Pith. sign in

REVIEW 3 major objections 5 minor 88 references

Visual hallucination detection in large vision-language models via evidential conflict

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that an evidential-conflict score computed from a model's own decoder features detects visual hallucinations in large vision-language models more reliably than standard internal uncertainty metrics, by average AUROC gains…

desk verdict Useful benchmark, plausible heuristic, but the theory rests on a wrong reading of transformer architecture. read the letter →

arxiv 2506.19513 v1 pith:MTSMKH5H submitted 2025-06-24 cs.CV cs.LG

classification cs.CVcs.LG
keywords visualhallucinationdetectionDempster-Shafertheoryevidentialconflictlargevision-languagemodelsuncertaintyquantificationPRE-HALbenchmarkrelationreasoningfeed-forwardnetworkfeatures
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 is trying to establish that a single-pass evidential-conflict score, computed from the top-layer features and feed-forward weights of an open large vision-language model, flags visual hallucinations more reliably than standard internal uncertainty metrics. It also builds the PRE-HAL benchmark, which separates perception from reasoning across instance, scene, and relation semantics, and shows that relation reasoning is where the strongest models fail most: hallucination rates of 49.44%, 50.42%, and 19.92% on LLaVA-v1.5, mPLUG-Owl2, and mPLUG-Owl3. On that benchmark the authors report that their $\kappa_{\max}$ score raises AUROC over the best internal baseline by 4%, 10%, and 7% on the three models. If the claim holds, hallucination detection becomes a training-free, single forward pass that needs no repeated sampling, no external verifier, and no access to calibration statistics.

What carries the argument

The central object is the evidential-conflict score $\kappa_{\max}$. The paper constructs it by pairing the decoder's top-layer feature vector $\Phi$ with the feed-forward network weight matrix $B$ and bias $A$, splitting $W = B \odot \Phi^\top + A$ into positive and negative weight-of-evidence pools, converting each into simple mass functions, and combining them with Dempster's rule. Simple mass functions with the same focal set combine by adding weights, so the power-set explosion of general Dempster-Shafer combination is avoided; the closed-form conflict of Eq. (20) measures disagreement between evidence supporting token $z_i$ and evidence supporting its complement, and $\kappa_{\max}$ aggregates per-token conflicts over the whole response.

What would settle it

Recompute $\kappa$ using the actual language-model head weights instead of the feed-forward matrix: if the AUROC advantage over entropy and length baselines largely disappears or reverses, the claimed Dempster-Shafer mechanism is not what carries the result.

Watch

Extended reading notes

Core claim

Visual hallucinations leave a measurable trace in the internal evidence of an LVLM. The paper treats the final hidden feature vector $\Phi$ as evidence and the feed-forward network parameters as a set of simple mass functions; positive and negative support are separated, combined under Dempster's rule, and summarized by the conflict $\kappa$. Taking the maximum token-level conflict over the generated response yields $\kappa_{\max}$, which the authors claim is the first Dempster-Shafer-based visual hallucination detector for LVLMs and outperforms five baseline uncertainty metrics and two external methods on the new PRE-HAL benchmark, with the largest gain on mPLUG-Owl2. The same evidence also drives the benchmark result: reasoning-based hallucinations, especially relation reasoning, are far more frequent than perception-based ones, and the detector is most robust on scene-perception items.

Load-bearing premise

The method assumes that the feed-forward network's weight matrix is the classifier that turns the top-layer features into next-token logits, whereas in the evaluated models a separate language-model head produces those logits.

Editorial extensions

If this is right

  • Hallucination detection becomes a single inference pass, so it scales to deployment on open LVLMs without repeated generation or external verifier calls.
  • The detector is most valuable exactly where probability-based measures fail, since conflict does not depend on calibration; mPLUG-Owl3's ECE of 0.48 sets up that case.
  • PRE-HAL supplies a shared protocol for measuring reasoning hallucinations separately from perception hallucinations, so progress on relation reasoning becomes testable.
  • The method transfers across model families and sizes: reported AUROC stays competitive over temperatures from 0.1 to 1.5 and across 7B, 13B, and 34B models.
  • The same evidential machinery can be pointed at other FFN-based prediction errors, such as classification and semantic-segmentation mistakes, per the paper's discussion.

