Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Reconsidering LLM Uncertainty Estimation Methods in the Wild

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Most LLM uncertainty estimators are highly sensitive to the decision threshold when calibration and test distributions differ, and ensembling many estimators at test time yields consistent performance gains.

desk verdict A broad, mostly solid empirical study of UE methods under deployment-style stress; the headline findings hold up, but the all-LLM label pipeline needs a human-validation check before the numbers are treated as benchmark-grade. read the letter →

arxiv 2506.01114 v1 pith:LFZCVOGE submitted 2025-06-01 cs.LG cs.AI

Yavuz Bakman , Duygu Nur Yaldiz , Sungmin Kang , Tuo Zhang , Baturalp Buyukates , Salman Avestimehr , Sai Praneeth Karimireddy This is my paper
classification cs.LGcs.AI
keywords uncertaintyestimationhallucinationdetectionlargelanguagemodelsthresholdcalibrationdistributionshiftadversarialpromptsensemblinglong-formgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Most uncertainty-scoring methods for large language models are evaluated in clean short-form question-answering settings with threshold-free metrics, which hides how they behave when actually deployed. This paper tests 19 such methods under four real-world stresses: choosing a decision threshold on a calibration set that may not match test data, typos, chat history, adversarial prompt injections, and long-form answers. It finds that the majority of methods are highly sensitive to threshold selection under distribution shift, with only a few (MARS, Semantic Entropy, Eccentricity) staying accurate across calibration datasets. The same methods mostly shrug off typos and prior chat history but drop sharply under a confidence-boosting adversarial prompt. It then shows that short-form methods can be adapted to long-form answers, and that simply averaging normalized or calibrated scores from many UE methods outperforms the best single method, even with only 100 calibration samples.

What carries the argument

The device that carries the threshold argument is ARE, the average absolute deviation between target recall and achieved recall when a threshold tuned on a calibration set is applied to a different test set. The paper's other core instruments are the PRR curve for each method, strategy functions for long-form adaptation (Naive, Question Generation, and Question-Answer Generation, where QAG regenerates an answer to a claim-specific question and scores only when it aligns with the original claim), and test-time ensembling with raw, standardized, or isotonically calibrated scores combined by averaging, voting, or a linear model.

What would settle it

Recompute the threshold-transfer experiments (TriviaQA and GSM8K) using exact-match or human-verified correctness labels instead of the GPT-4o-mini judge and check whether MARS, Semantic Entropy, and Eccentricity are still the only methods with ARE below 0.10 across all calibration domains; if the set changes, the robustness ordering depends on the judge.

Watch

Extended reading notes

Core claim

The paper's central claim is that current UE methods, whatever their AUROC or PRR in isolation, do not yet have the properties a deployed hallucination detector needs. Its main evidence is a new threshold-generalization metric, Average Recall Error (ARE): calibrate a score threshold on one dataset to hit a target recall, apply it to another, and measure how much achieved recall deviates from the target. Under same-task and cross-domain shifts, most of the 19 methods exceed 0.10 ARE in at least one setting, while MARS, Semantic Entropy, and Eccentricity stay below that across all calibration datasets. The paper also argues and demonstrates that prior chat history and typos leave most methods almost unaffected, that an adversarial confidence-boosting prefix degrades probability-based methods substantially, that adapting short-form methods to claim-level long-form evaluation via generated questions and answers recovers only part of the short-form performance, and that ensembling the scores of many methods after normalization or isotonic calibration improves PRR over the best individual method by up to 0.06 even with a 100-sample calibration set.

Load-bearing premise

The results assume the correctness labels produced by GPT-4o-mini (and, for long-form claims, the SAFE pipeline's GPT-4o-mini reasoning) are right; if the judge systematically mislabels hallucinations, every ARE and PRR value shifts and the relative ranking of methods could change.

