Elamur Research⌕
☰
home / papers / 2026 / Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers

Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers

Paper · 2026 · Van Yang, Shouren Wang, Debargha Ganguly et al.

Summary

Hybrid reasoning models such as Qwen3 are usually switched between Think and No-Think behavior with high-level instructions. This paper argues that the switch is actually controlled by a few specific tokens. Attention analysis on Qwen3-8B shows that generated tokens focus heavily on the opening "Okay" when the model reasons. In No-Think generation, attention instead concentrates on the "\n\n" that follows </think>. Controlled prompt variants evaluated on MATH500, AIME and GPQA confirm this pattern.

Building on this, the authors propose Mid-Think, a prompt format that needs no training. It places both cues together, the suppressing one followed by the activating one, so the model reasons at a medium budget. Mid-Think matches or beats Qwen3's fixed-token and prompt-based budget control on the accuracy–length trade-off.

Used as the generation mode during GRPO training after SFT, Mid-Think shortens training and yields stronger models. The result matters because it offers a cheap, interpretable way to control reasoning cost both at inference time and during RL.

Key points

  • Mid-Think format: <think>\n\n</think>\n\n<reason>\nOkay…. The wrapper tag itself is interchangeable; the authors test <reason>, <begin> and <less think>.
  • Effect of a leading "Okay" (Qwen3-8B, MATH500): without any <think> tag, the model recovers Think-level accuracy (94.1% vs 94.6%), compared with 83.2% for No-Think.
  • Conflicting cues: No-Think followed by "Okay" lands in between the two modes. For example, AIME accuracy is 55.3%, versus 71.6% for Think and 21.3% for No-Think.
  • RL training with GRPO on Qwen3-8B: compared with Think-mode training, Mid-Think training raises Think-mode AIME accuracy from 69.8% to 72.4% and GPQA from 58.5% to 61.1%. It also cuts training time from 54 to 46 hours, roughly 15% less. Qwen3-4B was tested as well.
  • Limitations: the method cannot target arbitrary budget levels. It also depends on overfitted trigger tokens that must be identified beforehand.

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

🎓Case Western Reserve UniversityCleveland · United States5 works · 280 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 Van Yang, Shouren Wang, Debargha Ganguly, Xinpeng Li, Chaoda Song, Vikash Singh, Vipin Chaudhary and Xiaotian Han. 2026. Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers. 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.299

@inproceedings{yang2026mid,
  title = {{Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers}},
  author = {Van Yang, Shouren Wang, Debargha Ganguly, Xinpeng Li, Chaoda Song, Vikash Singh, Vipin Chaudhary and Xiaotian Han},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.299},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.299}
}

