CoChat

Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph

Open access

The 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

Crossref · 10.1162/tacl_a_00737

Cites

78 of 78 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 1 Not valid · 16 Not checked · 78 Own results and interpretations

34 more claims

Create a free account to see all 96 claims and to verify the 19 references still open.

See all claims — free account
  • It may state two separate claims as one. 31
  • The quote only partly states this claim. 1
  • Probably not a substantial claim. 36

Create a free account to see these claims.

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?

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.

4 of 4

Citations

Cites

Show all 78 references

Research works printed but not linked: 4

  1. Yarin Gal. 2016. Uncertainty in Deep Learning . Ph.D. thesis, University of Cambridge.
  2. 16-20, 2020 , pages 8362-8372. Association for Computational Linguistics.
  3. 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 .
  4. 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

See all 37

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
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
Cited by
21
References
75
Citations by year
2026: 17 · 2025: 4
FWCI
8.42
Type
journal article
Language
en
Open access
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

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.