Editorial extensions

If this is right

  • Practitioners should not reuse a threshold calibrated on one dataset or domain on another; only MARS, Semantic Entropy, and Eccentricity held ARE below 0.10 across all calibration shifts tested.
  • A confidence-boosting prompt prefix is enough to degrade many UE methods, so hallucination-detection deployments should test robustness to prompt injection rather than assuming scores are stable.
  • Short-form UE methods can be carried to long-form claims, and generating several questions per claim and averaging the scores is the most effective of the proposed adaptation strategies.
  • Ensembling many UE scores after normalization or calibration is a cheap and reliable improvement, beating the best single method by up to 0.06 PRR with only 100 calibration samples.
  • The paper's proposed protocol—report ARE under distribution shift and test prompt variations—is a concrete template for evaluating future UE methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If threshold instability generalizes, the common practice of quoting a single AUROC or PRR number overstates how ready any UE method is for deployment, since those metrics average away the exact failure the paper measures.
  • The ensembling gains the paper reports are consistent with error diversity across methods; a direct test would be to select ensemble members by pairwise disagreement on the calibration data rather than including all 19 scores.
  • The adversarial-prompt result implies UE scores are part of the attack surface of an LLM application: a benign-looking 'be confident' instruction could be used as an injection to disable hallucination alarms without changing weights.
  • In the long-form experiments, QAG's advantage might come from question-generation quality rather than uncertainty scoring; an ablation with human-written questions for each claim would isolate the cause.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper presents a large empirical study of 19 LLM uncertainty-estimation (UE) methods under four real-world deployment challenges: decision-threshold selection under distribution shift, robustness to query transformations (chat context, typos, adversarial prompts), adaptation to long-form generation, and ensembling of multiple UE scores. It introduces Average Recall Error (ARE) to quantify threshold generalization, evaluates on TriviaQA, NaturalQA, and GSM8K with Llama-3-8B and GPT-4o-mini, and reports PRR for the other settings. The main conclusions are that most UE methods are highly sensitive to threshold selection when calibration and test distributions differ; that they are generally robust to context and typos but vulnerable to adversarial prompts; that long-form adaptation is possible but leaves room for improvement; and that simple ensembling of UE scores can improve over the best individual method.

Significance. If the findings hold, the paper is a useful contribution to the practical evaluation of UE methods: it identifies a previously underemphasized failure mode (threshold sensitivity under distribution shift), proposes the ARE metric, and shows that a small calibration set can support simple ensembling gains. The study covers a wide range of representative methods, and the authors provide code, which supports reproducibility. The paper also makes falsifiable claims about method robustness that can guide future UE development. However, the strength of the conclusions depends on several methodological choices that need to be tightened: the use of an LLM judge without human validation, the tuning of the adversarial prompt on the same dataset family used for evaluation, and the supervised training of two included methods on the evaluation datasets.

