Elamur Research⌕
☰
home / papers / 2026 / Confidence Should Be Calibrated More Than One Turn Deep

Confidence Should Be Calibrated More Than One Turn Deep

Paper · 2026 · Zhaohan Zhang, Chengzhengxu Li, Xiaoming Liu et al.

Summary

Research on LLM confidence calibration has mostly looked at single-turn question answering. Real deployments, however, involve dialogues where users push back. The authors define multi-turn calibration: confidence must match accuracy at every turn, given the conversation history. They introduce ECE@T, which measures calibration error at a fixed turn, and ECE@D, which aggregates over all turns.

In a pilot on TriviaQA with Llama3.1-8B-Instruct, a single persuasive follow-up made the model change answers while staying highly confident. Accuracy fell and calibration error rose sharply.

To fix this, they propose MTCal. It is a lightweight MLP probe that reads the model's hidden states and is trained with a differentiable surrogate for ECE@T, because ECE@T itself cannot be optimized directly. They also propose ConfChat, a decoding strategy. It reweights candidate tokens using calibrated confidence and combines later-turn scores with first-turn scores, helping the model decide whether to keep or revise its answer.

MTCal keeps ECE@T below 10% across turns, and ConfChat improves accuracy under persuasion. This matters for trustworthy dialogue systems in high-stakes settings.

Key points

  • Pilot result (Llama3.1-8B-Instruct, TriviaQA, length-normalized likelihood as confidence): after a challenging follow-up, ECE rose from 6.43% (turn 1) to 26.41% (turn 2), and accuracy dropped from 69.25% to 40.85%.
  • MTCal performance: MTCal keeps ECE@T consistently under 10.0%. With access to the full dialogue history, its ECE@1/2/3 is 5.04/2.31/7.24, compared with 10.45/12.87/13.72 when it sees only the current turn.
  • ConfChat performance: ConfChat outperforms the baselines Reminder Prompt (RP) and CARG. It raises first-turn accuracy by 1.30% on average and gives further gains in later turns.
  • Effect of persuasion style: follow-up messages were embedded with all-minilm-l6-v2 and grouped by hierarchical clustering. Blunt, decisive appeals caused the largest ECE increase (about 13.1 points), while mild, hesitant ones caused the smallest (about 7.8).
  • Limitations: MTCal needs white-box access to hidden states, so it cannot be applied to closed models such as the GPT series. It outputs one confidence score per turn, so long-form, multi-claim answers are out of scope. ConfChat adds decoding cost. Code is released at github.com/petezone/Multiturn-Calibration.

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

🎓Queen Mary University of LondonLondon · United Kingdom54 works · 3,542 citations🎓Xi'an Jiaotong UniversityXi'an · China22 works · 36,046 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

Cite

Cite as Zhaohan Zhang, Chengzhengxu Li, Xiaoming Liu, Chao Shen, Ziquan Liu and Ioannis Patras. 2026. Confidence Should Be Calibrated More Than One Turn Deep. 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.1787

@inproceedings{zhang2026confidence,
  title = {{Confidence Should Be Calibrated More Than One Turn Deep}},
  author = {Zhaohan Zhang, Chengzhengxu Li, Xiaoming Liu, Chao Shen, Ziquan Liu and Ioannis Patras},
  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.1787},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1787}
}

