Pith. sign in

REVIEW 3 major objections 6 minor 35 references

TECP: Token-Entropy Conformal Prediction for LLMs

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read TECP claims that token entropy, calibrated with conformal quantiles, constructs LLM answer sets whose coverage is at least 1−α.

desk verdict The TECP coverage guarantee doesn't survive contact with the paper's own procedure: the calibration scores are not one per exchangeable unit, so the split-CP quantile controls the wrong event, and the 'logit-free' claim is contradicted by Eq. (4). read the letter →

arxiv 2509.00461 v2 pith:FT7VHKTE submitted 2025-08-30 cs.CL cs.AI

Beining Xu , Yongming Lu This is my paper
classification cs.CLcs.AI
keywords conformalpredictionuncertaintyquantificationtokenentropysetslargelanguagemodelscoverageguaranteeblack-boxLLMquestionanswering
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

This paper argues that the per-token entropy of sampled LLM generations can serve as a calibrated uncertainty score for open-ended question answering. The proposed method, TECP, collects token-entropy scores from generations that match the ground truth in a calibration split, takes a conformal quantile, and then keeps every test candidate whose entropy is below that threshold. It claims this construction yields prediction sets with nominal coverage of at least 1−α for the true answer under standard exchangeability assumptions. If correct, black-box API users of LLMs would gain a principled, model-agnostic way to control error rates while shrinking the answer set as confidence grows. The paper itself notes in its limitations that the entropy scores come from output probabilities, so the word 'logit-free' in the abstract is not reflected in the method as written.

What carries the argument

The central object is the token-entropy nonconformity score, U(ŷ_m) = −Σ_t Σ_v p_t(v) log p_t(v), combined with the split-conformal quantile threshold of Eq. (6). The score ranks candidate answers by the model's internal predictive uncertainty, and the quantile calibration converts that ranking into a set with an error-rate interpretation. The paper claims this machinery is logit-free and reference-free, though Eq. (4) actually requires access to the token predictive distribution.

What would settle it

Run TECP on a benchmark where the proportion of questions with no semantically correct candidate differs sharply between calibration and test splits, and compare realized coverage to 1−α. If coverage falls below 1−α when test questions without a correct candidate are included, the stated unconditional guarantee is false; if coverage only holds after filtering those test questions out, the guarantee is conditional, not the nominal one claimed.

Watch

Extended reading notes

Core claim

The paper's central claim is that token-level cumulative entropy is a better nonconformity score for LLM prediction sets than frequency-based semantic consistency, and that plugging it into split conformal prediction gives formal coverage guarantees. Concretely, for each sampled candidate ŷ_m the uncertainty score is U(ŷ_m) = −Σ_t Σ_v p_t(v) log p_t(v), the summed entropy of the model's predictive distribution over tokens. The calibration set contributes one entropy score per semantically correct candidate, the threshold q̂_α is the higher-interpolated ⌈(1−α)(n+1)⌉-th order statistic, and a test candidate enters the prediction set when its entropy is at most q̂_α. The paper asserts that by c

Load-bearing premise

The coverage guarantee holds only if the empirical quantile of per-candidate entropy scores from calibration examples that passed the semantic filter controls the probability that a test prediction set contains at least one correct candidate; the paper asserts this by construction without deriving it from the standard split-conformal exchangeability condition.

Editorial extensions

If this is right

  • A user can set an allowable error rate α and receive a set of candidate answers whose miscoverage is nominally at most α, under the stated exchangeability and filtering assumptions.
  • Prediction-set size becomes a meaningful uncertainty signal: larger sets signal lower confidence, smaller sets signal higher confidence, and the size shrinks monotonically with α in the reported results.
  • The method transfers across model families and scales without retraining or internal hidden-state access, as long as token-level probabilities are obtainable.
  • TECP's empirical coverage is more stable across random calibration-test splits than the ConU baseline, which the paper attributes to token entropy avoiding the hallucination bias of frequency-based scores.
  • In high-stakes QA applications, TECP would replace heuristic confidence scores with a threshold that carries a statistical coverage guarantee.

