Elamur Research⌕
☰
home / papers / 2026 / Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification

Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification

Paper · 2026 · Elena Merdjanovska, Omar Zaidan and Andreas Rücklé

Summary

When prompted LLMs act as classifiers, users often need a confidence score to decide which predictions to trust or reject. This paper examines a neglected problem: verbalized confidences take only a handful of distinct values. Across four datasets (SST-2, SST-5, Amazon ESCI, Yahoo! answers topics), the authors test five estimation methods with Qwen3-32B and Claude 3.7 Sonnet: vanilla verbalization, top-k verbalization, consistency sampling, verbalization sampling and token logprobs. Such sparse confidences give few usable rejection thresholds. They also distort evaluation. Computing AUARC with linear interpolation inflates the scores of sparse methods and can reorder the rankings, so the authors recommend stepwise interpolation as the standard. They also propose verbalization logprobs, which computes the expected value of the verbalized confidence digits using their token probabilities. This turns a few discrete values into a near-continuous score. Under stepwise AUARC, the method performs best and adds no inference cost over vanilla verbalization. Practitioners get more usable thresholds, and researchers get a fairer basis for comparing methods.

Key points

  • On SST-2, Qwen3-32B with vanilla verbalization produced only eight distinct confidence values, and more than half were exactly 95%. Across the datasets, 45–93% of verbalized confidences fell into the five most frequent values.
  • Consistency sampling ranks first under linear AUARC (0.808) but last under stepwise AUARC (0.683). Token logprobs, which produces continuous scores, is unaffected by the interpolation choice.
  • Verbalization logprobs reaches a stepwise AUARC of 0.736, versus 0.713 for vanilla verbalization (+2.3 points). This roughly matches verbalization sampling (0.734) at 1.7× instead of 6.8× relative cost.
  • The number of distinct confidence values grows from 8–12 (vanilla) to 630–2,227 (verbalization logprobs), while Spearman ρ between the two methods averages 0.89. To reach ≥98% accuracy on SST-2, only 28.1% of predictions need to be rejected instead of 39.7%.
  • The method approximates the digits as independent, assumes each digit is a single token, and needs logprob access. It is therefore unavailable for purely black-box APIs.

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

🎓Humboldt-Universität zu BerlinBerlin · Germany19 works · 41,895 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 1Research subjects and languages the work is about.Evaluation & benchmarks · 1,220 works

Cite

Cite as Elena Merdjanovska, Omar Zaidan and Andreas Rücklé. 2026. Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1671

@inproceedings{merdjanovska2026evaluation,
  title = {{Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification}},
  author = {Elena Merdjanovska, Omar Zaidan and Andreas Rücklé},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1671},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1671}
}