References

64 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv–2407.
  2. Zhihao Fan, Lai Wei, Jialong Tang, Wei Chen, Siyuan Wang, Zhongyu Wei, and Fei Huang. 2025. Ai hospital: Benchmarking large language models in a multi-agent medical interaction simulator. In COLING.
  3. W Brier Glenn and 1 others. 1950. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1–3. find
  4. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  5. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321–1330. PMLR. find
  6. Jaroslaw Blasiok and Preetum Nakkiran. Smooth ece: Principled reliability diagrams via kernel smoothing. In The Twelfth International Conference on Learning Representations. find
  7. Sophia Hager, David Mueller, Kevin Duh, and Nicholas Andrews. 2025. Uncertainty distillation: Teaching language models to express semantic confidence. arXiv preprint arXiv:2503.14749.
  8. Umesh Bodhwani, Yuan Ling, Shujing Dong, Yarong Feng, and Hongfei Li. 2025. A calibrated reflection approach for enhancing confidence estimation in llms. In Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025), pages 399–411. find
  9. Pengcheng He, Jianfeng Gao, and Weizhu Chen. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. In The Eleventh International Conference on Learning Representations. Yongchao Chen, Harsh Jhamtani, Srinagesh Sharma, Chuchu Fan, and Chi Wang. St…
  10. Chengsong Huang, Langlin Huang, Jixuan Leng, Jiacheng Liu, and Jiaxin Huang. 2025a. Efficient test-time scaling via self-calibration. arXiv preprint arXiv:2503.00031. — [arXiv:2503.00031](https://arxiv.org/abs/2503.00031) find
  11. Young-Min Cho, Sharath Chandra Guntuku, and Lyle Ungar. 2025. Herd behavior: Investigating peer influence in llm-based multi-agent systems. arXiv preprint arXiv:2505.21588. — [arXiv:2505.21588](https://arxiv.org/abs/2505.21588) find
  12. Robert B Cialdini and 1 others. 2009. Influence: Science and practice, volume 4. Pearson education Boston. find
  13. Vincent Cohen-Addad, Varun Kanade, Frederik Mallmann-Trenn, and Claire Mathieu. 2019. Hierarchical clustering: Objective functions and algorithms. Journal of the ACM (JACM), 66(4):1–42. find
  14. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  15. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Liangjie Huang, Dawei Li, Huan Liu, and Lu Cheng. 2025b. Beyond accuracy: The role of calibration in self-improving large language models. arXiv preprint arX… — [arXiv:2504.02902](https://arxiv.org/abs/2504.02902) find
  16. Mingjian Jiang, Yangjun Ruan, Sicong Huang, Saifei Liao, Silviu Pitis, Roger Baker Grosse, and Jimmy Ba. 2023. Calibrating language models via augmented prompt ensembles. In International Conference on Machine Learning. find
  17. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computatio…
  18. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, and 1 others. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221. find
  19. Sanyam Kapoor, Nate Gruver, Manley Roberts, Katie Collins, Arka Pal, Umang Bhatt, Adrian Weller, Samuel Dooley, Micah Goldblum, and Andrew G Wilson. 2024. Large language models must be taught to know what they don’t know. Advances in Neural Information Processing Systems, 37:85932–85972. find
  20. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur… — Natural Questions: A Benchmark for Question Answering Research
  21. Chengzu Li, Han Zhou, Goran Glavaš, Anna Korhonen, and Ivan Vulić. 2025a. Large language models are miscalibrated in-context learners. In Findings of the Association for Computational Linguistics: ACL 2025, pages 11575–11596. find
  22. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36:41451–41530.
  23. Stella Li, Vidhisha Balachandran, Shangbin Feng, Jonathan Ilgen, Emma Pierson, Pang Wei W Koh, and Yulia Tsvetkov. 2024. Mediq: Question-asking llms and a benchmark for reliable interactive clinical reasoning. Advances in Neural Information Processing Systems, 37:28858–28888.
  24. Yibo Li, Miao Xiong, Jiaying Wu, and Bryan Hooi. 2025b. Conftuner: Training large language models to express their confidence verbally. arXiv preprint arXiv:2508.18847. — [arXiv:2508.18847](https://arxiv.org/abs/2508.18847) find
  25. Yubo Li, Yidi Miao, Xueying Ding, Ramayya Krishnan, and Rema Padman. 2025c. Firm or fickle? evaluating large language models consistency in sequential interactions. arXiv preprint arXiv:2503.22353. find
  26. Yubo Li, Xiaobin Shen, Xinyu Yao, Xueying Ding, Yidi Miao, Ramayya Krishnan, and Rema Padman. 2025d. Beyond single-turn: A survey on multi-turn interactions with large language models. arXiv preprint arXiv:2504.04717. — [arXiv:2504.04717](https://arxiv.org/abs/2504.04717) find
  27. Zhenwen Liang, Dian Yu, Wenhao Yu, Wenlin Yao, Zhihan Zhang, Xiangliang Zhang, and Dong Yu. 2024. Mathchat: Benchmarking mathematical reasoning and instruction following in multi-turn interactions. arXiv preprint arXiv:2405.19444. find
  28. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024a. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. find
  29. Ben Liu, Jihai Zhang, Fangquan Lin, Xu Jia, and Min Peng. 2025. One size doesn’t fit all: A personalized conversational tutoring agent for mathematics instruction. In Companion Proceedings of the ACM on Web Conference 2025, pages 2401–2410. find
  30. Xin Liu, Farima Fatahi Bayat, and Lu Wang. 2024b. Enhancing language model factuality via activation-based confidence calibration and guided decoding. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 10436–10448. find
  31. Andrey Malinin and Mark Gales. Uncertainty estimation in autoregressive structured prediction. In International Conference on Learning Representations. find
  32. Zhiting Mei, Christina Zhang, Tenny Yin, Justin Lidard, Ola Shorinwa, and Anirudha Majumdar. 2025. Reasoning about uncertainty: Do reasoning models know when they don’t know? arXiv preprint arXiv:2506.18183.
