Elamur Research⌕
☰
home / papers / 2025 / Read Your Own Mind: Reasoning Helps Surface Self-Confidence Signals in LLMs

Read Your Own Mind: Reasoning Helps Surface Self-Confidence Signals in LLMs

Paper · 2025 · Jakub Podolak and Rajeev Verma

Summary

The paper asks where an LLM's verbalized confidence (VC) comes from. One possibility is that the model reads out a hidden internal belief. The other is that it only gains a usable sense of its own uncertainty after explicitly exploring alternative answers, which is what Semantic Entropy (SE) does by sampling many responses. The authors test this on DeepSeek R1-32B using a small QA set of fact-retrieval and math questions. They use a forced-reasoning technique to make the model think for progressively longer budgets before answering and reporting its confidence. They then compare VC with SE at matched compute.

With no reasoning, the stated confidence is barely better than chance. As the reasoning budget grows, VC approaches SE's ability to flag wrong answers, even for simple factual questions where accuracy does not change. An external reader model that sees only the reasoning trace recovers nearly the same confidence signal. The conclusion is that reliable self-reported uncertainty depends on test-time exploration rather than on the specific uncertainty heuristic. This undercuts VC's main appeal as a cheap estimate.

Key points

  • Setup: DeepSeek R1-32B was evaluated on 270 QA items: 126 fact-retrieval, 134 mathematical reasoning and 10 others. Uncertainty-estimation quality was measured as ROC-AUC for detecting incorrect answers.
  • Effect of short reasoning: A budget of only 100–500 reasoning tokens raised accuracy from 41% to 63% and VC ROC-AUC from 0.56 to 0.80. With longer budgets, ROC-AUC eventually reached about 0.88.
  • Parity with SE: VC roughly matched SE at about 200 tokens for fact retrieval and about 3,500 tokens for math. SE consumed about 218 tokens per sample on average, so the cost is comparable for factual questions.
  • External reader: GPT-4o-mini read only DeepSeek's chain of thought. At 3.4k tokens it reached ROC-AUC 0.841, versus 0.851 for DeepSeek's own VC, with |ρ| = 0.57 between the two scores. This supports the "self-sampling" view over an intrinsic latent confidence.
  • Limitations: The study covers only one model and a compact dataset. Future work could look for more efficient or structured ways for models to explore their own uncertainty.

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

🎓University of AmsterdamAmsterdam · Netherlands144 works · 173,906 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 works

Cite

Cite as Jakub Podolak and Rajeev Verma. 2025. Read Your Own Mind: Reasoning Helps Surface Self-Confidence Signals in LLMs. arXiv:2505.23845 · UncertaiNLP Workshop at Empirical Methods in Natural Language Processing 2025 (EMNLP 2025).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.23845 · arXiv 2505.23845

@misc{podolak2025read,
  title = {{Read Your Own Mind: Reasoning Helps Surface Self-Confidence Signals in LLMs}},
  author = {Jakub Podolak and Rajeev Verma},
  year = {2025},
  doi = {10.48550/arxiv.2505.23845},
  eprint = {2505.23845},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.23845}
}

References

25 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. AIME. 2024. Aime 2024 dataset. https://huggingface.co/datasets/Maxwell-Jia/ AIME_2024. Accessed: 2025-05-07. ↗
  2. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. triviaqa: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. arXiv e-prints, arXiv:1705.03551.
  3. Neil Band, Xuechen Li, Tengyu Ma, and Tatsunori Hashimoto. 2024. Linguistic calibration of long-form generations. Preprint, arXiv:2404.00474. — [arXiv:2404.00474](https://arxiv.org/abs/2404.00474) find
  4. William Jurayj, Jeffrey Cheng, and Benjamin Van Durme. 2025. Is that your final answer? test-time scaling improves selective question answering. Preprint, arXiv:2502.13962.
  5. 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. Preprint, arXiv:2110.14168.
  6. 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. Preprint, arXiv:2501.19393.
  7. 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…
  8. Shehzaad Dhuliawala, Leonard Adolphs, Rajarshi Das, and Mrinmaya Sachan. 2022. Calibration of machine reading systems at scale. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1682–1693, Dublin, Ireland. Association for Computational Linguistics. find
  9. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, and 1 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  10. S. Farquhar, J. Kossen, L. Kuhn, and Y. Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625–630. find
  11. Hasan Abed Al Kader Hammoud, Hani Itani, and Bernard Ghanem. 2025. Beyond the last answer: Your reasoning trace uncovers more than you think. Preprint, arXiv:2504.20708. — [arXiv:2504.20708](https://arxiv.org/abs/2504.20708) find
  12. Shiyu Ni, Keping Bi, Lulu Yu, and Jiafeng Guo. 2024. Are large language models more honest in their probabilistic or verbalized confidence? Preprint, arXiv:2408.09773. — [arXiv:2408.09773](https://arxiv.org/abs/2408.09773) find
  13. OpenAI. 2024. Hello, gpt-4o. Accessed: 2025-01-05. find
  14. Yudi Pawitan and Chris Holmes. 2024. Confidence in the reasoning of large language models. Preprint, arXiv:2412.15296. — [arXiv:2412.15296](https://arxiv.org/abs/2412.15296) find
  15. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test-time compute optimally can be more effective than scaling model parameters. Preprint, arXiv:2408.03314.
  16. Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D. Manning. 2023. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. Preprint, arXiv:2305.14975. find
  17. Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. 2024. Measuring short-form factuality in large language models. Preprint, arXiv:2411.04368.
  18. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Preprint, arXiv:2009.03300.
  19. 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. Preprint, arXiv:2201.11903.
  20. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. Preprint, arXiv:2311.05232.
  21. Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. 2024. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. Preprint, arXiv:2306.13063. find
  22. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2024. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems.
  23. Daniel Yang, Yao-Hung Hubert Tsai, and Makoto Yamada. 2024. On verbalized confidence scores for llms. Preprint, arXiv:2412.14737. — [arXiv:2412.14737](https://arxiv.org/abs/2412.14737) find
  24. Qingcheng Zeng, Weihao Xuan, Leyang Cui, and Rob Voigt. 2025. Do reasoning models show better verbalized calibration? Preprint, arXiv:2504.06564. — [arXiv:2504.06564](https://arxiv.org/abs/2504.06564) find
  25. Xinran Zhao, Hongming Zhang, Xiaoman Pan, Wenlin Yao, Dong Yu, Tongshuang Wu, and Jianshu Chen. 2024. Fact-and-reflection (FaR) improves confidence calibration of large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 8702–8718, Bangkok, Thailand. Associ… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/podolak-2025-read-own-mind-reasoning-helps/