References

44 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jiefeng Chen, Jinsung Yoon, Sayna Ebrahimi, Sercan Arik, Tomas Pfister, and Somesh Jha. 2023. Adaptation with self-evaluation to improve selective prediction in LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5190–5213, Singapore. find
  2. Wenyu Chen, Chen Miao, Zhenghao Zhang, Cathy Sin-Hang Fung, Ran Wang, Yizhen Chen, Yan Qian, Lixin Cheng, Kevin Y Yip, Stephen Kwok-Wing Tsui, and 1 others. 2024. Commonly used software tools produce conflicting and overly-optimistic auprc values. Genome Biology, 25(1):118. find
  3. Javier Coronado-Blázquez. 2025. Deterministic or probabilistic? the psychology of llms as random number generators. Preprint, arXiv:2502.19965. — [arXiv:2502.19965](https://arxiv.org/abs/2502.19965) find
  4. Jinhao Duan, Hao Cheng, Shiqi Wang, Alex Zavalny, Chenan Wang, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. 2024. Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models. In Proceedings of the 62nd Annual Meeting of the Association for Com… find
  5. Ran El-Yaniv and Yair Wiener. 2010. On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11(53):1605–1641. find
  6. Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2024. A survey of confidence estimation and calibration in large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language… find
  7. James A Hanley and Barbara J McNeil. 1982. The meaning and use of the area under a receiver operating characteristic (roc) curve. Radiology, 143(1):29– 36. find
  8. Yuheng Huang, Jiayang Song, Zhijie Wang, Shengming Zhao, Huaming Chen, Felix Juefei-Xu, and Lei Ma. 2025. Look before you leap: An exploratory study of uncertainty analysis for large language models. IEEE Transactions on Software Engineering. find
  9. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, and 17 others. 2022.… find
  10. Seungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin, Jamin Shin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, and Minjoon Seo. 2024. Prometheus 2: An open source language model specialized in evaluating other language models. In Proceedings of the 2024 Conference on Empirical Metho…
  11. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NeurIPS ’22, Red Hook, NY, USA. Curran Associates Inc.
  12. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations (ICLR 2023). find
  13. Yukyung Lee, JoongHoon Kim, Jaehee Kim, Hyowon Cho, Jaewook Kang, Pilsung Kang, and Najoung Kim. 2025. CheckEval: A reliable LLM-as-a-judge framework for evaluating text generation using checklists. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 1578…
  14. Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2024a. Contextualized sequence likelihood: Enhanced confidence scores for natural language generation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 10351– 10368, Miami, Florida, USA. find
  15. Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2024b. Generating with confidence: Uncertainty quantification for black-box large language models. Transactions on Machine Learning Research. find
  16. Chen Ling, Xujiang Zhao, Xuchao Zhang, Wei Cheng, Yanchi Liu, Yiyou Sun, Mika Oishi, Takao Osaki, Katsushi Matsuda, Jie Ji, Guangji Bai, Liang Zhao, and Haifeng Chen. 2024. Uncertainty quantification for in-context learning of large language models. In Proceedings of the 2024 Conference of the North… find
  17. Gabrielle Kaili-May Liu, Gal Yona, Avi Caciularu, Idan Szpektor, Tim G. J. Rudner, and Arman Cohan. 2025a. MetaFaith: Faithful natural language uncertainty expression in LLMs. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 29600–29644, Suzhou, China.
  18. Xiaoou Liu, Tiejin Chen, Longchao Da, Chacha Chen, Zhen Lin, and Hua Wei. 2025b. Uncertainty quantification and confidence calibration in large language models: A survey. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining.
  19. Joseph Marvin Imperial and Harish Tayyar Madabushi. 2025. Scaling policy compliance assessment in language models with policy reasoning traces. arXiv e-prints, pages arXiv–2509. find
  20. John Muschelli III. 2020. Roc and auc with a binary predictor: a potentially misleading metric. Journal of classification, 37(3):696–708. find
  21. Malik Sajjad Ahmed Nadeem, Jean-Daniel Zucker, and Blaise Hanczar. 2009. Accuracy-rejection curves (arcs) for comparing classification methods with a reject option. In Proceedings of the third International Workshop on Machine Learning in Systems Biology, volume 8 of Proceedings of Machine Learning… find
  22. Huy Nghiem, Advik Sachdeva, and Hal Daumé III. 2025. Smarter: A data-efficient framework to improve toxicity detection with explanation via self-augmenting large language models. arXiv preprint arXiv:2509.15174.
