---
title: "When More Thinking Hurts — paper by Zhou et al. (2026)"
description: "The authors propose a cost-aware utility metric that penalizes compute use. 76.3% precision. 26 references, cited by 2 papers here."
authors: ["Shu Zhou", "Rui Ling", "Junan Chen", "Xin Wang", "Tao Fan", "Hao Wang"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.1199/"
doi: "10.18653/v1/2026.findings-acl.1199"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/zhou-2026-when-more-thinking-hurts-overthinking/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhou-2026-when-more-thinking-hurts-overthinking/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/9/) / When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling

# When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling

**Authors:** Shu Zhou, [Rui Ling](https://md.research.elamur.ai/free/en/authors/rui-ling/), Junan Chen, [Xin Wang](https://md.research.elamur.ai/free/en/authors/xin-wang/), [Tao Fan](https://md.research.elamur.ai/free/en/authors/tao-fan/) and [Hao Wang](https://md.research.elamur.ai/free/en/authors/hao-wang/) · **Year:** 2026 · **Venue:** [Findings of the Association for Computational Linguistics: ACL 2026](https://md.research.elamur.ai/free/en/venues/findings/) ([Findings of the Association for Computational Linguistics 2026](https://md.research.elamur.ai/free/en/venues/findings/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.findings-acl.1199/) · [PDF](https://aclanthology.org/2026.findings-acl.1199.pdf) · [DOI 10.18653/v1/2026.findings-acl.1199](https://doi.org/10.18653/v1/2026.findings-acl.1199) · [OpenAlex](https://openalex.org/W7166849418)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · 2 in our library, of the 149,439 papers whose bibliographies we read

## Summary
Test-time scaling research usually assumes that longer reasoning always gives better answers. This paper tests that assumption directly. It measures how much each extra 500 reasoning tokens adds to accuracy. Two open-weight 32B models, DeepSeek-R1-32B and s1-32B, are run under budget forcing at 32 budgets from 500 to 16,000 tokens. The benchmarks are AIME 2024/2025, MATH-500 (split by difficulty level) and GPQA Diamond.

The authors follow each problem as its budget grows and record "flip events", where an answer turns from wrong to right or from right to wrong. They also track textual signs of overthinking, such as hesitation words and answers that keep changing. Gains shrink as budgets grow and eventually become negative: longer reasoning can push a model to drop an answer it already had right. Easy problems reach this point much sooner than hard ones.

The authors propose a cost-aware utility metric that penalizes compute use. They argue that papers should report efficiency frontiers, not only accuracy curves. Moderate, difficulty-adaptive budgets can cut compute substantially at little cost to accuracy.

## Key points
- **Accuracy peaks before the maximum budget.** On AIME, R1-32B goes from 28.2% at 500 tokens to a peak of 55.8% at 12K, then falls to 54.9% at 16K. Marginal utility drops below zero past roughly 12K for both models. s1-32B degrades earlier than R1-32B.
- **Harmful flips take over at about 7K tokens.** Beyond this point, right-to-wrong flips outnumber wrong-to-right ones. MATH-500 Level 1–2 problems cross this threshold at around 2K tokens, while Level 5 problems cross it at around 8K.
- **The overthinking signals are predictive.** Combined indicators predict negative flips with 76.3% precision at 80% recall. A manual review of 80 negative flips found that 67.5% were genuine reconsideration of a correct answer, and 12.5% were degradation artifacts.
- **The effect is not limited to math or to budget forcing.** On GPQA Diamond, accuracy peaks near 10K tokens. In 312 samples where R1-32B naturally reasoned past 8K tokens, the flip ratio was 1.31, close to the 1.42 seen under forced continuation.
- **Cost-aware metric shows large savings.** The metric is U = Acc − λ·t/t_max.
  - At λ=0.5, stopping near 6K tokens saves about 50% of compute for about a 6% accuracy loss.
  - At λ=1.0, the best budget is around 2K tokens.
  - Early stopping based on the indicators keeps 97% of peak accuracy while using 60% of the compute.

