Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph
Open accessThe DOI matches its Crossref record; 78 of the 78 references it cites were found in a registry or an index; 1 of the 20 sources its claims cite were checked for support.
Abstract
The rapid proliferation of large language models (LLMs) has stimulated researchers to seek effective and efficient approaches to deal with LLM hallucinations and low-quality outputs. Uncertainty quantification (UQ) is a key element of machine learning applications in dealing with such challenges. However, research to date on UQ for LLMs has been fragmented in terms of techniques and evaluation methodologies. In this work, we address this issue by introducing a novel benchmark that implements a collection of state-of-the-art UQ baselines and offers an environment for controllable and consistent evaluation of novel UQ techniques over various text generation tasks. Our benchmark also supports the assessment of confidence normalization methods in terms of their ability to provide interpretable scores. Using our benchmark, we conduct a large-scale empirical investigation of UQ and normalization techniques across eleven tasks, identifying the most effective approaches.
Overview
CoChat Summary
TL;DR
The paper introduces LM-Polygraph, a benchmark for consistent evaluation of uncertainty quantification (UQ) and confidence normalization methods for LLM text generation across selective QA/generation, claim-level fact-checking, and calibration. A large-scale empirical study finds sample-diversity methods strong overall, information-based methods best for short outputs, CCP best for claim-level fact-checking, and isotonic PCC useful for interpretable confidence without degrading ranking performance.
Key findings
- In selective QA/generation, sample-diversity methods generally perform well, especially for longer outputs; for short generations (<7 symbols), information-based methods such as MSP, Perplexity, Mean Token Entropy, CCP, and TokenSAR perform best. SAR is highlighted as robust across output lengths.
- For claim-level fact-checking, CCP achieves the best ROC-AUC/PR-AUC for English Mistral 7B (0.74/0.46), English Jais 13B (0.68/0.47), Arabic Jais 13B (0.73/0.30), and Russian Vikhr 7B (0.67/0.80); for Chinese Yi 6B, CCP and P(True) are comparable (both 0.61 ROC-AUC, 0.25 PR-AUC).
- MSP is a strong and robust baseline across tasks and should not be discarded; P(True) often performs poorly, especially with smaller LLMs.
- Binned and isotonic PCC confidently outperform linear and quantile normalization for calibration (MSE), and isotonic PCC preserves PRR relative to raw uncertainty scores (Table 13 shows mostly 0.000 difference).
Methods
- The benchmark is built on the LM-Polygraph framework and implements white-box and black-box UQ baselines, claim-level extensions, and normalization methods including linear scaling, quantile scaling, Binned PCC, and Isotonic PCC.
- Selective QA/generation uses CoQA, TriviaQA, MMLU, GSM8k, XSum, WMT14 Fr-En, and WMT19 De-En; claim-level fact-checking uses automatically generated biographies in English, Arabic, Russian, and Chinese, with GPT-4 for claim extraction, token mapping, and annotation.
- Models include Mistral 7B v0.2, Stable LM 2 12B, GPT-4o-mini, Jais 13B, Vikhr 7B-0.2, and Yi 7B/6B; evaluation metrics include PRR, ROC-AUC/PR-AUC, and MSE between normalized confidence and quality.
Limitations
- The paper does not include a dedicated limitations section; it states that about 5% of claims cannot be mapped to tokens and are excluded, and that no more than 8% of claims were classified as unknown and discarded.
- GPT-4 is used as an automatic annotator and judge; the paper validates it against manual annotation (e.g., English Mistral-7B-v0.1 accuracy 0.98, F1 0.93; Fleiss Kappa 0.90), but this remains a proxy for human judgment.
- Evident from the text: results are tied to the specific models, datasets, languages, and prompt formats tested, so generalization to other LLMs, tasks, or domains is not established.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
1 Valid · 1 Not valid · 16 Not checked · 78 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“We argue that such an approach introduces a big discrepancy between the generated text and what a proxy LLM actually wants to generate, which results in biased UQ performance.”
What the sources say
- Supportedcites Manakul et al., 2023 matched to “SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models” (2023) source
The Proxy LLMs section (p. 3) describes using another LLM to approximate the generating model’s token probabilities. Section 7.1 (p. 6) reports worse proxy-LLM detection performance and suggests that differing generation patterns can assign low probabilities to the generated text. The citing authors frame the resulting bias as their own argument, rather than attributing that conclusion directly to Manakul et al. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Despite being a simplistic baseline, MSP demonstrates strong performance.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, for MMLU, these methods fall significantly behind informationbased techniques.”