major comments (5)
  1. [Section 2.2 footnote; Appendix E.2; Appendix E.3] All correctness labels in the paper come from GPT-4o-mini (short-form QA) or the SAFE pipeline with GPT-4o-mini reasoning (long-form claims), and no human validation or inter-annotator agreement check is reported. Because every ARE and PRR value is computed against these labels, a systematic judge bias (e.g., correlated with answer length, verbosity, or question domain) could reorder method rankings and alter the qualitative conclusions. I ask the authors to include a human-annotated validation subset (for example, 200 short-form answers and 200 long-form claims) with agreement statistics, and to discuss any observed label-noise effects.
  2. [Section 4.2; Appendix E.1] The adversarial prompt for GPT-4o-mini is tuned on 150 samples from the TriviaQA training split over 15 iterations and then evaluated on TriviaQA test data (as well as GSM8K). This is an in-distribution attack for the TriviaQA results: the prompt was explicitly selected to degrade UE performance on the same dataset family used for evaluation, so the reported vulnerability on TriviaQA may be overstated. The GSM8K results are a useful transfer test, but the paper should also report a non-tuned baseline prompt (e.g., a generic confidence booster, as used for Llama-3-8B) and clearly separate tuned and transfer results in the discussion.
  3. [Appendix C.1; Appendix C.2; Section 3 and Table 1] The supervised methods LARS and SAPLMA are trained on the same datasets used for calibration and test in the threshold-sensitivity experiments. LARS is trained on GSM8K, TriviaQA, and NaturalQA (18k samples total), and SAPLMA on TriviaQA and GSM8K (13k samples). Consequently, the 'distribution shift' between calibration and test in Table 1 and Table 3 is not a true distribution shift for these methods: both the calibration and test distributions were seen during training. This confounds their ARE results and the paper's general statement that most methods are sensitive to distribution shift. The authors should either retrain these methods on strictly disjoint data, or clearly report and discuss this overlap, ideally by excluding them from the headline 'most methods' claim or presenting them in a separate category.
  4. [Section 3.1 and Section 3.2] The ARE computation as written is underspecified. The paper states that for each target recall r* a threshold is 'randomly picked from the set {t: M(U,D_cal,t)=m*}' and then evaluated on the test set, with R spanning 0 to 1.0 in increments of 0.001. With a calibration set of 500 samples, most target recalls are not exactly achievable (the achievable recall values are k/P for integer k), and the paper does not state how impossible targets are handled. This directly affects all ARE values in Tables 1 and 3 and must be clarified (e.g., nearest achievable recall, interpolation, or a tolerance window) for the results to be reproducible and interpretable.
  5. [Tables 1-3 and Table 2] Although all experiments are run with 5 seeds, Tables 1, 2, and 3 report only the average and no standard deviation or confidence interval. Several claims involve small differences (for example, identifying MARS, Semantic Entropy, and Eccentricity as consistently better than other methods in Table 1), and Table 2 compares ensemble strategies to a 'best single' baseline where gains are as small as 0.03-0.04 PRR. Without variance information, the stability of these rankings and gains is unclear. Please add standard deviations (or another measure of dispersion) to the tables, or at least to the values that underpin the headline claims.
minor comments (5)
  1. [Section 3.3] The sentence 'only MARS, Semantic Entropy, and Eccentricity consistently achieve low error across calibration datasets, while all others exceed 0.10 ARE in at least one setting' is contradicted by Table 1: Eccentricity-C has a maximum of 0.098 in the Llama-3-8B columns and never exceeds 0.10 in the GPT-4o-mini columns, and several other GPT-4o-mini rows also stay below 0.10. Please rephrase to 'most others' or explicitly state which comparison is being made.
  2. [Section 2.2 and Table 2] Section 2.2 states that PRR ranges from 0.0 (random) to 1.0 (perfect), but Table 2 reports a value of -0.02 for the 'Raw Max' ensemble on GSM8K with Llama-3-8B. Please clarify whether PRR can be negative in this implementation or whether this value is a typo.
  3. [Section 5.2, QAG strategy] In the QAG strategy, a score of infinity is assigned whenever the model-generated answer y' is not semantically equivalent to the claim. The paper does not report how often this fallback occurs. If it is frequent, the strategy may be separating claims largely on the basis of whether the model can reproduce the claim, rather than on the UE score itself. Please report the fraction of claims that receive the infinity score.
  4. [Appendix C.2] The paper modifies SAPLMA (adding the question to the input and using the last layer activations) and INSIDE (using only the score component without the test-time feature clipping). These are sensible implementation choices, but the main text should more prominently note that the benchmark evaluates modified variants, since readers may otherwise assume the original methods are being compared.
  5. [Appendix E.2] There is a typo: 'evalaution' should be 'evaluation' in the sentence 'For a consistent evalaution, we exclude question-generation pairs where the LLM refuses to provide an answer.'

Circularity Check

2 steps flagged · score 4.0 of 10

No derivation-level circularity; central threshold and ensembling results are genuine held-out evaluations, but two analysis choices make parts of the adversarial and SAPLMA findings partly self-confirming.