*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

- [Baidu](https://md.research.elamur.ai/free/en/companies/baidu/) — Company, Beijing, China · 74 works · 15,263 citations
- [Nanjing University](https://md.research.elamur.ai/free/en/universities/nanjing-university/) — University, Nanjing, China · 74 works · 7,814 citations

Also on the byline: [Nanjing University of Finance and Economics](https://ror.org/031y8am81).

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works

## Cite

Cite as: Shu Zhou, Rui Ling, Junan Chen, Xin Wang, Tao Fan and Hao Wang. 2026. When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{zhou2026when,
  title = {{When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling}},
  author = {Shu Zhou, Rui Ling, Junan Chen, Xin Wang, Tao Fan and Hao Wang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1199},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1199}
}
```

- [← previous When Flores Bloomz Wrong: Cross-Direction Contamination in Machine Translation Evaluation](https://md.research.elamur.ai/free/en/papers/tan-2026-when-flores-bloomz-wrong-cross/)
- [next → When Seeing Overrides Knowing: Disentangling Knowledge Conflicts in Vision-Language Models](https://md.research.elamur.ai/free/en/papers/ortu-2026-when-seeing-overrides-knowing-disentangling/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/9/)

## References

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

1. 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](https://md.research.elamur.ai/free/en/papers/chen-2025-towards-reasoning-era-survey-long/). arXiv preprint arXiv:2503.09567.
2. 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. 2024a. [Do NOT think that much for 2+3=? on the overthinking of o1-like LLMs](https://md.research.elamur.ai/free/en/papers/chen-2024-think-that-much-2-3/). Preprint, arXiv:2412.21187.
3. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024b. [Do not think that much for 2+ 3=?](https://md.research.elamur.ai/free/en/papers/chen-2024-think-that-much-2-3/) on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187.
4. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. [DeepSeek-R1: Incentivizing Reasoning Capability in LL…](https://md.research.elamur.ai/free/en/papers/deepseek-ai-2025-deepseek-r1-incentivizing-reasoning-capability/)
5. Yonatan Geifman and Ran El-Yaniv. 2017. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 4885–4894, Red Hook, NY, USA. Curran Associates Inc. [find](https://scholar.google.com/scholar?q=Yonatan+Geifman+and+Ran+El%2DYaniv%2E+2017%2E+Selective+classification+for+deep+neural+networks%2E+In+Proceedings+of+the+31st+International+Conference+on+Neural+Information+Processing+Systems%2C+NIPS%E2%80%9917%2C+page+4)
6. Soumya Suvra Ghosal, Souradip Chakraborty, Avinash Reddy, Yifu Lu, Mengdi Wang, Dinesh Manocha, Furong Huang, Mohammad Ghavamzadeh, and Amrit Singh Bedi. 2025. Does thinking more always help? understanding test-time scaling in reasoning models. arXiv preprint arXiv:2506.04210, 2. — [arXiv:2506.04210](https://arxiv.org/abs/2506.04210) [find](https://scholar.google.com/scholar?q=Soumya+Suvra+Ghosal%2C+Souradip+Chakraborty%2C+Avinash+Reddy%2C+Yifu+Lu%2C+Mengdi+Wang%2C+Dinesh+Manocha%2C+Furong+Huang%2C+Mohammad+Ghavamzadeh%2C+and+Amrit+Singh+Bedi%2E+2025%2E+Does+thinking+more+always+help%3F+understa)
7. 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](https://md.research.elamur.ai/free/en/papers/hendrycks-2021-measuring-mathematical-problem-solving-math/). Advances in Neural Information Processing Systems, 34:28304– 28318.