Reading between the lines

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

  • Because calibration collects one score per semantically correct candidate while the test event is 'the set contains at least one correct candidate,' the coverage argument is not the standard split-conformal one; a rigorous exchangeability-based guarantee would require one score per example, and the paper does not supply that derivation.
  • A genuinely logit-free implementation would need to estimate token entropy from sampling statistics rather than from p_t(v); the paper's Eq. (4) does not provide such an estimate.
  • The calibration set is filtered to 'assessable' samples via a semantic threshold τ, which makes the guarantee conditional on a question having a correct candidate; applying the same quantile to test questions that may be unassessable is an extrapolation worth testing explicitly.
  • A testable extension would apply TECP under distribution shift or non-exchangeable data and check whether weighted conformal corrections, as developed in the beyond-exchangeability literature, are needed to preserve coverage.
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

3 major / 6 minor

Summary. TECP proposes to build prediction sets for open-ended LLM question answering by combining token-level entropy as a nonconformity score with split conformal prediction. For each input, M candidate outputs are generated; the uncertainty score U(ŷ_m) is defined either as a semantic-consistency combination (Eq. 3) or as cumulative token entropy (Eq. 4). Calibration examples are first filtered to those containing at least one semantically correct candidate, the nonconformity multiset R in Eq. (5) collects the scores of all semantically correct candidates, and the threshold q̂_α is the empirical quantile of R (Eq. 6). A test prediction set Γ(x) consists of all candidates whose score is at most q̂_α (Eq. 7). The paper claims that, by construction, Γ(x) has nominal coverage at least 1−α under standard exchangeability assumptions. Experiments on TriviaQA and CoQA with six open LLMs compare TECP with ConU in terms of empirical miscoverage rate and average prediction set size.

Significance. The paper targets a real problem: providing finite-sample risk control for free-form LLM outputs rather than relying on heuristic uncertainty scores. The experimental scope—six models, two QA benchmarks, comparison to a recent conformal baseline—is commendable, and the authors correctly identify that held-out calibration, rather than test-set reuse, is needed. However, the central formal claim is not established. The construction in §3.3 is not a valid split-conformal argument: the calibration multiset contains multiple dependent scores per unit, and the filtering step conditions on ground truth. In addition, the abstract's logit-free/black-box claim is contradicted by Eq. (4), which requires the model's full per-token predictive distribution. Because the coverage guarantee is the paper's main selling point, the current manuscript does not support its stated contribution.

major comments (3)
  1. [§3.3, Eqs (5)–(7)] The nominal coverage claim does not follow from split conformal prediction. R in Eq. (5) contains one score for every semantically correct candidate, so a calibration unit (x_i, y*_i) contributes K_i dependent scores that share both the input and the ground truth; n=|R| is not the number of exchangeable units. The standard split-CP quantile lemma requires one exchangeable score per calibration unit and controls a single-score event. In contrast, Eq. (7) requires the union event 'at least one correct candidate has U ≤ q̂_α' over K_test dependent scores, and the filtering step into D_filtered conditions on ground truth via τ. A concrete failure shows the threshold can be biased: exchangeable units with two correct candidates of score 0.4 with probability 0.4 and one correct candidate of score 0.6 with probability 0.6 yield R with about 57% scores 0.4, so q̂=0.4 at α=0.5, but only 40% of te
  2. [§3.1 Eq. (2) vs §3.3 Eq. (7)] There is a mismatch between the formal coverage event and the evaluated event. Eq. (2) promises P(y* ∈ Γ(x)) ≥ 1−α, but Γ(x) in Eq. (7) is a subset of sampled candidates, and the reference answer y* is not itself a generated candidate. The experiments instead evaluate EMR, defined as containment of at least one output whose semantic similarity to y* exceeds a threshold (0.7 in §4.1), while the calibration multiset in Eq. (5) uses 'semantically correct' candidates with an unspecified matching threshold. These are different events, and the guarantee in Eq. (2) is therefore vacuous or misstated. The paper should define the target event explicitly and state any conditioning, e.g., restriction to assessable test examples.
  3. [Abstract/§3.2 Eq. (4)/Limitations] The paper's headline claim that token entropy is 'logit-free' and suitable for black-box LLM access is contradicted by Eq. (4), which requires p_t(v), the model's predictive distribution over the vocabulary at every token position. That is exactly the kind of internal probability signal the abstract says is inaccessible. The Limitations section concedes this: it describes the approach as a 'white-box conformal prediction framework' relying on 'uncertainty scores derived from model output probabilities.' This contradiction undermines the stated contribution; either the black-box/logit-free claim must be removed, or the scoring function must be replaced with one computable from sampled text alone.
minor comments (6)
  1. [Abstract] The first sentence says 'internal model signals are accessible' under black-box constraints; this should presumably read 'inaccessible.' Also, 'episodic uncertainty' in the abstract appears to be a typo for 'epistemic uncertainty.'
  2. [§3.2, Eqs (3)–(4)] U(ŷ_m) is defined twice with different meanings: Eq. (3) is a convex combination of semantic-consistency frequency and average similarity, while Eq. (4) is cumulative token entropy. The relationship between these definitions and which one is used in the experiments should be clarified.
  3. [§4.3] The ablation study is described as varying the calibration–test split ratio (0.3, 0.5, 0.7), but the text refers to Figures 1 and 3, which show alpha variation rather than split-ratio variation. The actual ablation results appear to be missing or unreferenced.
  4. [§4.1 and Figures 1–4] The acronym EMR is defined as 'Expected Metric Recall' in the metrics list but labeled 'Empirical Miscoverage Rate' in the figures. These are different quantities; the notation should be made consistent.
  5. [§4.1, Table 1] Table 1 reports average prediction set sizes without standard deviations, confidence intervals, or the number of test examples, despite the text claiming stability across random seeds. This weakens the empirical comparison.
  6. [§4.1] Reproducibility details are incomplete: the sampling size M=10 is mentioned, but temperature, top-p, number of calibration samples, and the exact split ratios used in the main experiments are not specified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TECP follows the standard split-CP calibration protocol, and its central coverage claim, while statistically unsupported, does not reduce to its inputs by construction.