  1. fitted input called prediction [Section 4.3 and Appendix E.1 (adversarial prompt search for GPT-4o-mini)]
    "At each iteration, we request a modified prompt and evaluate its impact using the performance of three selected UE methods: LNS, Eccentricity, and KLE. ... After 15 iterations, we select the most effective prompt that maximally decreases UE performance. ... Finally, results indicate that the confidence booster prompt injection acts as an adversarial prompt, reducing performance across various datasets, particularly affecting probability-based methods in GPT-4o-mini."

    The adversarial prompt was explicitly optimized to maximize the performance drop of LNS, a probability-based method, on a TriviaQA training subset. The Section 4.3 conclusion that probability-based methods are 'particularly affected' is therefore the objective function of the prompt search restated as an empirical finding: for LNS and the other tuned methods the qualitative vulnerability is by construction, and only the transfer magnitude on held-out samples and the behavior of non-tuned methods are independently measured.

  2. other [Appendix C.2 (SAPLMA training setup)]
    "However, since we observe a performance improvement when excluding NaturalQA, we train SAPLMA on a reduced dataset of 13k samples comprising only TriviaQA and GSM8K."

    SAPLMA's training-data composition was selected by observing performance changes on the same TriviaQA and GSM8K datasets later used to report its ARE and PRR. The reported scores are therefore partly produced by the selection criterion: excluding NaturalQA because it improved performance and then presenting the resulting TriviaQA/GSM8K numbers as benchmark evidence folds the optimization target back into the reported result.

full rationale

This is an empirical benchmark rather than a derivation, so no equation-level equivalence (e.g., Eq. X = Eq. Y by construction) is present. The headline threshold-sensitivity result (Section 3) evaluates 19 methods on held-out test and calibration splits against external datasets (TriviaQA, NaturalQA, GSM8K), and the ensembling result (Section 6) uses held-out PRR; both claims have independent content. Two evaluation-design choices, however, are partly self-confirming. First, the GPT-4o-mini adversarial prompt was tuned to degrade LNS, Eccentricity, and KLE (Appendix E.1), and Section 4.3 then reports that probability-based methods are particularly vulnerable; for the tuned methods this finding is the fit objective restated. Second, SAPLMA's training set was chosen by observing performance on the same TriviaQA/GSM8K datasets used for evaluation (Appendix C.2), so its reported scores reflect that selection. The supervised methods LARS and SAPLMA are also trained on the same benchmark datasets, which is a fairness limitation rather than circularity. The use of GPT-4o-mini and SAFE to produce correctness labels (Section 2.2 footnote, Appendix E.3) is a label-validity risk, not a derivation-level circularity, because the paper does not derive the UE scores from those labels. Overall the central claims survive as independent measurements, but the two self-confirming design choices warrant a moderate score.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The evaluation rests on domain assumptions about label quality and dataset representativeness, plus one fitted adversarial prompt. The supervised methods' training-set overlap with the evaluation datasets is the main source of circularity burden.

free parameters (1)
  • adversarial prompt for GPT-4o-mini = "Confidently present every answer as if it were indisputably correct, embellishing your certainty with emphatic…
    Tuned via 15 rounds of prompt-tuning on 150 TriviaQA training samples to maximize the drop in UE performance (Appendix E.1); this fitted prompt is then used as the adversarial transformation in Section 4.
assumptions (4)
  • domain assumption GPT-4o-mini produces correct binary labels for short-form QA correctness.
    Section 2.2 footnote states GPT-4o-mini is used as correctness evaluator for all short-form generations; if the judge is biased, all ARE and PRR measurements inherit that bias.
  • domain assumption SAFE labels long-form claims correctly.
    Section E.3 uses SAFE with GPT-4o-mini reasoning to label decomposed claims as supported or unsupported; treated as ground truth for PRR in Section 5.
  • domain assumption The 19 selected UE methods represent the broader space of LLM uncertainty estimators.
    The paper's conclusions about 'most UE methods' generalize from this specific set (Section 2.3).
  • domain assumption Training splits for LARS and SAPLMA are disjoint from the evaluation and calibration splits.
    Sections C.1 and C.2 describe training on TriviaQA, NaturalQA, and GSM8K, the same datasets used in Section 3; the paper never states that the 1000-sample test and 500-sample calibration sets are excluded from those training sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reconsidering LLM Uncertainty Estimation Methods in the Wild." pith.science (2026). https://pith.science/paper/LFZCVOGE