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“State-of-the-art methods based on sample diversity typically perform well on CoQA and TriviaQA.”
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The majority of density-based methods demonstrate poor results across all tasks.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
34 more claims
Create a free account to see all 96 claims and to verify the 19 references still open.
- It may state two separate claims as one. 31
- The quote only partly states this claim. 1
- Probably not a substantial claim. 36
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“Notably, there is growing attention on assessing LLM bias [63-64] and robustness [323-324].”
From A Survey of Large Language Models (2026) · cites this paper as [323]
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 5.4 Models · p. 9
“Figure 1: Mean PRR ↑ aggregated over all selective QA tasks for each white-box LLM separately (the lower the better).”
§ 5.1 Selective QA / Generation · p. 7
“A higher PRR indicates a better uncertainty score.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 6.2 Fact-Checking · p. 11
“In the fact-checking evaluation, we generated biographies for 100 individuals across five languages: English (using Mistral 7B-v0.1 and Jais 13B), Arabic (using Jais 13B), Russian (using Vikhr 7B-0.2), and Chinese (using Yi 6B).”
§ 5.2 Claim-Level Fact-Checking · p. 8
“The pipeline supports English, Mandarin Chinese, Arabic, and Russian.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 5.4 Models · p. 9
“Yi 7B (for Chinese).”
§ 6.2 Fact-Checking · p. 11
“Chinese (using Yi 6B).”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 6.1 Selective QA and Generation · p. 11
“For both models, on the text summarization task, the best results are achieved by sample-diversity techniques in terms of both metrics.”
§ 6.1 Selective QA and Generation · p. 11
“One exception is the performance of Conditional PMI in terms of ROUGE-L-based PRR, which is drastically better than the performance of all other methods.”
Citations
Cites
Advances in Neural Information Processing Systems 19
(2007) · The MIT Press eBooks · doi:10.52202/085713 · Cited by 15,539
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929 · cited 2×
“Ren et al. (2023) showed that it might be useful to adjust the Mahalanobis distance score by subtracting from it the other Mahalanobis distance MD 0 ( x ) computed for some large general purpose dataset covering many domains like C4 (Raffel et al., 2020).”
Least Median of Squares Regression
Rousseeuw, P. J. (1984) · Journal of the American Statistical Association · doi:10.1080/01621459.1984.10477105 · Cited by 3,835 · cited 1×
“Additionally, computing of the covariance matrix Σ for each individual class is done by us-ing the Minimum Covariance Determinant estimation (Rousseeuw, 1984).”
A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks
Lee, K., Lee, K., Lee, H., & Shin, J. (2018) · arXiv · doi:10.48550/arxiv.1807.03888 · Cited by 2,450 · cited 2×
“Several variations of these methods have been proposed recently (Lee et al., 2018; Yoo et al., 2022; Kotelevskii et al., 2022; Ren et al., 2023).”
DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION
He, P., Liu, X., Gao, J., & Chen, W. (2021) · International Conference on Learning Representations · Cited by 2,445
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., et al. (2023) · arXiv · doi:10.48550/arxiv.2306.05685 · Cited by 1,888 · cited 1×
“We provide a more detailed discussion on the specifics of the normalization methods in Appendix B.”
TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension
Joshi, M., Choi, E., Weld, D., & Zettlemoyer, L. (2017) · Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p17-1147 · Cited by 1,690 · cited 1×
“The first two datasets are used in many works on UQ, while the last two are popular datasets for evaluating the English-language generation quality of LLMs.”
TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension
Joshi, M., Choi, E., Weld, D. S., & Zettlemoyer, L. (2017) · arXiv · doi:10.48550/arxiv.1705.03551 · Cited by 1,650 · cited 1×
“The first two datasets are used in many works on UQ, while the last two are popular datasets for evaluating the English-language generation quality of LLMs.”
Weight Uncertainty in Neural Network
Blundell, C., Cornebise, J., Kavukcuoglu, K., & Wierstra, D. (2015) · International Conference on Machine Learning · Cited by 1,335
Narayan, S., Cohen, S. B., & Lapata, M. (2018) · Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d18-1206 · Cited by 1,207 · cited 1×
“For selective generation, we use two MT datasets: WMT-14 French to English (Bojar et al., 2014), WMT-19 German to English (Barrault et al., 2019), and one text summarization dataset – XSum (Narayan et al., 2018).”