full rationale

The paper's threshold is computed on a held-out calibration split and applied to a separate test split, with coverage evaluated on test data. This is the standard conformal prediction protocol, not a circular reuse of the test set or a fitted parameter renamed as a prediction. The main formal problem is that Eq. (5) pools multiple dependent nonconformity scores per calibration unit and conditions on ground-truth correctness via the threshold τ, so the exchangeability argument for the stated coverage guarantee is not valid; however, this is a statistical flaw, not a circularity. Additional inconsistencies — Eq. (3) defining uncertainty via semantic consistency versus Eq. (4) using token-level entropy, and the abstract's 'logit-free' claim contradicted by Eq. (4) and the Limitations section's admission of white-box reliance on output probabilities — are internal contradictions rather than circular reductions. No load-bearing self-citations or imported uniqueness theorems appear. Under the specified rules, a non-circular but flawed derivation should receive a low circularity score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method depends on hand-chosen thresholds for filtering and correctness, on the sampling count M, and on an unproved assumption that correct-candidate score quantiles control test-time coverage. The most consequential axiom is the availability of token-level distributions, which contradicts the paper's black-box framing.

free parameters (4)
  • lambda (Eq 3)
    Convex combination weight in the semantic-consistency score that is then replaced by token entropy; value never specified or reported.
  • tau_filter (Section 3.3 step 1) = 0.8-0.9
    Semantic-matching threshold that decides whether a sample is assessable and enters calibration; changes the calibration multiset.
  • tau_correct (Section 4.1) = 0.7
    Similarity threshold that defines a generated answer as correct; all coverage and EMR numbers depend on this choice.
  • M (Section 3.1) = 10
    Fixed number of sampled generations per prompt; fixes the candidate set size and therefore the maximum prediction set size.
assumptions (5)
  • standard math Calibration and test examples are exchangeable
    Invoked in Section 3.3 via 'standard exchangeability assumptions'; standard for split conformal prediction.
  • domain assumption Token-level predictive distribution p_t(v) is available
    Eq (4) sums over v in V, which requires per-token probabilities from the model, contradicting the logit-free black-box claim in the abstract.
  • ad hoc to paper Quantile of correct-candidate scores yields coverage of a correct candidate in the test set
    No proof is given; standard CP exchangeability does not apply because Eq (5) takes one score per correct candidate rather than one score per example.
  • domain assumption Filtering to assessable samples preserves the coverage guarantee
    Section 3.3 step 1 discards samples without a candidate similar to ground truth; the proportion discarded is not reported.
  • domain assumption DistillRoBERTa cross-encoder at threshold 0.7 correctly identifies semantic correctness
    Section 4.1 uses this as the sole correctness oracle; errors in this oracle directly propagate to all coverage numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TECP: Token-Entropy Conformal Prediction for LLMs." pith.science (2026). https://pith.science/paper/FT7VHKTE