8. William Jurayj, Jeffrey Cheng, and Benjamin Van Durme. 2025. [Is that your final answer? test-time scaling improves selective question answering](https://md.research.elamur.ai/free/en/papers/jurayj-2025-is-that-your-final-answer/). Preprint, arXiv:2502.13962.
9. Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective question answering under domain shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684– 5696, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Amita+Kamath%2C+Robin+Jia%2C+and+Percy+Liang%2E+2020%2E+Selective+question+answering+under+domain+shift%2E+In+Proceedings+of+the+58th+Annual+Meeting+of+the+Association+for+Computational+Linguistics%2C+pages+5684%E2%80%93)
10. Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023. Fast inference from transformers via speculative decoding. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 19274–19286. PMLR. [find](https://scholar.google.com/scholar?q=Yaniv+Leviathan%2C+Matan+Kalman%2C+and+Yossi+Matias%2E+2023%2E+Fast+inference+from+transformers+via+speculative+decoding%2E+In+Proceedings+of+the+40th+International+Conference+on+Machine+Learning%2C+volume+202+of)
11. Jianglin Lu, Hailing Wang, Yi Xu, Yizhou Wang, Kuo Yang, and Yun Fu. 2025. Representation potentials of foundation models for multimodal alignment: A survey. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 16680–16695. [find](https://scholar.google.com/scholar?q=Jianglin+Lu%2C+Hailing+Wang%2C+Yi+Xu%2C+Yizhou+Wang%2C+Kuo+Yang%2C+and+Yun+Fu%2E+2025%2E+Representation+potentials+of+foundation+models+for+multimodal+alignment%3A+A+survey%2E+In+Proceedings+of+the+2025+Conference+on+E)
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](https://md.research.elamur.ai/free/en/papers/muennighoff-2025-s1-simple-test-time-scaling/). arXiv preprint. ArXiv:2501.19393 [cs].
13. OpenAI. 2024. Learning to reason with LLMs. [https://openai.com/index/](https://openai.com/index/) learning-to-reason-with-llms/. Accessed: 2024-09-12. [↗](https://openai.com/index/)
14. 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](https://md.research.elamur.ai/free/en/papers/rein-2023-gpqa-graduate-level-google-proof/). In First Conference on Language Modeling.
15. Roy Schwartz, Gabriel Stanovsky, Swabha Swayamdipta, Jesse Dodge, and Noah A. Smith. 2020. [The right tool for the job: Matching model and instance complexities](https://md.research.elamur.ai/free/en/papers/schwartz-2020-right-tool-job-matching-model/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6640–6651. Association for Computational…
16. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. [Scaling llm test-time compute optimally can be more effective than scaling model parameters](https://md.research.elamur.ai/free/en/papers/snell-2024-scaling-llm-test-time-compute/). Preprint, arXiv:2408.03314. learning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 24011–24028.
17. Shu Zhou, Rui Zhao, Zhengda Zhou, Haohan Yi, Xuhui Zheng, and Hao Wang. 2025c. Enhancing extractive question answering in multiparty dialogues with logical inference memory network. In Proceedings of the 31st International Conference on Computational Linguistics, pages 8725–8738. [find](https://scholar.google.com/scholar?q=Shu+Zhou%2C+Rui+Zhao%2C+Zhengda+Zhou%2C+Haohan+Yi%2C+Xuhui+Zheng%2C+and+Hao+Wang%2E+2025c%2E+Enhancing+extractive+question+answering+in+multiparty+dialogues+with+logical+inference+memory+network%2E+In+Proceedings+of+)
