Elamur Research⌕
☰
home / papers / 2026 / Thermometer of Thoughts: Enhancing LLM's Exploration via Attention Temperature Modulation

Thermometer of Thoughts: Enhancing LLM's Exploration via Attention Temperature Modulation

Paper · 2026 · Zhiyuan Yu, Shijian Xiao, Cam-Tu Nguyen et al.

Summary

The paper tackles weak exploration in LLM reasoning. Standard stochastic decoding only samples from the model's next-token distribution, so the model keeps reasoning along the same habitual lines. The authors instead change the attention temperature, the factor that sets how sharply attention weights concentrate on particular key states. This alters the model's internal focus rather than just its output probabilities.

Their experiments show that moderate shifts away from the default value leave reasoning ability intact. The useful direction depends on problem difficulty. Hotter attention spreads focus and helps on hard problems by widening the search. Cooler attention sharpens retrieval and reduces overthinking on easy ones.

Building on this, they propose a two-stage inference procedure:

  1. Sample reasoning traces under several attention temperatures.
  2. Merge the candidates with a weighted vote that takes estimated difficulty into account.

Across 7 reasoning benchmarks, this raises Pass@10 by 6.78–14.20% and aggregation accuracy by 9.74%. The work offers a training-free way to scale test-time compute, but it requires access to the model's internals.

Key points

  • Main experiments use small Qwen3 models (Qwen3-0.6B, Qwen3-1.7B), with attention temperatures around 0.9–1.1.
  • The difficulty-induced aggregation beats majority voting by 9.74% on average. Gains reach about 16.2% on competition sets (AIME, HMMT), versus 2.6% on GSM8K and 4.2% on GPQA. The best hyperparameters found were C_low = 0.45 and β = 1.25.
  • A new Continuation Tendency Score (CTS), judged by DeepSeek-R1, measures how often a reasoning chain continues linearly versus switching course. On AIME 2024, CTS falls as attention temperature rises, meaning the model reflects and tries alternatives more often. Raising decoding temperature produces no such trend.
  • Randomly varying temperature per layer or per head lowers Pass@10 (−2.37% layer-wise, a further −2.91% head-wise). A single uniform value works better.
  • Limitations: the method cannot be applied to API-only models such as GPT-5, and it was not tested on larger models like Qwen3-32B.

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

🎓Fudan UniversityShanghai · China202 works · 22,509 citations🎓Nanjing UniversityNanjing · China99 works · 8,510 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksReasoning · 1,685 works

Cite

Cite as Zhiyuan Yu, Shijian Xiao, Cam-Tu Nguyen, Zhangyue Yin, Lekai Xing, Wenzhong Li and Sanglu Lu. 2026. Thermometer of Thoughts: Enhancing LLM's Exploration via Attention Temperature Modulation. 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.200

@inproceedings{yu2026thermometer,
  title = {{Thermometer of Thoughts: Enhancing LLM's Exploration via Attention Temperature Modulation}},
  author = {Zhiyuan Yu, Shijian Xiao, Cam-Tu Nguyen, Zhangyue Yin, Lekai Xing, Wenzhong Li and Sanglu Lu},
  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.200},
  url = {https://doi.org/10.18653/v1/2026.acl-long.200}
}

