clearly worse.
  • Evaluation covered GSM8K, MATH500, AIME2024/2025, GPQA-Diamond, ARC-Challenge, BFCL and Meta-Tool. Accuracy held steady across difficulty levels, and most predictions kept the same correctness after early stopping.
  • DPO setup: 1.3K preference pairs built from Omni-MATH, trained in about 20 minutes on 2×H200 GPUs. The resulting model uses 9–28% fewer thinking tokens, and this transfers to out-of-domain tasks.
  • Limitations:
    • The method requires explicit </think> delimiters and access to logits, so it does not work with black-box APIs.
    • τ may need tuning.
    • It may stop early when the model is confidently wrong.
  • Code: github.com/holi-lab/ThinkBrake

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

🎓Seoul National UniversitySeoul · South Korea41 works · 45,671 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 Sangjun Song, Minjae Oh, Seungkyu Lee, Sungmin Jo and Yohan Jo. 2026. ThinkBrake: Efficient Reasoning via Log-Probability Margin Guided Decoding. 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.1095

@inproceedings{song2026thinkbrake,
  title = {{ThinkBrake: Efficient Reasoning via Log-Probability Margin Guided Decoding}},
  author = {Sangjun Song, Minjae Oh, Seungkyu Lee, Sungmin Jo and Yohan Jo},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1095},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1095}
}