18. Zhengda Zhou and Shu Zhou. 2025. Reasoning-guided prompt learning with historical knowledge injection for ancient chinese relation extraction. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 172–184. Springer. [find](https://scholar.google.com/scholar?q=Zhengda+Zhou+and+Shu+Zhou%2E+2025%2E+Reasoning%2Dguided+prompt+learning+with+historical+knowledge+injection+for+ancient+chinese+relation+extraction%2E+In+CCF+International+Conference+on+Natural+Language+Proce)
19. Gaurav Srivastava, Aafiya Hussain, Sriram Srinivasan, and Xuan Wang. 2025. [Do llms overthink basic math reasoning? benchmarking the accuracy-efficiency tradeoff in language models](https://md.research.elamur.ai/free/en/papers/srivastava-2026-do-llms-overthink-basic-math/). arXiv preprint arXiv:2507.04023.
20. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, Hanjie Chen, and Xia Hu. 2025. [Stop overthinking: A survey on efficient reasoning for large language models](https://md.research.elamur.ai/free/en/papers/sui-2025-stop-overthinking-survey-efficient-reasoning/). Transactions on Machine Learning Research.
21. Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. 2025a. Inference scaling laws: An empirical analysis of compute-optimal inference for LLM problem-solving. In The Thirteenth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Yangzhen+Wu%2C+Zhiqing+Sun%2C+Shanda+Li%2C+Sean+Welleck%2C+and+Yiming+Yang%2E+2025a%2E+Inference+scaling+laws%3A+An+empirical+analysis+of+compute%2Doptimal+inference+for+LLM+problem%2Dsolving%2E+In+The+Thirteenth+Interna)
22. Yuyang Wu, Yifei Wang, Ziyu Ye, Tianqi Du, Stefanie Jegelka, and Yisen Wang. 2025b. When more is less: Understanding chain-of-thought length in LLMs. Preprint, arXiv:2502.07266. [find](https://scholar.google.com/scholar?q=Yuyang+Wu%2C+Yifei+Wang%2C+Ziyu+Ye%2C+Tianqi+Du%2C+Stefanie+Jegelka%2C+and+Yisen+Wang%2E+2025b%2E+When+more+is+less%3A+Understanding+chain%2Dof%2Dthought+length+in+LLMs%2E+Preprint%2C+arXiv%3A2502%2E07266%2E)
23. Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, and Shyam Upadhyay. 2025. [Do llms really need 10+ thoughts for" find the time 1000 days later"?](https://md.research.elamur.ai/free/en/papers/zhang-2026-do-llms-really-need-10/) towards structural understanding of llm overthinking. arXiv preprint a…
24. Shu Zhou, Yuxuan Ao, Yunyang Xuan, Xin Wang, Tao Fan, and Hao Wang. 2026. Inference scaling law for retrieval augmented generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 16522–16530. [find](https://scholar.google.com/scholar?q=Shu+Zhou%2C+Yuxuan+Ao%2C+Yunyang+Xuan%2C+Xin+Wang%2C+Tao+Fan%2C+and+Hao+Wang%2E+2026%2E+Inference+scaling+law+for+retrieval+augmented+generation%2E+In+Proceedings+of+the+AAAI+Conference+on+Artificial+Intelligence%2C+vo)
25. Shu Zhou, Xin Wang, Jingwen Qiu, Xiaomin Li, Bin Shi, and Hao Wang. 2025a. Losdf: a logical optimization and semantic decoupling framework for question answering in multi-party conversations. Information Processing & Management, 62(5):104200. [find](https://scholar.google.com/scholar?q=Shu+Zhou%2C+Xin+Wang%2C+Jingwen+Qiu%2C+Xiaomin+Li%2C+Bin+Shi%2C+and+Hao+Wang%2E+2025a%2E+Losdf%3A+a+logical+optimization+and+semantic+decoupling+framework+for+question+answering+in+multi%2Dparty+conversations%2E+Informat)
26. Shu Zhou, Yunyang Xuan, Yuxuan Ao, Xin Wang, Tao Fan, and Hao Wang. 2025b. Merit: Multi-agent collaboration for unsupervised time series representation [find](https://scholar.google.com/scholar?q=Shu+Zhou%2C+Yunyang+Xuan%2C+Yuxuan+Ao%2C+Xin+Wang%2C+Tao+Fan%2C+and+Hao+Wang%2E+2025b%2E+Merit%3A+Multi%2Dagent+collaboration+for+unsupervised+time+series+representation)

*Updated 31 October 2026*