Reading between the lines

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

  • Because Eq. (7) identifies the FFN with the classifier even though the evaluated models use a separate LM head, the practical signal may come from generic feature statistics rather than from Dempster-Shafer semantics; comparing $\kappa$ computed with the actual LM head would isolate the mechanism.
  • The benchmark's multiple-choice format may understate or overstate reasoning hallucination rates relative to open-ended VQA; converting a subset of PRE-HAL to free-form answers would test whether the 49% and 50% relation-reasoning numbers are format-dependent.
  • The max aggregation throws away where in the sentence the conflict peaks; using per-token $\kappa$ as a localization signal could identify the specific phrase that hallucinates, which the paper does not explore.
  • The method should transfer to detecting hallucinations caused by low-resolution or adversarial images, since intra- and inter-modality uncertainty are both claimed to surface as feature conflict; a direct OOD test would check that claim.
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 / 5 minor

Summary. The paper introduces PRE-HAL, a multiple-choice visual hallucination benchmark for LVLMs that spans perception and reasoning across instance, scene, and relation semantics, and proposes a Dempster-Shafer theory (DST) based method for detecting visual hallucinations during inference. The method extracts high-level features from the last transformer layer, treats them as evidence through simple mass functions, and computes a conflict value κ that is aggregated as the maximum over generated tokens (κmax). The authors evaluate LLaVA-v1.5, mPLUG-Owl2, and mPLUG-Owl3 on PRE-HAL, reporting that κmax outperforms five internal baselines (PS, LPS, PE, LN-PE, length) with average AUROC improvements of 4%, 10%, and 7%, and also compares favorably to two external baselines (SE, SC) while requiring only a single forward pass. They additionally report high relation-reasoning hallucination rates (49.44%, 50.42%, 19.92%) and an ablation over temperature and model size.

Significance. If the claimed results are correct, the proposed κmax is a useful training-free, single-pass internal uncertainty signal for hallucination detection in open-source LVLMs, and PRE-HAL adds a reasoning-focused benchmark dimension that existing perception-only benchmarks lack. Strengths include that no parameter is fitted to hallucination labels; the conflict formula is a closed-form function of pretrained weights and features; the benchmark is assembled from external sources; and the code is openly released. The AUROC gains over internal baselines are consistent across three different LVLM families, and the method's stability across temperatures and model sizes is a positive sign. However, the central derivation rests on an architectural identification of the FFN as the final classifier that does not match the evaluated LLaMA-style decoders, and the evaluation lacks error bars or significance tests, so the magnitude and theoretical grounding of the reported gains are not yet established.