@misc{pith2026250900461,
  author       = {Pith},
  title        = {Pith review of: TECP: Token-Entropy Conformal Prediction for LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FT7VHKTE}},
  note         = {Machine review of arXiv:2509.00461}
}
read the original abstract

Uncertainty quantification (UQ) for open-ended language generation remains a critical yet underexplored challenge, especially under black-box constraints where internal model signals are inaccessible. In this paper, we introduce Token-Entropy Conformal Prediction (TECP), a novel framework that leverages token-level entropy as a logit-free, reference-free uncertainty measure and integrates it into a split conformal prediction (CP) pipeline to construct prediction sets with formal coverage guarantees. Unlike existing approaches that rely on semantic consistency heuristics or white-box features, TECP directly estimates epistemic uncertainty from the token entropy structure of sampled generations and calibrates uncertainty thresholds via CP quantiles to ensure provable error control. Empirical evaluations across six large language models and two benchmarks (CoQA and TriviaQA) demonstrate that TECP consistently achieves reliable coverage and compact prediction sets, outperforming prior self-consistency-based UQ methods. Our method provides a principled and efficient solution for trustworthy generation in black-box LLM settings.

Figures

Figures reproduced from arXiv: 2509.00461 by the authors.

Figure 1
Figure 1. EMR vs. Alpha for six models on the TriviaQA dataset using TECP it captures the model’s intrinsic uncertainty during genera￾tion, avoiding the overconfidence bias and hallucination sen￾sitivity inherent in frequency-based diversity measures.This reference-free criterion evaluates each candidate’s confi￾dence independently of the ground truth. 3.3. Uncertainty Calibration and Prediction-Set Construction Given an unce… view at source ↗
Figure 2
Figure 2. EMR vs. Alpha for six models on the TriviaQA dataset using ConU 4. Experiments 4.1. Experimental Setup Backbone LLMs and Evaluation Tasks. Since we incor￾porate token entropy into the conformal prediction frame￾work for open-ended Question Answering (QA) tasks, it is necessary to evaluate the performance of the proposed architecture rigorously. To this end, we select a diverse set of open-source large language model… view at source ↗
Figure 3
Figure 3. EMR vs. Alpha for six models on the CoQA dataset using TECP paired with its corresponding reference answer and assessed using a semantic matching model. To this end, we leverage a DistillRoBERTa-based cross-encoder, which scores the semantic closeness between the two texts. A prediction is deemed correct if its similarity score surpasses a fixed threshold of 0.7, ensuring only semantically faithful outputs are retai… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: EMR vs. Alpha for six models on the CoQA dataset using ConU different proportions of calibration data. All evaluations are performed under a fixed risk level of α = 0.1, and the results are averaged over 100 random seeds to mitigate the influence of sampling variabilit…
Figure 5
Figure 5. Figure 5: CoQA 0.3 0.5 0.7 Split Ratio 0.92 0.94 0.96 0.98 1.00 1.02 Coverage Rate (1 EMR) 0.97 0.97 0.97 0.96 0.96 0.96 0.940.94 0.940.94 0.940.94 1.00 1.00 1.00 0.96 0.96 0.96 EMR Coverage at =0.1 TriviaQA Llama-3.1-8B Llama-3.2-1B Qwen-2.5-3B Qwen-2.5-7B Vicuna-13B Vicuna-7B …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    N., Bates, S., et al

    Angelopoulos, A. N., Bates, S., et al. Conformal prediction: A gentle introduction. Foundations and trends in machine learning , 2023

    work page 2023

  3. [3]

    F., Candes, E

    Barber, R. F., Candes, E. J., Ramdas, A., and Tibshirani, R. J. Conformal prediction beyond exchangeability. The Annals of Statistics, 2023

    work page 2023

  4. [4]

    Does machine unlearning truly remove model knowledge? a framework for auditing unlearning in llms

    Chen, H., Zhang, Y., Bi, Y., Zhang, Y., Liu, T., Bi, J., Lan, J., Gu, J., Grosser, C., Krompass, D., et al. Does machine unlearning truly remove model knowledge? a framework for auditing unlearning in llms. arXiv preprint arXiv:2505.23270, 2025

    work page arXiv 2025

  5. [5]

