Elamur Research⌕
☰
home / papers / 2026 / Thinking Long, but Short: Stable Sequential Test-Time Scaling for Large Reasoning Models

Thinking Long, but Short: Stable Sequential Test-Time Scaling for Large Reasoning Models

Paper · 2026 · Michael R. Metel, Yufei Cui, Boxing Chen et al.

Summary

Sequential test-time scaling makes a large reasoning model keep thinking, for example by swapping its end-of-thinking token for "Wait". This helps only up to a point. Beyond that point, accuracy drops and outputs become repetitive and unstable. The authors propose Min-Seek, a training-free method that keeps only a small, changing subset of earlier thoughts (reconstruction cycles) in the KV cache, preferring the shortest ones. The idea rests on prior findings that flawed reasoning tends to be long, while sound reasoning is usually more direct. A custom KV cache stores keys without position embeddings and re-encodes them contiguously before each new thought. This lets reasoning run past the model's maximum context length with linear cost under mild assumptions. On five benchmarks with two distilled DeepSeek-R1 models, Min-Seek outperforms Budget Forcing across a wide range of induced thought counts and stays stable even with unlimited thoughts. It also runs faster. The method removes the need to tune reasoning length for each model and task.

Key points

  • Models tested: DeepSeek-R1-Distill-Qwen-1.5B and DeepSeek-R1-Distill-Qwen-7B. Tasks: AIME 2024, AMC (2022–2023, integer answers), GPQA Diamond, MATH-500 and MMLU-Pro. Sampling used temperature 0.6 and top_p 0.95.
  • Induced reconstruction cycles were capped at M ∈ {0, 2, 4, 6, 10, 20, 50, 100, inf}. Accuracy is normalized to standard generation (M = 0).
  • 1.5B model:
    • Budget Forcing peaks at 1.069 (M = 2).
    • Min-Seek stays between 1.076 and 1.148 for M ≥ 6.
    • Min-Seek settles at 1.137 for M ≥ 100.
  • 7B model:
    • Budget Forcing peaks at 1.090.
    • Min-Seek leads by at least 0.045 for every M > 0 and exceeds 1.095 for M ≥ 100.
    • AMC is the main per-task exception.
  • Efficiency: only one extra induced thought is kept in the KV cache. As a result, Budget Forcing is on average 44.0% (7B) and 29.4% (1.5B) slower than Min-Seek.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 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 Michael R. Metel, Yufei Cui, Boxing Chen and Prasanna Parthasarathi. 2026. Thinking Long, but Short: Stable Sequential Test-Time Scaling for Large Reasoning Models. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-eacl.153

@inproceedings{metel2026thinking,
  title = {{Thinking Long, but Short: Stable Sequential Test-Time Scaling for Large Reasoning Models}},
  author = {Michael R. Metel, Yufei Cui, Boxing Chen and Prasanna Parthasarathi},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.153},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.153}
}

References

25 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. DeepSeek. 2025a. deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. https://huggingface.co/ deepseek-ai/DeepSeek-R1-Distill-Qwen-1. 5B. Accessed: 2025-05-30. ↗
  2. DeepSeek. 2025b. deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. https://huggingface.co/ deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. Accessed: 2025-05-01. ↗
  3. DeepSeek. 2025c. Deepseek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv preprint arXiv:2501.12948.
  4. 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 NeurIPS Track on Datasets and Benchmarks.
  5. Hugging Face. 2025a. Caching. https://huggingface.co/docs/transformers/en/ cache_explanation. Accessed: 2025-07-10. ↗
  6. Hugging Face. 2025b. Utilities for Generation: class transformers.DynamicCache. https://huggingface.co/docs/transformers/ main/en/internal/generation_utils\# transformers.DynamicCache. Accessed: 2025-07-08. ↗
  7. Hugging Face H4. 2024. HuggingFaceH4/MATH-500. https://huggingface.co/datasets/ HuggingFaceH4/MATH-500. Accessed: 2025-05-30. ↗
  8. Hugging Face H4. 2025. HuggingFaceH4/aime_2024. https://huggingface.co/datasets/ HuggingFaceH4/aime_2024. Accessed: 2025-05-30. ↗
  9. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s Verify Step by Step. In ICLR.
  10. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-Refine: Iterative Refinement with… find
  11. Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han Lù, Nicholas Meade, Dongchan Shin, Amirhossein Kazemnejad, Gaurav Kamath, Marius Mosbach, Karolina Stańczak, and Siva Reddy. 2025. DeepS… find
  12. 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. In EMNPL, pages 20275–20321.
  13. OpenAI. 2024. OpenAI o1 system card. arXiv preprint arXiv:2412.16720.
  14. Project Numina. 2024. AI-MO/aimo-validation-amc. https://huggingface.co/datasets/AI-MO/ aimo-validation-amc. Accessed: 2025-06-03. ↗
  15. David Rein. 2023. Idavidrein/gpqa. https://huggingface.co/datasets/Idavidrein/gpqa. Accessed: 2025-05-21. ↗
  16. 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 COLM.
  17. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. In NeurIPS, pages 8634–8652.
  18. Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. RoFormer: Enhanced Transformer with Rotary Position Embedding. Neurocomputing, 568.
  19. TIGER Lab. 2024. TIGER-Lab/MMLU-Pro. https://huggingface.co/datasets/ TIGER-Lab/MMLU-Pro. Accessed: 2025-06-25. ↗
  20. 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 ICLR.
  21. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. 2024. MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark…
  22. 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 NeurIPS, pages 24824–24837.
  23. Yuchen Yan, Yongliang Shen, Yang Liu, Jin Jiang, Mengdi Zhang, Jian Shao, and Yueting Zhuang. 2025. Inftythink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models. arXiv preprint arXiv:2503.06692.
  24. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate Problem Solving with Large Language Models. In NeurIPS, pages 11809– 11822. find
  25. Zhiyuan Zeng, Qinyuan Cheng, Zhangyue Yin, Yunhua Zhou, and Xipeng Qiu. 2025. Revisiting the Test-Time Scaling of o1-like Models: Do they Truly Possess Test-Time Scaling Capabilities? In ACL, pages 4651–4665.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/metel-2026-thinking-long-but-short-stable/