References

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

  1. Vidhisha Balachandran, Jingya Chen, Lingjiao Chen, Shivam Garg, Neel Joshi, Yash Lara, John Langford, Besmira Nushi, Vibhav Vineet, Yue Wu, et al. 2025. Inference-time scaling for complex tasks: Where we stand and what lies ahead. arXiv preprint arXiv:2504.00294. find
  2. Mislav Balunović, Jasper Dekoninck, Ivo Petrov, Nikola Jovanović, and Martin Vechev. 2025. Matharena: Evaluating llms on uncontaminated math competitions.
  3. Marco Bronzini, Carlo Nicolini, Bruno Lepri, Jacopo Staiano, and Andrea Passerini. 2024. Unveiling llms: The evolution of latent representations in a dynamic knowledge graph. find
  4. Vivien Cabannes, Charles Arnal, Wassim Bouaziz, Xingyu Alice Yang, Francois Charton, and Julia Kempe. 2024. Iteration head: A mechanistic study of chain-of-thought. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. find
  5. Siyu Cao, Hangting Chen, Peng Chen, Yiji Cheng, Yutao Cui, Xinchi Deng, Ying Dong, Kipper Gong, Tianpeng Gu, Xiusen Gu, Tiankai Hang, Duojun Huang, Jie Jiang, Zhengkai Jiang, Weijie Kong, Changlin Li, Donghao Li, Junzhe Li, Xin Li, Yang Li, Zhenxi Li, Zhimin Li, Jiaxin Lin, Linus, Lucaz Liu, Shu Liu… — HunyuanImage 3.0 Technical Report
  6. Souradip Chakraborty, Mohammadreza Pourreza, Ruoxi Sun, Yiwen Song, Nino Scherrer, Furong Huang, Amrit Singh Bedi, Ahmad Beirami, Jindong Gu, Hamid Palangi, and Tomas Pfister. 2025. On the role of feedback in test-time scaling of agentic ai workflows. find
  7. Kaiyan Chang, Yonghao Shi, Chenglong Wang, Hang Zhou, Chi Hu, Xiaoqian Liu, Yingfeng Luo, Yuan Ge, Tong Xiao, and Jingbo Zhu. 2025. Step-level verifier-guided hybrid test-time scaling for large language models. find
  8. Lei Chen, Joan Bruna, and Alberto Bietti. 2025a. Distributional associations vs in-context reasoning: A study of feed-forward and attention layers. find
  9. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
  10. 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. 2025b. Do not think that much for 2+3=? on the overthinking of o1-like llms.
  11. Yaofo Chen, Zeng You, Shuhai Zhang, Haokun Li, Yirui Li, Yaowei Wang, and Mingkui Tan. 2025c. Core context aware transformers for long context language modeling. find
  12. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems.
  13. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.
  14. Mengru Ding, Hanmeng Liu, Zhizhang Fu, Jian Song, Wenbo Xie, and Yue Zhang. 2024. Break the chain: Large language models can be shortcut reasoners. find
  15. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. find
  16. 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).
  17. Zhongzhan Huang, Mingfu Liang, Jinghui Qin, Shanshan Zhong, and Liang Lin. 2023. Understanding self-attention mechanism via dynamical system perspective. find
  18. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.
  19. Mingyu Jin, Kai Mei, Wujiang Xu, Mingjie Sun, Ruixiang Tang, Mengnan Du, Zirui Liu, and Yongfeng Zhang. 2025. Massive values in self-attention modules are the key to contextual knowledge understanding. find
  20. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models.
  21. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large language models are zero-shot reasoners.
  22. Dacheng Li, Shiyi Cao, Chengkun Cao, Xiuyu Li, Shangyin Tan, Kurt Keutzer, Jiarong Xing, Joseph E Gonzalez, and Ion Stoica. 2025a. S*: Test time scaling for code generation. arXiv preprint arXiv:2502.14382. find
  23. Yang Li, Zhichen Dong, Yuhan Sun, Weixun Wang, Shaopan Xiong, Yijia Luo, Jiashun Liu, Han Lu, Jiamang Wang, Wenbo Su, Bo Zheng, and Junchi Yan. 2025b. Attention illuminates llm reasoning: The preplan-and-anchor rhythm enables fine-grained policy optimization. find
  24. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step.
  25. Jiahao Lu, Ziwei Xu, and Mohan Kankanhalli. 2025. Reasoning llms are wandering solution explorers. find
  26. 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
  27. Qian Niu, Junyu Liu, Ziqian Bi, Pohsun Feng, Benji Peng, Keyu Chen, Ming Li, Lawrence KQ Yan, Yichao Zhang, Caitlyn Heqi Yin, Cheng Fei, Tianyang Wang, Yunze Wang, Silin Chen, Ming Liu, Ziyuan Qin, Riyang Bao, Xinyuan Song, and Zekun Jiang. 2025. Large language models and cognitive science: A compre… find
  28. OpenAI. 2025. Gpt-5 is here. find
  29. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Train…
  30. Jianing Qi, Xi Ye, Hao Tang, Zhigang Zhu, and Eunsol Choi. 2025. Learning to reason across parallel samples for llm reasoning. find
  31. Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model.
  32. Dhananjay Ram, Wei Xia, and Stefano Soatto. 2025. Learning to focus: Focal attention for selective and scalable transformers. find