References

42 references, 20 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, and 1 others. 2025. Phi-4-reasoning technical report. arXiv preprint arXiv:2504.21318. find
  2. Pranjal Aggarwal and Sean Welleck. 2025. L1: Controlling how long a reasoning model thinks with reinforcement learning. In Second Conference on Language Modeling.
  3. Art of Problem Solving. 2025. AIME problems and solutions. Accessed: 2026-1-02. find
  4. 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. arXiv preprint arXiv:2503.09567.
  5. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. Preprint, arXiv:1803.05457.
  6. 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.
  7. 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
  8. Mehdi Fatemi, Banafsheh Rafiee, Mingjie Tang, and Kartik Talamadupula. 2025. Concise reasoning via reinforcement learning. arXiv preprint arXiv:2504.05185. — [arXiv:2504.05185](https://arxiv.org/abs/2504.05185) find
  9. Sicheng Feng, Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2025. Efficient reasoning models: A survey. Transactions on Machine Learning Research.
  10. Yichao Fu, Junda Chen, Yonghao Zhuang, Zheyu Fu, Ion Stoica, and Hao Zhang. 2025. Reasoning without self-doubt: More efficient chain-of-thought through certainty probing. In ICLR 2025 Workshop on Foundation Models in the Wild. find
  11. Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, and 1 others. 2024. Omni-math: A universal olympiad level mathematic benchmark for large language models. arXiv preprint arXiv:2410.07985.
  12. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. find
  13. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  14. 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).
  15. 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 International Conference on Learning Representations.
  16. James Y. Huang, Sailik Sengupta, Daniele Bonadiman, Yi-An Lai, Arshit Gupta, Nikolaos Pappas, Saab Mansour, Katrin Kirchhoff, and Dan Roth. 2025. DeAL: Decoding-time alignment for large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volum…
  17. Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhenqiang Gong, and Lichao Sun. 2024. Metatool benchmark for large language models: Deciding whether to use tools and which to use. In The Twelfth International Conference on Learning Representati… find
  18. Guochao Jiang, Guofeng Quan, Zepeng Ding, Ziqin Luo, Dixuan Wang, and Zheng Hu. 2025. Flashthink: An early exit method for efficient reasoning. arXiv preprint arXiv:2505.13949. — [arXiv:2505.13949](https://arxiv.org/abs/2505.13949) find
  19. 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. find
  20. Tomasz Korbak, Hady Elsahar, Germán Kruszewski, and Marc Dymetman. 2022. On reinforcement learning and distribution matching for fine-tuning language models with no catastrophic forgetting. Advances in Neural Information Processing Systems, 35:16203–16220. find
  21. 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
  22. Gengyang Li, Yifeng Gao, Yuming Li, and Yunfang Wu. 2025. Thinkless: A training-free inference-efficient method for reducing reasoning redundancy. arXiv preprint arXiv:2505.15684. — [arXiv:2505.15684](https://arxiv.org/abs/2505.15684) find
  23. Alisa Liu, Xiaochuang Han, Yizhong Wang, Yulia Tsvetkov, Yejin Choi, and Noah A. Smith. 2024a. Tuning language models by proxy. In First Conference on Language Modeling. find
  24. Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A. Smith, and Yejin Choi. 2021. DExperts: Decoding-time controlled text generation with experts and anti-experts. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th I…
  25. Tianlin Liu, Shangmin Guo, Leonardo Bianco, Daniele Calandriello, Quentin Berthet, Felipe Llinares-López, Jessica Hoffmann, Lucas Dixon, Michal Valko, and Mathieu Blondel. 2024b. Decoding-time realignment of language models. In Forty-first International Conference on Machine Learning. find
  26. 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, Suzhou, China. Association for Computational Linguistics.
  27. Minjia Mao, Bowen Yin, Yu Zhu, and Xiao Fang. 2025. Early stopping chain-of-thoughts in large language models. arXiv preprint arXiv:2509.14004. — [arXiv:2509.14004](https://arxiv.org/abs/2509.14004) find
  28. Eric Mitchell, Rafael Rafailov, Archit Sharma, Chelsea Finn, and Christopher D Manning. 2024. An emulator for fine-tuning large language models using small language models. In The Twelfth International Conference on Learning Representations.
  29. OpenAI. 2024. Learning to reason with llms. Associated Press (AP). find
  30. Shishir G Patil, Huanzhi Mao, Fanjia Yan, Charlie Cheng-Jie Ji, Vishnu Suresh, Ion Stoica, and Joseph E. Gonzalez. 2025. The berkeley function calling leaderboard (BFCL): From tool use to agentic evaluation of large language models. In Forty-second International Conference on Machine Learning. find
  31. 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.
  32. 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.
  33. Mingyang Song and Mao Zheng. 2025. Walk before you run! concise llm reasoning via reinforcement learning. arXiv preprint arXiv:2505.21178. the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3613–3635, Albuquerque, New Mex… — [arXiv:2505.21178](https://arxiv.org/abs/2505.21178) find
  34. 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. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 7459–7482, Suzhou, China. Associati… find
  35. 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 Advances in Neural Information Processing Systems.
  36. Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, and 1 others. 2025. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686. 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. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  38. Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Minghui Chen, Zheng Lin, and Weiping Wang. 2026. Dynamic early exit in reasoning models. In The Fourteenth International Conference on Learning Representations. find
  39. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, YuYue, Weinan Dai, Tiantian Fan, Gaohong Liu, Juncai Liu, LingJun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, and 17 others. 2025. DAPO: An open-source LLM reinforcement learning sy…
  40. Xuechen Zhang, Zijian Huang, Chenshun Ni, Ziyang Xiong, Jiasi Chen, and Samet Oymak. 2025a. Making small language models efficient reasoners: Intervention, supervision, reinforcement. In ICML 2025 Workshop on Long-Context Foundation Models. find
  41. Yunxiang Zhang, Muhammad Khalifa, Lechen Zhang, Xin Liu, Ayoung Lee, Xinliang Frederick Zhang, Farima Fatahi Bayat, and Lu Wang. 2025b. Logit arithmetic elicits long reasoning capabilities without training. arXiv preprint arXiv:2510.09354.
  42. Guangya Wan, Yuqi Wu, Jie Chen, and Sheng Li. 2025. Reasoning aware self-consistency: Leveraging reasoning paths for efficient LLM sampling. In Proceedings of the 2025 Conference of the Nations of

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/song-2026-thinkbrake-efficient-reasoning-log-probability/