    Can llms’ tuning methods work in medical multimodal domain? In MICCAI, 2024 a

    Chen, J., Jiang, Y., Yang, D., Li, M., Wei, J., Qian, Z., and Zhang, L. Can llms’ tuning methods work in medical multimodal domain? In MICCAI, 2024 a

    work page 2024

  6. [6]

    Miss: A generative pre-training and fine-tuning approach for med-vqa

    Chen, J., Yang, D., Jiang, Y., Lei, Y., and Zhang, L. Miss: A generative pre-training and fine-tuning approach for med-vqa. In International Conference on Artificial Neural Networks, 2024 b

    work page 2024

  7. [7]

    Detecting and evaluating medical hallucinations in large vision language models

    Chen, J., Yang, D., Wu, T., Jiang, Y., Hou, X., Li, M., Wang, S., Xiao, D., Li, K., and Zhang, L. Detecting and evaluating medical hallucinations in large vision language models. arXiv preprint arXiv:2406.10185, 2024 c

    arXiv 2024

  8. [8]

    Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models

    Duan, J., Cheng, H., Wang, S., Zavalny, A., Wang, C., Xu, R., Kailkhura, B., and Xu, K. 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 Computational Linguistics (Volume 1: Long Papers), 2024

    work page 2024

Show all 35 references
  1. [9]

    Detecting hallucinations in large language models using semantic entropy

    Farquhar, S., Kossen, J., Kuhn, L., and Gal, Y. Detecting hallucinations in large language models using semantic entropy. Nature, 2024

    2024

  2. [10]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 2025

    2025

  3. [11]

    and Mazoyer, A

    Iutzeler, F. and Mazoyer, A. Risk-controlling prediction with distributionally robust optimization. Transactions on Machine Learning Research, 2025. ISSN 2835-8856

    2025

  4. [12]

    Comt: Chain-of-medical-thought reduces hallucination in medical report generation

    Jiang, Y., Chen, J., Yang, D., Li, M., Wang, S., Wu, T., Li, K., and Zhang, L. Comt: Chain-of-medical-thought reduces hallucination in medical report generation. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE

    2025

  5. [13]

    R., Showk, S., et al

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Dodds, Z., German, M., Johnston, S., Kravec, S., Lovitt, L. R., Showk, S., et al. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221, 2022

    2022 arXiv

  6. [14]

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation

    Kuhn, L., Gal, Y., and Farquhar, S. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023

    2023

  7. [15]

    Knowledge graphs, large language models, and hallucinations: An nlp perspective

    Lavrinovics, E., Biswas, R., Bjerva, J., and Hose, K. Knowledge graphs, large language models, and hallucinations: An nlp perspective. Journal of Web Semantics, 2025

    2025

  8. [16]

    Truthfulqa: Measuring how models mimic human falsehoods

    Lin, S., Hilton, J., and Evans, O. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022

    2022

  9. [17]

    Mv-debate: Multi-view agent debate with dynamic reflection gating for multimodal harmful content detection in social media

    Lu, R., Bi, J., Ma, Y., Xiao, F., Du, Y., and Tian, Y. Mv-debate: Multi-view agent debate with dynamic reflection gating for multimodal harmful content detection in social media. arXiv preprint arXiv:2508.05557, 2025

    2025 arXiv

  10. [18]

    and Gales, M

    Malinin, A. and Gales, M. Uncertainty estimation in autoregressive structured prediction. In International Conference on Learning Representations, 2021

    2021

  11. [19]

    and Calderon-Ramirez, S

    Mora-Cross, M. and Calderon-Ramirez, S. Uncertainty estimation in large language models to support biodiversity conservation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volu...

    2024

  12. [20]

    Conformal prediction with cellwise outliers: A detect-then-impute approach

    Peng, Q., Bao, Y., Ren, H., Wang, Z., and Zou, C. Conformal prediction with cellwise outliers: A detect-then-impute approach. In Forty-second International Conference on Machine Learning, 2025

    2025

  13. [21]

    and Miikkulainen, R

    Qiu, X. and Miikkulainen, R. Semantic density: Uncertainty quantification for large language models through confidence measurement in semantic space. Advances in neural information processing systems, 2024

    2024

  14. [22]

    Backdoor cleaning without external guidance in mllm fine-tuning

    Rong, X., Huang, W., Liang, J., Bi, J., Xiao, X., Li, Y., Du, B., and Ye, M. Backdoor cleaning without external guidance in mllm fine-tuning. arXiv preprint arXiv:2505.16916, 2025