major comments (3)
  1. [Section 4.1, Eq. (7)] The derivation identifies the FFN parameter matrix B as the classifier weight matrix that produces next-token logits. In LLaMA/Vicuna-based models (LLaVA-v1.5, mPLUG-Owl2, mPLUG-Owl3), the final hidden state passes through a separate language-model head (lm_head) to produce logits; the FFN is inside the last transformer block and is not the classifier. If the implementation uses the last block's MLP weights as B, then the computed κ is not the conflict among the features that actually determine p(z | ...), and the theoretical connection to Denoeux's framework in Sections 4.2-4.3 does not hold as stated. If the implementation instead uses lm_head.weight, then the notation 'FFN parameters' in Figure 5 and Section 4.1 is misleading. The paper needs to state explicitly which weight matrix was used, justify the identification architecturally, and ideally provide an ablation comparing the two choices, since the reported AUROC gaps could be sensitive to this choice.
  2. [Tables 1 and 2, Figure 6] The central empirical claims (4%, 10%, 7% average AUROC improvements) are presented without error bars, confidence intervals, or significance tests. Since the improvements are over the best internal baseline and the datasets have thousands of samples, a bootstrap or paired test is feasible. Without such evidence, it is unclear whether the reported gains are reliable or could arise from sampling variability, especially where the margins are small (e.g., LLaVA-v1.5 total AUROC 0.59 vs. SC 0.65 in Table 2, where the external baseline actually scores higher).
  3. [Section 5.1, 'Evaluation Metrics' and Section 3.2] Hallucination labels are operationalized as a wrong choice in a multiple-choice VQA question. This conflates general predictive error with visual hallucination: a model may select the wrong option for reasons unrelated to visual content (e.g., priors, option ambiguity), and a correct answer does not guarantee the absence of hallucinated content in the response. The paper's title and claims are specifically about visual hallucination detection, so the detection target should be clearly distinguished from ordinary answer correctness, or the claims should be moderated to 'answer-error detection.'
minor comments (5)
  1. [Equation (16)] The product notation 'IM' appears to be a rendering error; it should be a product or fusion operator over i, and the index alignment in the displayed formulas should be corrected.
  2. [Section 5.3, Figure 7] The ablation on model size mixes model families (LLaVA-v1.5 for 7B/13B and LLaVA-v1.6 for 34B), so any observed trend in model size is confounded with architecture and training changes; this should be stated as a limitation.
  3. [Section 5.4] The comparison between ChatGPT-4o and human evaluators reports only that differences 'were not large enough to impact results' without providing agreement statistics or the number of volunteers; adding Cohen's kappa or a similar measure would make the claim verifiable.
  4. [Section 1, 'Statistical lower bound'] The sentence 'the occurrence of hallucinations has been proven to have a statistical lower bound' is a strong claim that should be more precisely phrased: the cited result states that calibrated language models must hallucinate under a specific formalization, not that all hallucinations are unavoidable in practice.
  5. [Table 2] The row for LLaVA-v1.5 shows SC total AUROC 0.65 while κmax is 0.59, yet the abstract reports improvement over internal baselines; the text should make clear that the 4%/10%/7% comparisons are against the best internal baseline, not against the external methods, to avoid an apparent inconsistency.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the evidential-conflict detector is a closed-form function of pretrained weights and features, and the disclosed BELIEF 2024 predecessor [55] is not load-bearing.

full rationale

The derivation chain is self-contained. The conflict score kappa is computed from the pretrained transformer weights and hidden features through the closed-form expressions in Eqs. (8)-(21), following the external Dempster-Shafer framework of Denoeux [19]; no parameter is fitted to hallucination labels, and the LCP estimates in Eqs. (10)-(12) reduce to row/column centering of the given weights rather than a learned fit. The PRE-HAL benchmark is assembled from external sources (MMBench, MMVP, POPE, R-Bench) plus manually constructed OOD data, and detection performance is evaluated against ground-truth answer correctness, not against the method's own outputs. The statement that the plausibility transformation in Eq. (22) is equivalent to the softmax in Eq. (7) is an internal identity of the DST construction, not the detection score; kappa_max is computed from feature-level positive and negative evidence pools, so it is not a mere relabeling of predictive entropy or token probability. The only self-citation is the disclosed BELIEF 2024 predecessor [55], which is presented as an extended version source and is not used to justify the new benchmark or detection results. The architectural concern about whether the relevant matrix is the FFN or the language-model head is a correctness risk rather than a circularity, and it does not raise the circularity score.

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

The method introduces no trained parameters or new entities; it reuses Denoeux's closed-form evidential estimation and depends on the assumption that transformer final-layer features behave like classifier features. The benchmark operationalizes hallucination as wrong multiple-choice answers.

assumptions (5)
  • standard math Dempster-Shafer theory and Dempster's rule of combination apply to feature-level evidence.
    Background from [15,69], used throughout Section 2.1 and Section 4.3.
  • domain assumption The final layer of an LVLM can be treated as a linear classifier with feature vector Phi and weight matrix B, as in Eq (7).
    Section 4.1; in standard transformer decoders logits come from the language-model head on hidden states, not the FFN, so this identification is not automatic.
  • domain assumption The Least Commitment Principle parameter estimates from Denoeux [19] are valid in this setting.
    Eqs (10)-(12) are adopted from [19] without re-derivation.
  • ad hoc to paper The maximum conflict over generated tokens, kappa_max, is the right sentence-level aggregation.
    Eq (21); no theoretical justification beyond empirical performance.
  • domain assumption Ground-truth answers in PRE-HAL identify visual hallucinations; an incorrect multiple-choice answer counts as a hallucination.
    Section 3.2 and 3.3; conflates all wrong answers with hallucination.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual hallucination detection in large vision-language models via evidential conflict." pith.science (2026). https://pith.science/paper/MTSMKH5H