  33. Rui Miao, Yixin Liu, Yili Wang, Xu Shen, Yue Tan, Yiwei Dai, Shirui Pan, and Xin Wang. 2026. Blind-guard: Safeguarding llm-based multi-agent systems under unknown attacks. find
  34. Alexandru Niculescu-Mizil and Rich Caruana. 2005. Predicting good probabilities with supervised learning. In Proceedings of the 22nd international conference on Machine learning, pages 625–632. find
  35. Daniel J O’keefe. 2015. Persuasion: Theory and research. Sage Publications. find
  36. Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, and 1 others. 2023. Discovering language model behaviors with model-written evaluations. In Findings of the association for computational linguistics:…
  37. John Platt and 1 others. 1999. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61–74. find
  38. Romain Puech, Jakub Macina, Julia Chatain, Mrinmaya Sachan, and Manu Kapur. 2024. Towards the pedagogical steering of large language models for tutoring: A case study with modeling productive failure. arXiv preprint arXiv:2410.03781. find
  39. Huachuan Qiu, Hongliang He, Shuai Zhang, Anqi Li, and Zhenzhong Lan. 2024. Smile: Single-turn to multi-turn inclusive language expansion via chatgpt for mental health support. In EMNLP (Findings).
  40. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  41. Ved Sirdeshmukh, Kaustubh Deshpande, Johannes Mols, Lifeng Jin, Ed-Yeremai Cardona, Dean Lee, Jeremy Kritz, Willow Primack, Summer Yue, and Chen Xing. 2025. Multichallenge: A realistic multi-turn conversation evaluation benchmark challenging to frontier llms. arXiv preprint arXiv:2501.17399. find
  42. Elias Stengel-Eskin, Peter Hase, and Mohit Bansal. 2025. Teaching models to balance resisting and accepting persuasion. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers),… find
  43. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118. find
  44. 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. In Proceedings of the 2023 Confer… find
  45. Dennis Thomas Ulmer, Martin Gubri, Hwaran Lee, Sangdoo Yun, and Seong Joon Oh. 2024. Calibrating large language models using their generations only. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 15440–15459. Association for Computational Linguistic… find
  46. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in neural information processing systems, 33:5776–5788. find
  47. Johannes Welbl, Nelson F Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 94–106.
  48. Wenqing Wu, Yi Zhao, Yuzhuo Wang, Siyou Li, Juexi Shao, Yunfei Long, and Chengzhi Zhang. 2026. Cnovbench: Evaluating large language models on academic paper novelty assessment. arXiv preprint arXiv:2604.11543. — [arXiv:2604.11543](https://arxiv.org/abs/2604.11543) find
  49. Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, and 1 others. 2024a. Finben: A holistic financial benchmark for large language models. Advances in Neural Information Processing Systems, 37:95716–95743. find
  50. Qiming Xie, Zengzhi Wang, Yi Feng, and Rui Xia. 2024b. Ask again, then fail: Large language models’ vacillations in judgment. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10709–10745. find
  51. Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. In The Twelfth International Conference on Learning Representations. find
  52. Rongwu Xu, Brian Lin, Shujian Yang, Tianqi Zhang, Weiyan Shi, Tianwei Zhang, Zhixuan Fang, Wei Xu, and Han Qiu. 2024. The earth is flat because...: Investigating llms’ belief towards misinformation via persuasive conversation. In Proceedings of the 62nd Annual Meeting of the Association for Computat… find
  53. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  54. Dongkeun Yoon, Seungone Kim, Sohee Yang, Sunkyoung Kim, Soyeon Kim, Yongil Kim, Eunbi Choi, Yireun Kim, and Minjoon Seo. 2025. Reasoning models better express their confidence. arXiv preprint arXiv:2505.14489. — [arXiv:2505.14489](https://arxiv.org/abs/2505.14489) find
  55. Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. 2025a. Reasoning models know when they’re right: Probing hidden states for self-verification. arXiv preprint arXiv:2504.05419. find
  56. Caiqi Zhang, Ruihan Yang, Xiaochen Zhu, Chengzu Li, Tiancheng Hu, Yijiang River Dong, Deqing Yang, and Nigel Collier. 2026. Confidence estimation for llms in multi-turn interactions. arXiv preprint arXiv:2601.02179.
  57. Caiqi Zhang, Xiaochen Zhu, Chengzu Li, Nigel Collier, and Andreas Vlachos. 2025b. Reinforcement learning for better verbalized confidence in long-form generation. arXiv preprint arXiv:2505.23912. — [arXiv:2505.23912](https://arxiv.org/abs/2505.23912) find
  58. Hanlin Zhang, Yifan Zhang, Yaodong Yu, Dhruv Madeka, Dean Foster, Eric Xing, Himabindu Lakkaraju, and Sham Kakade. 2024. A study on the calibration of in-context learning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lang… find
  59. Zhaohan Zhang, Ziquan Liu, and Ioannis Patras. 2025c. Get confused cautiously: Textual sequence memorization erasure with selective entropy maximization. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10924–10939.
  60. Zhaohan Zhang, Ziquan Liu, and Ioannis Patras. 2025d. Grace: A generative approach to better confidence elicitation in large language models. arXiv preprint arXiv:2509.09438. — [arXiv:2509.09438](https://arxiv.org/abs/2509.09438) find
  61. 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. find
  62. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems, 36:46595–46623.
  63. Han Zhou, Xingchen Wan, Lev Proleev, Diana Mincu, Jilin Chen, Katherine A Heller, and Subhrajit Roy. Batch calibration: Rethinking calibration for in-context learning and prompt engineering. In The Twelfth International Conference on Learning Representations. find
  64. Xiaochen Zhu, Caiqi Zhang, Tom Stafford, Nigel Collier, and Andreas Vlachos. 2024. Conformity in large language models. arXiv preprint arXiv:2410.12428.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-confidence-should-be-calibrated-more/