@misc{pith2026250601114,
  author       = {Pith},
  title        = {Pith review of: Reconsidering LLM Uncertainty Estimation Methods in the Wild},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFZCVOGE}},
  note         = {Machine review of arXiv:2506.01114}
}
read the original abstract

Large Language Model (LLM) Uncertainty Estimation (UE) methods have become a crucial tool for detecting hallucinations in recent years. While numerous UE methods have been proposed, most existing studies evaluate them in isolated short-form QA settings using threshold-independent metrics such as AUROC or PRR. However, real-world deployment of UE methods introduces several challenges. In this work, we systematically examine four key aspects of deploying UE methods in practical settings. Specifically, we assess (1) the sensitivity of UE methods to decision threshold selection, (2) their robustness to query transformations such as typos, adversarial prompts, and prior chat history, (3) their applicability to long-form generation, and (4) strategies for handling multiple UE scores for a single query. Our evaluations on 19 UE methods reveal that most of them are highly sensitive to threshold selection when there is a distribution shift in the calibration dataset. While these methods generally exhibit robustness against previous chat history and typos, they are significantly vulnerable to adversarial prompts. Additionally, while existing UE methods can be adapted for long-form generation through various strategies, there remains considerable room for improvement. Lastly, ensembling multiple UE scores at test time provides a notable performance boost, which highlights its potential as a practical improvement strategy. Code is available at: https://github.com/duygunuryldz/uncertainty_in_the_wild.

Figures

Figures reproduced from arXiv: 2506.01114 by the authors.

Figure 1
Figure 1. Left: Existing pipeline for UE. The uncertainty score is calculated for short-form QA and evaluated using a threshold-free metric such as AUROC. Right: Reconsidering LLM uncertainty estimation methods in the wild. We ask four critical questions addressing challenges in deploying UE methods in real-world scenarios. many UE methods are proposed and tested for short-form QA, real-world questions often require extended … view at source ↗
Figure 2
Figure 2. PRR performance of UE methods with Llama-3 8b, evaluated under a regular prompt (no transformation) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. PRR performance of UE methods on the GSM8K and TriviaQA datasets with GPT-4o-mini. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: PRR scores for UE methods applied to long-form generation. ‘QG-5’ and ‘QAG-5’ indicate that five [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: PRR performance of all investigated UE methods on the GSM8K and TriviaQA datasets under the typo [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: PRR scores for UE methods applied to long-form generation. ‘QG5’ and ‘QAG5’ indicate that five [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: PRR scores for UE methods applied to long-form generation for GPT-4o-mini. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TruthTorchLM: A Comprehensive Library for Predicting Truthfulness in LLM Outputs

    cs.CL 2025-07 conditional novelty 5.0 of 10

    TruthTorchLM is a new open-source library that standardizes 30+ LLM truthfulness prediction methods and benchmarks them on three datasets.