@misc{pith2026250619513,
  author       = {Pith},
  title        = {Pith review of: Visual hallucination detection in large vision-language models via evidential conflict},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MTSMKH5H}},
  note         = {Machine review of arXiv:2506.19513}
}
read the original abstract

Despite the remarkable multimodal capabilities of Large Vision-Language Models (LVLMs), discrepancies often occur between visual inputs and textual outputs--a phenomenon we term visual hallucination. This critical reliability gap poses substantial risks in safety-critical Artificial Intelligence (AI) applications, necessitating a comprehensive evaluation benchmark and effective detection methods. Firstly, we observe that existing visual-centric hallucination benchmarks mainly assess LVLMs from a perception perspective, overlooking hallucinations arising from advanced reasoning capabilities. We develop the Perception-Reasoning Evaluation Hallucination (PRE-HAL) dataset, which enables the systematic evaluation of both perception and reasoning capabilities of LVLMs across multiple visual semantics, such as instances, scenes, and relations. Comprehensive evaluation with this new benchmark exposed more visual vulnerabilities, particularly in the more challenging task of relation reasoning. To address this issue, we propose, to the best of our knowledge, the first Dempster-Shafer theory (DST)-based visual hallucination detection method for LVLMs through uncertainty estimation. This method aims to efficiently capture the degree of conflict in high-level features at the model inference phase. Specifically, our approach employs simple mass functions to mitigate the computational complexity of evidence combination on power sets. We conduct an extensive evaluation of state-of-the-art LVLMs, LLaVA-v1.5, mPLUG-Owl2 and mPLUG-Owl3, with the new PRE-HAL benchmark. Experimental results indicate that our method outperforms five baseline uncertainty metrics, achieving average AUROC improvements of 4%, 10%, and 7% across three LVLMs. Our code is available at https://github.com/HT86159/Evidential-Conflict.

Figures

Figures reproduced from arXiv: 2506.19513 by the authors.

Figure 1
Figure 1. Examples illustrating five types of capabilities of LVLMs (top row) and their [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The components of a data entry. (MMVP) benchmark, which includes images that CLIP misinterprets, expos￾ing the limitations of LVLMs in handling visually distinct patterns. More￾over, the variation in data formats and evaluation metrics across datasets complicates an extensive assessment of hallucinations in LVLMs [54]. To fill this gap, we propose a comprehensive benchmark for LVLMs’ hallucination detection, PRE-HAL… view at source ↗
Figure 3
Figure 3. The composition and the hierarchical capability dimensions of PRE-HAL. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Hallucination rates and ECE scores on five types of hallucinations. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Scheme of the proposed method. The upper subfigure illustrates the next-token [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Visual hallucination detection rates across the five hallucination types on the [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: We compare models with 7B, 13B, and 34B parameters. As LLaVA [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 7
Figure 7. Figure 7: Ablation study of AUROC on LLaVA-v1.5 under varying temperatures and [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: AUROC scores based on the correctness results of LVLM outputs assessed [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 58 canonical work pages

  1. [1]

    https://lmsys.org/blog/2023-03-30-vicuna

    Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* Chat- GPT Quality | LMSYS Org. https://lmsys.org/blog/2023-03-30-vicuna

  2. [2]

    International Journal of Man-Machine Studies, 30(5):525–536, 1989

    A computationally efficient approximation of dempster-shafer theory. International Journal of Man-Machine Studies, 30(5):525–536, 1989

  3. [3]

    Resource bounded and anytime approximation of belief function com- putations.International Journal of Approximate Reasoning, 31(1):103– 154, 2002

  4. [4]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

  5. [5]

    Continual evidential deep learning for out-of-distribution de- tection

    Eduardo Aguilar, Bogdan Raducanu, Petia Radeva, and Joost Van de Weijer. Continual evidential deep learning for out-of-distribution de- tection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3444–3454, 2023

  6. [6]

    Flamingo: avisuallanguagemodelforfew-shot learning.Advances in neural information processing systems, 35:23716– 23736, 2022

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, MalcolmReynolds, etal. Flamingo: avisuallanguagemodelforfew-shot learning.Advances in neural information processing systems, 35:23716– 23736, 2022