References

52 references, 17 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. Dibyanayan Bandyopadhyay, Soham Bhattacharjee, and Asif Ekbal. 2025. Thinking machines: A survey of llm based reasoning strategies. arXiv preprint arXiv:2503.10814. find
  3. 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.
  4. Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2025. Thinkless: Llm learns when to think. arXiv preprint arXiv:2505.13379. — [arXiv:2505.13379](https://arxiv.org/abs/2505.13379) find
  5. Debargha Ganguly, Sreehari Sankar, Biyao Zhang, Vikash Singh, Kanan Gupta, Harshini Kavuru, Alan Luo, Weicong Chen, Warren Morningstar, Raghu Machiraju, and 1 others. 2026. Trust the typical. arXiv preprint arXiv:2602.04581. find
  6. Xiangming Gu, Tianyu Pang, Chao Du, Qian Liu, Fengzhuo Zhang, Cunxiao Du, Ye Wang, and Min Lin. 2024. When attention sink emerges in language models: An empirical view. arXiv preprint arXiv:2410.10781. — [arXiv:2410.10781](https://arxiv.org/abs/2410.10781) find
  7. 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.
  8. Zhezheng Hao, Hong Wang, Haoyang Liu, Jian Luo, Jiarui Yu, Hande Dong, Qiang Lin, Can Wang, and Jiawei Chen. 2025. Rethinking entropy interventions in rlvr: An entropy change perspective. arXiv preprint arXiv:2510.10150.
  9. Jujie He, Jiacai Liu, Chris Yuhao Liu, Rui Yan, Chaojie Wang, Peng Cheng, Xiaoyu Zhang, Fuxiang Zhang, Jiacheng Xu, Wei Shen, Siyuan Li, Liang Zeng, Tianwen Wei, Cheng Cheng, Bo An, Yang Liu, and Yahui Zhou. 2025a. Skywork open reasoner 1 technical report. arXiv preprint arXiv:2505.22312. find
  10. Jujie He, Jiacai Liu, Chris Yuhao Liu, Rui Yan, Chaojie Wang, Peng Cheng, Xiaoyu Zhang, Fuxiang Zhang, Jiacheng Xu, Wei Shen, Siyuan Li, Liang Zeng, Tianwen Wei, Cheng Cheng, Yang Liu, and Yahui Zhou. 2025b. Skywork open reasoner series. Notion Blog. find
  11. Hugging Face. 2025. Open r1: A fully open reproduction of deepseek-r1. find
  12. 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.
  13. Lingjie Jiang, Xun Wu, Shaohan Huang, Qingxiu Dong, Zewen Chi, Li Dong, Xingxing Zhang, Tengchao Lv, Lei Cui, and Furu Wei. 2025. Think only when you need with large hybrid-reasoning models. arXiv preprint arXiv:2505.14631. — [arXiv:2505.14631](https://arxiv.org/abs/2505.14631) find
  14. Feiyang Kang, Michael Kuchnik, Karthik Padthe, Marin Vlastelica, Ruoxi Jia, Carole-Jean Wu, and Newsha Ardalani. 2025. Quagmires in sft-rl post-training: When high sft scores mislead and what to use instead. arXiv preprint arXiv:2510.01624. find
  15. 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. 2025. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419.
  16. Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, and 1 others. 2024. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434. find
  17. Keliang Liu, Dingkang Yang, Ziyun Qian, Weijie Yin, Yuchi Wang, Hongsheng Li, Jun Liu, Peng Zhai, Yang Liu, and Lihua Zhang. 2025a. Reinforcement learning meets large language models: A survey of advancements and applications across the llm lifecycle. arXiv preprint arXiv:2509.16679. — [arXiv:2509.16679](https://arxiv.org/abs/2509.16679) find
  18. 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. arXiv preprint arXiv:2503.20783. 6033 10
  19. Ivan Moshkov, Darragh Hanley, Ivan Sorokin, Shubham Toshniwal, Christof Henkel, Benedikt Schifferer, Wei Du, and Igor Gitman. 2025. Aimo-2 winning solution: Building state-of-the-art mathematical reasoning models with openmathreasoning dataset. arXiv preprint arXiv:2504.16891. find
  20. Gongrui Nan, Siye Chen, Jing Huang, Mengyu Lu, Dexun Wang, Chunmei Xie, Weiqi Xiong, Xianzhou Zeng, Qixuan Zhou, Yadong Li, and 1 others. 2025. Ngrpo: Negative-enhanced group relative policy optimization. arXiv preprint arXiv:2509.18851. — [arXiv:2509.18851](https://arxiv.org/abs/2509.18851) find
  21. NVIDIA, :, Aarti Basant, Abhijit Khairnar, Abhijit Paithankar, Abhinav Khattar, Adithya Renduchintala, Aditya Malte, Akhiad Bercovich, and 1 others. 2025. Nvidia nemotron nano 2: An accurate and efficient hybrid mamba-transformer reasoning model. Preprint, arXiv:2508.14444.
  22. Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. 2024. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511.
  23. 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.
  24. Brian S Robinson, Nathan Drenkow, Colin Conwell, and Michael Bonner. 2024. A sparse null code emerges in deep neural networks. In Proceedings of UniReps: The First Workshop on Unifying Representations in Neural Models, pages 302–314. PMlR. find
  25. 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.
  26. Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. 2024. Massive activations in large language models. arXiv preprint arXiv:2402.17762. find
  27. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025a. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  28. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, and 1 others. 2025b. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599. — Kimi k1.5: Scaling Reinforcement Learning with LLMs
  29. NovaSky Team. 2025. Think less, achieve more: Cut reasoning costs by 50 https://novasky-ai.github.io/posts/reduce-overthinking. Accessed: 2025-01-23. ↗