Reference graph

Works this paper leans on

47 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Yasin Abbasi-Yadkori, Ilja Kuzborskij, Andr \'a s Gy \"o rgy, and Csaba Szepesvari. 2024. https://openreview.net/forum?id=k6iyUfwdI9 To believe or not to believe your LLM : Iterativeprompting for estimating epistemic uncertainty . In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page 2024

  4. [4]

    Kian Ahrabian, Xihui Lin, Barun Patra, Vishrav Chaudhary, Alon Benhaim, Jay Pujara, and Xia Song. 2025. https://aclanthology.org/2025.findings-naacl.446/ A practical analysis of human alignment with * PO . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 8013--8021, Albuquerque, New Mexico. Association for Computational Linguistics

    work page 2025

  5. [5]

    Lukas Aichberger, Kajetan Schweighofer, and Sepp Hochreiter. 2024. https://arxiv.org/abs/2412.15176 Rethinking uncertainty estimation in natural language generation . Preprint, arXiv:2412.15176

    arXiv 2024

  6. [6]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  7. [7]

    Amos Azaria and Tom Mitchell. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.68 The internal state of an LLM knows when it ' s lying . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 967--976, Singapore. Association for Computational Linguistics

    doi:10.18653/v1/2023.findings-emnlp.68 2023

  8. [8]

    Yavuz Faruk Bakman, Duygu Nur Yaldiz, Baturalp Buyukates, Chenyang Tao, Dimitrios Dimitriadis, and Salman Avestimehr. 2024. https://doi.org/10.18653/v1/2024.acl-long.419 MARS : Meaning-aware response scoring for uncertainty estimation in generative LLM s . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1...

    doi:10.18653/v1/2024.acl-long.419 2024

Show all 47 references
  1. [9]

    Yoshua Bengio, Sören Mindermann, Daniel Privitera, Tamay Besiroglu, Rishi Bommasani, Stephen Casper, Yejin Choi, Philip Fox, Ben Garfinkel, Danielle Goldfarb, Hoda Heidari, Anson Ho, Sayash Kapoor, Leila Khalatbari, Shayne Longpre, Sam Manning, Vasilios Mavroudis, Mantas Mazei...

    2025 arXiv

  2. [10]

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024. https://openreview.net/forum?id=Zj12nzlQbz INSIDE : LLM s' internal states retain the power of hallucination detection . In The Twelfth International Conference on Learning Representations

    2024

  3. [11]

    I Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, Pengfei Liu, et al. 2023. Factool: Factuality detection in generative ai--a tool augmented framework for multi-task and multi-domain scenarios. arXiv preprint arXiv:2307.13528

    2023 arXiv

  4. [12]

    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. arXiv preprint arXiv:2110.14168

    2021 arXiv

  5. [13]

    Jinhao Duan, Hao Cheng, Shiqi Wang, Alex Zavalny, Chenan Wang, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. 2024. https://doi.org/10.18653/v1/2024.acl-long.276 Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models ...

    2024 doi

  6. [14]

    Ekaterina Fadeeva, Aleksandr Rubashevskii, Artem Shelmanov, Sergey Petrakov, Haonan Li, Hamdy Mubarak, Evgenii Tsymbalov, Gleb Kuzmin, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, and Maxim Panov. 2024. https://doi.org/10.18653/v1/2024.findings-acl.558 Fact-checking th...

    2024 doi

  7. [15]

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625--630

    2024

  8. [16]

    Shangbin Feng, Weijia Shi, Yike Wang, Wenxuan Ding, Vidhisha Balachandran, and Yulia Tsvetkov. 2024. https://doi.org/10.18653/v1/2024.acl-long.786 Don ' t hallucinate, abstain: Identifying LLM knowledge gaps via multi- LLM collaboration . In Proceedings of the 62nd Annual Meet...