  7. [7]

    Vqa: Visual ques- tion answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual ques- tion answering. InProceedings of the IEEE international conference on computer vision, pages 2425–2433, 2015

  8. [8]

    Evidential classification for defending against adver- sarial attacks on network traffic.Information Fusion, 92:115–126, 2023

    Matthew Beechey, Sangarapillai Lambotharan, and Konstantinos G Kyriakopoulos. Evidential classification for defending against adver- sarial attacks on network traffic.Information Fusion, 92:115–126, 2023. 28

Show all 88 references
  1. [9]

    Let there be a clock on the beach: Reducing object hallucination in image captioning

    Ali Furkan Biten, Lluís Gómez, and Dimosthenis Karatzas. Let there be a clock on the beach: Reducing object hallucination in image captioning. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1381–1390, 2022

  2. [10]

    Learning horn envelopes via queries from language models.International Journal of Approximate Reasoning, 171:109026, 2024

    Sophie Blum, Raoul Koudijs, Ana Ozaki, and Samia Touileb. Learning horn envelopes via queries from language models.International Journal of Approximate Reasoning, 171:109026, 2024

  3. [11]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

  4. [12]

    A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology, 15(3):1–45, 2024

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology, 15(3):1–45, 2024

  5. [13]

    Inside: Llms’ internal states retain the power of hallucination detection

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. Inside: Llms’ internal states retain the power of hallucination detection. InThe Twelfth International Conference on Learning Representations, 2024

  6. [14]

    In- structBLIP: Towards General-purpose Vision-Language Models with In- struction Tuning, June 2023

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. In- structBLIP: Towards General-purpose Vision-Language Models with In- struction Tuning, June 2023

  7. [15]

    Upper and lower probability inferences based on a sample from a finite univariate population.Biometrika, 54(3-4):515–528, 1967

    Arthur P Dempster. Upper and lower probability inferences based on a sample from a finite univariate population.Biometrika, 54(3-4):515–528, 1967

  8. [16]

    Generalized evidence theory.Applied Intelligence, 43(3):530–543, 2015

    Yong Deng. Generalized evidence theory.Applied Intelligence, 43(3):530–543, 2015

  9. [17]

    A neural network classifier based on dempster-shafer theory.IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 30(2):131–150, 2000

    Thierry Denoeux. A neural network classifier based on dempster-shafer theory.IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 30(2):131–150, 2000. 29

  10. [18]

    Decision-making with belief functions: A review

    Thierry Denoeux. Decision-making with belief functions: A review. International Journal of Approximate Reasoning, 109:87–110, 2019

  11. [19]

    Logistic regression, neural networks and dempster– shafer theory: A new perspective.Knowledge-Based Systems, 176:54–67, 2019

    Thierry Denœux. Logistic regression, neural networks and dempster– shafer theory: A new perspective.Knowledge-Based Systems, 176:54–67, 2019

  12. [20]

    Bert: Pre-trainingofdeepbidirectionaltransformersforlanguageunder- standing

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-trainingofdeepbidirectionaltransformersforlanguageunder- standing. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language te...

  13. [21]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition ...

  14. [22]

    A Survey of Vision-Language Pre-Trained Models, July 2022

    Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A Survey of Vision-Language Pre-Trained Models, July 2022

  15. [23]

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

    Jinhao Duan, Hao Cheng, Shiqi Wang, Alex Zavalny, Chenan Wang, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models. InProceedings of the 62nd Annual Meeting of the Ass...

  16. [24]

    De- tecting hallucinations in large language models using semantic entropy

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

  17. [25]

    Make me a bnn: A simple strategy for estimating bayesian uncertainty from pre-trained models

    Gianni Franchi, Olivier Laurent, Maxence Leguéry, Andrei Bursuc, An- drea Pilzer, and Angela Yao. Make me a bnn: A simple strategy for estimating bayesian uncertainty from pre-trained models. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

  18. [26]

    Dropout as a bayesian approxima- tion: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approxima- tion: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016

  19. [27]

    Llama- adapter v2: Parameter-efficient visual instruction model.arXiv preprint arXiv:2304.15010, 2023

    Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama- adapter v2: Parameter-efficient visual instruction model.arXiv preprint arXiv:2304.15010, 2023

  20. [28]

    Hallusionbench: an advanced diagnostic suite for entangled lan- guage hallucination and visual illusion in large vision-language models

    Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xi- aoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled lan- guage hallucination and visual illusion in large vision-language models. InPr...