  23. Dang Nguyen, Ali Payani, and Baharan Mirzasoleiman. 2025. Beyond semantic entropy: Boosting LLM uncertainty quantification with pairwise semantic similarity. In Findings of the Association for Computational Linguistics: ACL 2025, pages 4530–4540, Vienna, Austria. find
  24. Alexander Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. 2024. Kernel language entropy: fine-grained uncertainty quantification for llms from semantic similarities. In Proceedings of the 38th International Conference on Neural Information Processing Systems, NeurIPS ’24, Red Hook, NY, USA.… find
  25. Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard Duchesnay. 2011. Scikit-learn:… find
  26. Edward Phillips, Sean Wu, Soheila Molaei, Danielle Belgrave, Anshul Thakur, and David Clifton. 2025. Geometric uncertainty for detecting and correcting hallucinations in llms. arXiv preprint arXiv:2509.13813. — [arXiv:2509.13813](https://arxiv.org/abs/2509.13813) find
  27. Chandan K. Reddy, Lluís Màrquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. Shopping queries dataset: A large-scale ESCI benchmark for improving product search. Preprint, arXiv:2206.06588. — [arXiv:2206.06588](https://arxiv.org/abs/2206.06588) find
  28. Jie Ren, Yao Zhao, Tu Vu, Peter J. Liu, and Balaji Lakshminarayanan. 2023. Self-evaluation improves selective generation in large language models. In Proceedings on "I Can’t Believe It’s Not Better: Failure Modes in the Age of Foundation Models" at NeurIPS 2023 Workshops, volume 239 of Proceedings o… find
  29. Jiandong Shao, Yao Lu, and Jianfei Yang. 2025. Benford’s curse: Tracing digit bias to numerical hallucination in llms. arXiv preprint arXiv:2506.01734. — [arXiv:2506.01734](https://arxiv.org/abs/2506.01734) find
  30. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–…
  31. Tao Sun, Jian Xu, Yuanpeng Li, Zhao Yan, Ge Zhang, Lintao Xie, Lu Geng, Zheng Wang, Yueyan Chen, Qin Lin, Wenbo Duan, Kaixin Sui, and Yuanshuo Zhu. 2025. Bitsai-cr: Automated code review via llm in practice. In Proceedings of the 33rd ACM International Conference on the Foundations of Software Engin… find
  32. Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher 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 Conferen… find
  33. Francesco Tonolini, Nikolaos Aletras, Jordan Massiah, and Gabriella Kazai. 2024. Bayesian prompt ensembles: Model uncertainty estimation for black-box large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12229–12272, Bangkok, Thailand. find
  34. Roman Vashurin, Ekaterina Fadeeva, Artem Vazhentsev, Lyudmila Rvanova, Daniil Vasilev, Akim Tsvigun, Sergey Petrakov, Rui Xing, Abdelrahman Sadallah, Kirill Grishchenkov, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, Maxim Panov, and Artem Shelmanov. 2025. Benchmarking uncertainty quantificat…
  35. Artem Vazhentsev, Ekaterina Fadeeva, Rui Xing, Gleb Kuzmin, Ivan Lazichny, Alexander Panchenko, Preslav Nakov, Timothy Baldwin, Maxim Panov, and Artem Shelmanov. 2025. Unconditional truthfulness: Learning unconditional uncertainty of large language models. In Proceedings of the 2025 Conference on Em…
  36. Mozhi Zhang, Mianqiu Huang, Rundong Shi, Linsen Guo, Chong Peng, Peng Yan, Yaqian Zhou, and Xipeng Qiu. 2024. Calibrating the confidence of large language models by eliciting fidelity. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2959– 2979, Miami,… find
  37. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems (NeurIPS 2015).
  38. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. In Proceedings of the 37th International Conference on Neu…
  39. Cheng Wang, Gyuri Szarvas, Georges Balazs, Pavel Danchenko, and Patrick Ernst. 2024. Calibrating verbalized probabilities for large language models. Preprint, arXiv:2410.06707. — [arXiv:2410.06707](https://arxiv.org/abs/2410.06707) find
  40. 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. In The Twelfth International Conference on Learning Representations (ICLR 2024). find
  41. Weihao Xuan, Qingcheng Zeng, Heli Qi, Junjue Wang, and Naoto Yokoya. 2025. Seeing is believing, but how much? a comprehensive analysis of verbalized calibration in vision-language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 1408–1450, Suzh…
  42. 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.
  43. Fan Yin, Philippe Laban, XIANGYU PENG, Yilun Zhou, Yixin Mao, Vaibhav Vats, Linnea Ross, Divyansh Agarwal, Caiming Xiong, and Chien-Sheng Wu. 2025. Bingoguard: LLM content moderation tools with risk levels. In The Thirteenth International Conference on Learning Representations (ICLR 2025). find
  44. Qingcheng Zeng, Weihao Xuan, Leyang Cui, and Rob Voigt. 2025. Thinking out loud: Do reasoning models know when they’re right? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 1394–1407, Suzhou, China.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/merdjanovska-2026-evaluation-pitfalls-sparsity-limitations-llm/