    2025 arXiv

  15. [23]

    Snell, J. C. and Griffiths, T. L. Conformal prediction as bayesian quadrature. In Forty-second International Conference on Machine Learning, 2025

    2025

  16. [24]

    Large language models: A comprehensive survey on architectures, applications, and challenges

    Veeramachaneni, V. Large language models: A comprehensive survey on architectures, applications, and challenges. Advanced Innovations in Computer Programming Languages, 2025

    2025

  17. [25]

    Uncertainty quantification for safe and reliable autonomous vehicles: A review of methods and applications

    Wang, K., Shen, C., Li, X., and Lu, J. Uncertainty quantification for safe and reliable autonomous vehicles: A review of methods and applications. IEEE Transactions on Intelligent Transportation Systems, 2025 a

    2025

  18. [26]

    Sample then identify: A general framework for risk control and assessment in multimodal large language models

    Wang, Q., Geng, T., Wang, Z., Wang, T., Fu, B., and Zheng, F. Sample then identify: A general framework for risk control and assessment in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2025 b

    2025

  19. [27]

    Ascd: Attention-steerable contrastive decoding for reducing hallucination in mllm

    Wang, Y., Bi, J., Ma, Y., and Pirk, S. Ascd: Attention-steerable contrastive decoding for reducing hallucination in mllm. arXiv preprint arXiv:2506.14766, 2025 c

    2025 arXiv

  20. [28]

    Machinery multimodal uncertainty-aware rul prediction: A stochastic modeling framework for uncertainty quantification and informed fusion

    Wang, Y., Lei, Y., Li, N., Feng, K., Wang, Z., Tan, Y., and Li, H. Machinery multimodal uncertainty-aware rul prediction: A stochastic modeling framework for uncertainty quantification and informed fusion. IEEE Internet of Things Journal, 2025 d

    2025

  21. [29]

    T., and Zhu, X

    Wang, Z., Duan, J., Cheng, L., Zhang, Y., Wang, Q., Shi, X., Xu, K., Shen, H. T., and Zhu, X. C on U : Conformal uncertainty in large language models with correctness coverage guarantees. In Findings of the Association for Computational Linguistics: EMNLP 2024, 2024

    2024

  22. [30]

    Coin: Uncertainty-guarding selective question answering for foundation models with provable risk guarantees

    Wang, Z., Duan, J., Wang, Q., Zhu, X., Chen, T., Shi, X., and Xu, K. Coin: Uncertainty-guarding selective question answering for foundation models with provable risk guarantees. arXiv preprint arXiv:2506.20178, 2025 e

    2025 arXiv

  23. [31]

    Word-sequence entropy: Towards uncertainty estimation in free-form medical question answering applications and beyond

    Wang, Z., Duan, J., Yuan, C., Chen, Q., Chen, T., Zhang, Y., Wang, R., Shi, X., and Xu, K. Word-sequence entropy: Towards uncertainty estimation in free-form medical question answering applications and beyond. Engineering Applications of Artificial Intelligence, 2025 f

    2025

  24. [32]

    SC on U : Selective conformal uncertainty in large language models

    Wang, Z., Wang, Q., Zhang, Y., Chen, T., Zhu, X., Shi, X., and Xu, K. SC on U : Selective conformal uncertainty in large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025 g

    2025

  25. [33]

    Pediatricsgpt: Large language models as chinese medical assistants for pediatric applications

    Yang, D., Wei, J., Xiao, D., Wang, S., Wu, T., Li, G., Li, M., Wang, S., Chen, J., Jiang, Y., et al. Pediatricsgpt: Large language models as chinese medical assistants for pediatric applications. Advances in Neural Information Processing Systems, 2024

    2024

  26. [34]

    Spot! revisiting video-language models for event understanding

    Zhang, G., Bi, J., Gu, J., Chen, Y., and Tresp, V. Spot! revisiting video-language models for event understanding. arXiv preprint arXiv:2311.12919, 2023

    2023 arXiv

  27. [35]

    Hademif: Hallucination detection and mitigation in large language models

    Zhou, X., Zhang, M., Lee, Z., Ye, W., and Zhang, S. Hademif: Hallucination detection and mitigation in large language models. In The Thirteenth International Conference on Learning Representations, 2025

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.