  21. [29]

    Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation

    Nuno M Guerreiro, Elena Voita, and André FT Martins. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation. InProceedings of the 17th Conference of the Eu- ropean Chapter of the Association for Computational Linguistics, page...

  22. [30]

    On calibra- tion of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibra- tion of modern neural networks. InInternational conference on machine learning, pages 1321–1330. PMLR, 2017

  23. [31]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforce- ment learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforce- ment learning.arXiv preprint arXiv:2501.12948, 2025

  24. [32]

    Divert more attention to vision-language object tracking.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024

    Mingzhe Guo, Zhipeng Zhang, Liping Jing, Haibin Ling, and Heng Fan. Divert more attention to vision-language object tracking.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024

  25. [33]

    Ciem: Con- trastive instruction evaluation method for better instruction tuning

    Hongyu Hu, Jiyuan Zhang, Minyi Zhao, and Zhenbang Sun. Ciem: Con- trastive instruction evaluation method for better instruction tuning. In NeurIPS 2023 Workshop on Instruction Tuning and Instruction Follow- ing. 31

  26. [34]

    Automated trading systems statistical and machine learning methods and hardware implementation: a survey.Enterprise Information Sys- tems, 13(1):132–144, 2019

    Boming Huang, Yuxiang Huan, Li Da Xu, Lirong Zheng, and Zhuo Zou. Automated trading systems statistical and machine learning methods and hardware implementation: a survey.Enterprise Information Sys- tems, 13(1):132–144, 2019

  27. [35]

    Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development.Advances in neural information processing systems, 2021

    Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Connor Coley, Cao Xiao, Jimeng Sun, and Marinka Zit- nik. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development.Advances in neural information processing sy...

  28. [36]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232, 2023

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232, 2023

  29. [37]

    Lymphoma segmentation from 3D PET-CT images using a deep evidential network

    LingHuang, SuRuan, PierreDecazes, andThierryDenœux. Lymphoma segmentation from 3D PET-CT images using a deep evidential network. International Journal of Approximate Reasoning, 149:39–60, October 2022

  30. [38]

    Surveyof hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38, 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, EtsukoIshii, YeJinBang, AndreaMadotto, andPascaleFung. Surveyof hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38, 2023

  31. [39]

    Language models (mostly) know what they know.CoRR, 2022

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova Das- Sarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know.CoRR, 2022

  32. [40]

    Calibrated language mod- elsmusthallucinate

    Adam Tauman Kalai and Santosh S Vempala. Calibrated language mod- elsmusthallucinate. InProceedings of the 56th Annual ACM Symposium on Theory of Computing, pages 160–171, 2024

  33. [41]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Ben- jamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020. 32

  34. [42]

    Large language models must be taught to know what they don’t know.arXiv preprint arXiv:2406.08391, 2024

    Sanyam Kapoor, Nate Gruver, Manley Roberts, Katherine Collins, Arka Pal, Umang Bhatt, Adrian Weller, Samuel Dooley, Micah Goldblum, and Andrew Gordon Wilson. Large language models must be taught to know what they don’t know.arXiv preprint arXiv:2406.08391, 2024

  35. [43]

    Deep visual-semantic alignments for generating image descriptions

    Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3128–3137, 2015

  36. [44]

    Semantic uncer- tainty: Linguistic invariances for uncertainty estimation in natural lan- guagegeneration

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncer- tainty: Linguistic invariances for uncertainty estimation in natural lan- guagegeneration. InThe Eleventh International Conference on Learning Representations, 2023

  37. [45]

    Simple and scalable predictive uncertainty estimation using deep en- sembles.Advances in neural information processing systems, 30, 2017

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep en- sembles.Advances in neural information processing systems, 30, 2017

  38. [46]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 292–305, 2023

  39. [47]

    Teaching models to express their uncertainty in words.Transactions on Machine Learning Research, 2022

    Stephanie Lin, Jacob Hilton, and Owain Evans. Teaching models to express their uncertainty in words.Transactions on Machine Learning Research, 2022

  40. [48]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Per- ona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

  41. [49]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Per- ona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors,Computer Vision – ECCV 2014, pag...