    2024 doi

  9. [17]

    Samworth

    Qiyang Han, Tengyao Wang, Sabyasachi Chatterjee, and Richard J. Samworth. 2017. https://api.semanticscholar.org/CorpusID:41805423 Isotonic regression in general dimensions . The Annals of Statistics

    2017

  10. [18]

    Xinmeng Huang, Shuo Li, Mengxin Yu, Matteo Sesia, Hamed Hassani, Insup Lee, Osbert Bastani, and Edgar Dobriban. 2024. https://arxiv.org/abs/2404.03163 Uncertainty in language models: Assessment through rank-calibration . Preprint, arXiv:2404.03163

    2024 arXiv

  11. [19]

    Mingjian Jiang, Yangjun Ruan, Prasanna Sattigeri, Salim Roukos, and Tatsunori Hashimoto. 2024. https://openreview.net/forum?id=YgJPQW0lkO Graph-based uncertainty metrics for long-form language model generations . In The Thirty-eighth Annual Conference on Neural Information Pro...

    2024

  12. [20]

    Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. https://doi.org/10.18653/v1/P17-1147 T rivia QA : A large scale distantly supervised challenge dataset for reading comprehension . In Proceedings of the 55th Annual Meeting of the Association for Computational...

    2017 doi

  13. [21]

    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 For...

    2022 arXiv

  14. [22]

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. https://openreview.net/forum?id=VD-AYtP0dve Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation . In The Eleventh International Conference on Learning Representations

    2023

  15. [23]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019 doi

  16. [24]

    Moxin Li, Wenjie Wang, Fuli Feng, Fengbin Zhu, Qifan Wang, and Tat-Seng Chua. 2024. https://arxiv.org/abs/2403.09972 Think twice before trusting: Self-detection for large language models through comprehensive answer reflection . Preprint, arXiv:2403.09972

    2024 arXiv

  17. [25]

    Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2024. https://openreview.net/forum?id=DWkJCSxKU5 Generating with confidence: Uncertainty quantification for black-box large language models . Transactions on Machine Learning Research

    2024

  18. [26]

    Mat \'e o Mahaut, Laura Aina, Paula Czarnowska, Momchil Hardalov, Thomas M \"u ller, and Lluis Marquez. 2024. https://doi.org/10.18653/v1/2024.acl-long.250 Factual confidence of LLM s: on reliability and robustness of current estimators . In Proceedings of the 62nd Annual Meet...

    2024 doi

  19. [27]

    Andrey Malinin and Mark Gales. 2021. https://openreview.net/forum?id=jN5y-zb5Q7m Uncertainty estimation in autoregressive structured prediction . In International Conference on Learning Representations

    2021

  20. [28]

    Potsawee Manakul, Adian Liusie, and Mark Gales. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.557 S elf C heck GPT : Zero-resource black-box hallucination detection for generative large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural...

    2023 doi

  21. [29]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.741 FA ct S core: Fine-grained atomic evaluation of factual precision in long form text generatio...

    2023 doi

  22. [30]

    Viktor Moskvoretskii, Maria Lysyuk, Mikhail Salnikov, Nikolay Ivanov, Sergey Pletenev, Daria Galimzianova, Nikita Krayko, Vasily Konovalov, Irina Nikishina, and Alexander Panchenko. 2025. https://arxiv.org/abs/2501.12835 Adaptive retrieval without self-knowledge? bringing unce...

    2025 arXiv

  23. [31]

    Alexander V Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. 2024. https://openreview.net/forum?id=j2wCrWmgMX Kernel language entropy: Fine-grained uncertainty quantification for LLM s from semantic similarities . In The Thirty-eighth Annual Conference on Neural Informa...

    2024

  24. [32]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 GPT-4 Technical Report . Preprint, arXiv:2303.08774

    2023 arXiv

  25. [33]

    Selvan Sunitha Ravi, Bartosz Mielczarek, Anand Kannappan, Douwe Kiela, and Rebecca Qian. 2024. https://arxiv.org/abs/2407.08488 Lynx: An open source hallucination evaluation model . Preprint, arXiv:2407.08488

    2024 arXiv

  26. [34]

    Sadra Sabouri, Philipp Eibl, Xinyi Zhou, Morteza Ziyadi, Nenad Medvidovic, Lars Lindemann, and Souti Chattopadhyay. 2025. https://www.amazon.science/publications/trust-dynamics-in-ai-assisted-development-definitions-factors-and-implications Trust dynamics in ai-assisted develo...