  30. Chenlong Wang, Yuanning Feng, Dongping Chen, Zhaoyang Chu, Ranjay Krishna, and Tianyi Zhou. 2025a. Wait, we don’t need to" wait"! removing thinking tokens improves reasoning efficiency. arXiv preprint arXiv:2506.08343. find
  31. Shouren Wang, Wang Yang, Xianxuan Long, Qifan Wang, Vipin Chaudhary, and Xiaotian Han. 2025b. Demystifying hybrid thinking: Can llms truly switch between think and no-think? arXiv preprint arXiv:2510.12680. — [arXiv:2510.12680](https://arxiv.org/abs/2510.12680) find
  32. Yihong Wu, Liheng Ma, Lei Ding, Muzhi Li, Xinyu Wang, Kejia Chen, Zhan Su, Zhanguang Zhang, Chenyang Huang, Yingxue Zhang, and 1 others. 2025. It takes two: Your grpo is secretly dpo. arXiv preprint arXiv:2510.00977. — [arXiv:2510.00977](https://arxiv.org/abs/2510.00977) find
  33. Zhiheng Xi, Xin Guo, Yang Nan, Enyu Zhou, Junrui Shen, Wenxiang Chen, Jiaqi Liu, Jixuan Huang, Zhihao Zhang, Honglin Guo, and 1 others. 2025. Bapo: Stabilizing off-policy reinforcement learning for llms via balanced policy optimization with adaptive clipping. arXiv preprint arXiv:2510.18927. — [arXiv:2510.18927](https://arxiv.org/abs/2510.18927) find
  34. Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. Preprint, arXiv:2502.12067.
  35. Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2023. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453. find
  36. Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, and 1 others. 2025a. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686. find
  37. Xin Xu, Cliveb AI, Kai Yang, Tianhao Chen, Yang Wang, Saiyong Yang, and Can Yang. 2025b. Thinking-free policy initialization makes distilled reasoning models more effective and efficient reasoners. arXiv preprint arXiv:2509.26226. — [arXiv:2509.26226](https://arxiv.org/abs/2509.26226) find
  38. Yuhui Xu, Hanze Dong, Lei Wang, Doyen Sahoo, Junnan Li, and Caiming Xiong. 2025c. Scalable chain of thoughts via elastic reasoning. arXiv preprint arXiv:2505.05315. — [arXiv:2505.05315](https://arxiv.org/abs/2505.05315) find
  39. 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.
  40. Rubing Yang, Huajun Bai, Song Liu, Guanghua Yu, Runzhi Fan, Yanbin Dang, Jiejing Zhang, Kai Liu, Jianchen Zhu, and Peng Chen. 2025b. Specexit: Accelerating large reasoning model via speculative exit. arXiv preprint arXiv:2509.24248. 6034 11 — [arXiv:2509.24248](https://arxiv.org/abs/2509.24248) find
  41. Van Yang, Hongye Jin, Shaochen Zhong, Song Jiang, Qifan Wang, Vipin Chaudhary, and Xiaotian Han. 2025c. 100-longbench: Are de facto long-context benchmarks literally evaluating long-context ability? In Findings of the Association for Computational Linguistics: ACL 2025, pages 17560–17576. find
  42. Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, and 1 others. 2025. Group sequence policy optimization. arXiv preprint arXiv:2507.18071.
  43. Van Yang, Xiang Yue, Vipin Chaudhary, and Xiaotian Han. 2025d. Speculative thinking: Enhancing small-model reasoning with large model guidance at inference time. In Second Conference on Language Modeling. find
  44. Wang Yang, Zirui Liu, Hongye Jin, Qingyu Yin, Vipin Chaudhary, and Xiaotian Han. 2025e. Longer context, deeper thinking: Uncovering the role of long-context ability in reasoning. arXiv preprint arXiv:2505.17315. — [arXiv:2505.17315](https://arxiv.org/abs/2505.17315) find
  45. Wang Yang, Chaoda Song, Xinpeng Li, Debargha Ganguly, Chuang Ma, Shouren Wang, Zhihao Dou, Yuli Zhou, Vipin Chaudhary, and Xiaotian Han. 2026. Ace-bench: Agent configurable evaluation with scalable horizons and controllable difficulty under lightweight environments. arXiv preprint arXiv:2604.06111. — [arXiv:2604.06111](https://arxiv.org/abs/2604.06111) find
  46. Wang Yang, Xiang Yue, Vipin Chaudhary, and Xiaotian Han. 2025f. 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
  47. 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.
  48. Linan Yue, Yichao Du, Yizhi Wang, Weibo Gao, Fangzhou Yao, Li Wang, Ye Liu, Ziyu Xu, Qi Liu, Shimin Di, and 1 others. 2025. Don’t overthink it: A survey of efficient r1-style large reasoning models. arXiv preprint arXiv:2508.02120. find
  49. Charlie Zhang, Graham Neubig, and Xiang Yue. 2025a. On the interplay of pre-training, mid-training, and rl on reasoning language models. arXiv preprint arXiv:2512.07783. — [arXiv:2512.07783](https://arxiv.org/abs/2512.07783) find
  50. Jintian Zhang, Yuqi Zhu, Mengshu Sun, Yujie Luo, Shuofei Qiao, Lun Du, Da Zheng, Huajun Chen, and Ningyu Zhang. 2025b. Lightthinker: Thinking step-by-step compression. arXiv preprint arXiv:2502.15589.
  51. Kaiyan Zhang, Yuxin Zuo, Bingxiang He, Youbang Sun, Runze Liu, Che Jiang, Yuchen Fan, Kai Tian, Guoli Jia, Pengfei Li, and 1 others. 2025c. A survey of reinforcement learning for large reasoning models. arXiv preprint arXiv:2509.08827. — [arXiv:2509.08827](https://arxiv.org/abs/2509.08827) find
  52. Yiqun Zhang, Hao Li, Jianhao Chen, Hangfan Zhang, Peng Ye, Lei Bai, and Shuyue Hu. 2025d. Beyond gpt-5: Making llms cheaper and better via performance-efficiency optimized routing. In Proceedings of the 2025 7th International Conference on Distributed Artificial Intelligence, pages 122–129. 6035 12 find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-mid-think-training-free-intermediate/