  33. 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.
  34. Jie Ren, Qipeng Guo, Hang Yan, Dongrui Liu, Quanshi Zhang, Xipeng Qiu, and Dahua Lin. 2024. Identifying semantic induction heads to understand in-context learning. find
  35. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models.
  36. Rylan Schaeffer, Joshua Kazdan, John Hughes, Jordan Juravsky, Sara Price, Aengus Lynch, Erik Jones, Robert Kirk, Azalia Mirhoseini, and Sanmi Koyejo. 2025. How do large language monkeys get their power (laws)? find
  37. Yuheng Wu, Azalia Mirhoseini, and Thierry Tambe. 2025a. On the role of temperature sampling in test-time scaling. find
  38. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.
  39. Haizhou Shi, Ye Liu, Bo Pang, Zeyu Leo Liu, Hao Wang, Silvio Savarese, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. 2025a. Ssr: Socratic self-refine for large language model reasoning. find
  40. Weijie Shi, Han Zhu, Jiaming Ji, Mengze Li, Jipeng Zhang, Ruiyuan Zhang, Jia Zhu, Jiajie Xu, Sirui Han, and Yike Guo. 2025b. Legalreasoner: Step-wised verification-correction for legal judgment reasoning.
  41. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling llm test-time compute optimally can be more effective than scaling model parameters. In The Thirteenth International Conference on Learning Representations. OpenReview.
  42. 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.
  43. AoPSWiki Team. 2025. Aime problems and solutions. find
  44. Shangqing Tu, Yaxuan Li, Yushi Bai, Lei Hou, and Juanzi Li. 2025. Deepprune: Parallel scaling without inter-trace redundancy.
  45. Jian Wang, Boyan Zhu, Chak Tou Leong, Yongqi Li, and Wenjie Li. 2025a. Scaling over scaling: Exploring test-time scaling pareto in large reasoning models. arXiv preprint arXiv:2505.20522. — [arXiv:2505.20522](https://arxiv.org/abs/2505.20522) find
  46. Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, Yuqiong Liu, An Yang, Andrew Zhao, Yang Yue, Shiji Song, Bowen Yu, Gao Huang, and Junyang Lin. 2025b. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcem…
  47. Xu Wang, Yan Hu, Wenyu Du, Reynold Cheng, Benyou Wang, and Difan Zou. 2025c. Towards understanding fine-tuning mechanisms of llms via circuit analysis. find
  48. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models.
  49. 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.
  50. Wei Xiong, Wenting Zhao, Weizhe Yuan, Olga Golovneva, Tong Zhang, Jason Weston, and Sainbayar Sukhbaatar. 2025. Stepwiser: Stepwise generative judges for wiser reasoning. find
  51. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  52. Zhangyue Yin, Qiushi Sun, Qipeng Guo, Zhiyuan Zeng, Xiaonan Li, Junqi Dai, Qinyuan Cheng, Xuanjing Huang, and Xipeng Qiu. 2024. Reasoning in flux: Enhancing large language models reasoning through uncertainty-aware adaptive guidance. In ACL (1), pages 2401–2416. find
  53. Zhaojian Yu, Yinghao Wu, Yilun Zhao, Arman Cohan, and Xiao-Ping Zhang. 2025. Z1: Efficient test-time scaling with code. arXiv preprint arXiv:2504.00810. find
  54. Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Wenyue Hua, Haolun Wu, Zhihan Guo, Yufei Wang, Niklas Muennighoff, et al. 2025a. A survey on test-time scaling in large language models: What, how, where, and how well? arXiv preprint arXiv:2503.24235.
  55. Zefeng Zhang, Xiangzhao Hao, Hengzhu Tang, Zhenyu Zhang, Jiawei Sheng, Xiaodong Li, Zhenyang Li, Li Gao, Daiting Shi, Dawei Yin, and Tingwen Liu. 2025b. Cooper: A unified model for cooperative perception and reasoning in spatial intelligence. find
  56. Zheng Zhang. 2025. Comprehension without competence: Architectural limits of llms in symbolic computation and reasoning. find
  57. Zifan Zheng, Yezhaohui Wang, Yuxin Huang, Shichao Song, Mingchuan Yang, Bo Tang, Feiyu Xiong, and Zhiyu Li. 2024. Attention heads of large language models: A survey.
  58. Jialun Zhong, Wei Shen, Yanzeng Li, Songyang Gao, Hua Lu, Yicheng Chen, Yang Zhang, Wei Zhou, Jinjie Gu, and Lei Zou. 2025. A comprehensive survey of reward models: Taxonomy, applications, challenges, and future. find
  59. King Zhu, Hanhao Li, Siwei Wu, Tianshun Xing, Dehua Ma, Xiangru Tang, Minghao Liu, Jian Yang, Jiaheng Liu, Yuchen Eleanor Jiang, et al. 2025. Scaling test-time compute for llm agents. arXiv preprint arXiv:2506.12928. — [arXiv:2506.12928](https://arxiv.org/abs/2506.12928) find
  60. Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. 2024. Attention temperature matters in vit-based cross-domain few-shot learning. In Advances in Neural Information Processing Systems, volume 37, pages 116332–116354. Curran Associates, Inc. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yu-2026-thermometer-thoughts-enhancing-llm-s/