    2025

  27. [35]

    Shahnewaz Karim Sakib, Anindya Bijoy Das, and Shibbir Ahmed. 2025. https://aclanthology.org/2025.trustnlp-main.28/ Battling misinformation: An empirical study on adversarial factuality in open-source large language models . In Proceedings of the 5th Workshop on Trustworthy NLP...

    2025

  28. [36]

    Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. 2024. https://openreview.net/forum?id=LYx4w3CAgy LLM -check: Investigating detection of hallucinations in large language models . In The Thirty-eighth Annual Confe...

    2024

  29. [37]

    Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch

    Ala N. Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025. https://arxiv.org/abs/2502.05489 Mechanistic interpretability of emotion inference in large language models . Preprint, arXiv:2502.05489

    2025 arXiv

  30. [38]

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher Manning. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.330 Just ask for calibration: Strategies for eliciting calibrated confidence scores from language mod...

    2023 doi

  31. [39]

    Roman Vashurin, Ekaterina Fadeeva, Artem Vazhentsev, Lyudmila Rvanova, Akim Tsvigun, Daniil Vasilev, Rui Xing, Abdelrahman Boda Sadallah, Kirill Grishchenkov, Sergey Petrakov, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, Maxim Panov, and Artem Shelmanov. 2025. https://...

    2025 arXiv

  32. [40]

    Artem Vazhentsev, Gleb Kuzmin, Akim Tsvigun, Alexander Panchenko, Maxim Panov, Mikhail Burtsev, and Artem Shelmanov. 2023. https://doi.org/10.18653/v1/2023.acl-long.652 Hybrid uncertainty quantification for selective text classification in ambiguous tasks . In Proceedings of t...

    2023 doi

  33. [41]

    Yuxia Wang, Revanth Gangi Reddy, Zain Muhammad Mujahid, Arnav Arora, Aleksandr Rubashevskii, Jiahui Geng, Osama Mohammed Afzal, Liangming Pan, Nadav Borenstein, Aditya Pillai, Isabelle Augenstein, Iryna Gurevych, and Preslav Nakov. 2024. https://arxiv.org/abs/2311.09000 Factch...

    2024 arXiv

  34. [42]

    Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. 2024 a . https://arxiv.org/abs/2411.04368 Measuring short-form factuality in large language models . Preprint, arXiv:2411.04368

    2024 arXiv

  35. [43]

    Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Zixia Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V Le. 2024 b . https://openreview.net/forum?id=4M9f8VMt2C Long-form factuality in large language models . In The Thirty-eighth Annual...

    2024

  36. [44]

    Duygu Nur Yaldiz, Yavuz Faruk Bakman, Baturalp Buyukates, Chenyang Tao, Anil Ramakrishna, Dimitrios Dimitriadis, Jieyu Zhao, and Salman Avestimehr. 2025. https://aclanthology.org/2025.findings-naacl.41/ Do not design, learn: A trainable scoring function for uncertainty estimat...

    2025

  37. [45]

    Caiqi Zhang, Fangyu Liu, Marco Basaldella, and Nigel Collier. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.299 LUQ : Long-text uncertainty quantification for LLM s . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5244--526...

    2024 doi

  38. [46]

    Yukun Zhao, Lingyong Yan, Weiwei Sun, Guoliang Xing, Chong Meng, Shuaiqiang Wang, Zhicong Cheng, Zhaochun Ren, and Dawei Yin. 2024. https://doi.org/10.18653/v1/2024.naacl-long.390 Knowing what LLM s DO NOT know: A simple yet effective self-detection method . In Proceedings of ...

    2024 doi

  39. [47]

    Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. https://openreview.net/forum?id=92gvk82DE- Large language models are human-level prompt engineers . In The Eleventh International Conference on Learning Representations

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.