Research works printed but not linked: 4
- Yarin Gal. 2016. Uncertainty in Deep Learning . Ph.D. thesis, University of Cambridge.
- 16-20, 2020 , pages 8362-8372. Association for Computational Linguistics.
- Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. DeBERTa: Decodingenhanced bert with disentangled attention. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 .
- Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. CoRR , abs/2310.06825. · arXiv
Cited by
A Survey of Large Language Models
Zhao, W. X., Zhou, K., Li, J., Tang, T., Dong, Z., Hou, Y., Zhang, B., Min, Y., Zhang, J., Liu, P., et al. (2026) · Frontiers of Computer Science · doi:10.1007/s11704-026-60308-3 · Cited by 2,093
Shorinwa, O., Mei, Z., Lidard, J., Ren, A. Z., & Majumdar, A. (2026) · ACM Computing Surveys · doi:10.1145/3744238 · Cited by 61
“Lastly, we note that there has been some work that aims to standardize the tasks for evaluating the performance of LLMs by explicitly accounting for the uncertainty of LLMs in specific tasks, e.g., based on selective classification and generation [209] or conformal prediction [234].”
Uncertainty Quantification and Confidence Calibration in Large Language Models: A Survey
Liu, X., Chen, T., Da, L., Chen, C., Lin, Z., & Wei, H. (2025) · Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 · doi:10.1145/3711896.3736569 · Cited by 53
A Survey of Uncertainty Estimation Methods on Large Language Models
Xia, Z., Xu, J., Zhang, Y., & Liu, H. (2025) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2025.findings-acl.1101 · Cited by 120
A Survey on Recent Advances in Conversational Data Generation
Soudani, H., Petcu, R., Kanoulas, E., & Hasibi, F. (2026) · ACM Computing Surveys · doi:10.1145/3795686 · Cited by 39
Metadata
- Authors
- Vashurin, R., Fadeeva, E., Vazhentsev, A., Rvanova, L., Vasilev, D., Tsvigun, A., Petrakov, S., Xing, R., Sadallah, A., Grishchenkov, K., Panchenko, A., Baldwin, T., Nakov, P., Panov, M., & Shelmanov, A.
- Affiliations
- Mohamed bin Zayed University of Artificial Intelligence (AE)
- ETH Zurich (CH)
- Center for Artificial Intelligence Technology, Russia. vazhentsev@airi.net, artiomvazh99@gmail.com
- Laboratory for Analysis and Controllable Text Generation Technologies RAS, Russia. Milarv99@gmail.com
- National Research University Higher School of Economics (RU)
- Nebius. aktsvigun@gmail.com
- Center for Artificial Intelligence Technology, Russia. sergeypetrakof@gmail.com
- The University of Melbourne (AU)
- Independent Researcher. kirillgrish@gmail.com
- Center for Artificial Intelligence Technology, Russia. panchenkoalexander@gmail.com, panchenko@airi.net
- ORCID
- Venue
- Transactions of the Association for Computational Linguistics
- Publisher
- MIT Press
- Published
- 2025
- Volume
- 13
- Pages
- 220-248
- ISSN
- 2307-387X (ISSN-L)
- OpenAlex
- W4409218928
- Semantic Scholar
- cc0c6f4dbbfc163cfae15724da1d7e3042fa099c
- Cited by
- 21
- References
- 75
- Citations by year
- 2026: 17 · 2025: 4
- FWCI
- 8.42
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Licence
- Locations
- Transactions of the Association for Computational Linguistics (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion, cc-by)
- Minerva Access (University of Melbourne) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- uncertainty quantification, large language models, LLM hallucination, benchmark, text generation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-04, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-04 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph. https://app.cochat.ai/papers/10.1162/tacl_a_00737
- The paper
- Vashurin, R., Fadeeva, E., Vazhentsev, A., Rvanova, L., Vasilev, D., Tsvigun, A., Petrakov, S., Xing, R., Sadallah, A., Grishchenkov, K., Panchenko, A., Baldwin, T., Nakov, P., Panov, M., & Shelmanov, A. (2025). Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph. Transactions of the Association for Computational Linguistics. https://doi.org/10.1162/tacl_a_00737
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.