  42. [50]

    Generating with confi- dence: Uncertainty quantification for black-box large language models

    Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. Generating with confi- dence: Uncertainty quantification for black-box large language models. Transactions on Machine Learning Research, 2023

  43. [51]

    A sur- vey on hallucination in large vision-language models.arXiv preprint arXiv:2402.00253, 2024

    Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A sur- vey on hallucination in large vision-language models.arXiv preprint arXiv:2402.00253, 2024

  44. [52]

    Im- proved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Im- proved baselines with visual instruction tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024

  45. [53]

    Visual instruction tuning.Advances in neural information processing systems, 36, 2024

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36, 2024

  46. [54]

    Mmbench: Is your multi-modal model an all-around player? InEuro- pean conference on computer vision, pages 216–233

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? InEuro- pean conference on computer vision, pages 216–233. Springer, 2024

  47. [55]

    Object halluci- nation detection in large vision language models via evidential conflict

    Zhekun Liu, Tao Huang, Rui Wang, and Liping Jing. Object halluci- nation detection in large vision language models via evidential conflict. InInternational Conference on Belief Functions, pages 58–67. Springer, 2024

  48. [56]

    Negative object presence evaluation (NOPE) to measure object hallucination in vision-language models

    Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (NOPE) to measure object hallucination in vision-language models. In Jing Gu, Tsu-Jui (Ray) Fu, Drew Hudson, Asli Celikyilmaz, and William Wang, editors,Proceed- ing...

  49. [57]

    AI Halluci- nations: A Misnomer Worth Clarifying

    Negar Maleki, Balaji Padmanabhan, and Kaushik Dutta. AI Halluci- nations: A Misnomer Worth Clarifying. In2024 IEEE Conference on Artificial Intelligence (CAI), pages 133–138, June 2024. 34

  50. [58]

    Uncertainty estimation in autoregressive structured prediction

    Andrey Malinin and Mark John Francis Gales. Uncertainty estimation in autoregressive structured prediction. InInternational Conference on Learning Representations, 2021

  51. [59]

    Selfcheckgpt: Zero- resource black-box hallucination detection for generative large language models

    Potsawee Manakul, Adian Liusie, and Mark Gales. Selfcheckgpt: Zero- resource black-box hallucination detection for generative large language models. InThe 2023 Conference on Empirical Methods in Natural Lan- guage Processing, 2023

  52. [60]

    On faithfulness and factuality in abstractive summarization

    Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. On faithfulness and factuality in abstractive summarization. InProceed- ings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906–1919, 2020

  53. [61]

    Concise thoughts: Impact of output length on llm reasoning and cost.CoRR, 2024

    Sania Nayab, Giulio Rossolini, Giorgio C Buttazzo, Nicolamaria Manes, and Fabrizio Giacomelli. Concise thoughts: Impact of output length on llm reasoning and cost.CoRR, 2024

  54. [62]

    Gpt-4 technical report, 2024

    OpenAI. Gpt-4 technical report, 2024

  55. [63]

    Dinov2: Learning robust visual fea- tures without supervision.Transactions on Machine Learning Research, 2024

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual fea- tures without supervision.Transactions on Machine Learning Research, 2024

  56. [64]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

  57. [65]

    Semantic consistency for assuring reliability of large language models

    Harsh Raj, Vipul Gupta, Domenic Rosati, and Subhabrata Majumdar. Semantic consistency for assuring reliability of large language models. arXiv preprint arXiv:2308.09138, 2023

  58. [66]

    The Troubling Emergence of Hallucination in Large Language 35 Models - An Extensive Definition, Quantification, and Prescriptive Re- mediations

    Vipula Rawte, Swagata Chakraborty, Agnibh Pathak, Anubhav Sarkar, S.M Towhidul Islam Tonmoy, Aman Chadha, Amit Sheth, and Amitava Das. The Troubling Emergence of Hallucination in Large Language 35 Models - An Extensive Definition, Quantification, and Prescriptive Re- mediation...

  59. [67]

    Object hallucination in image captioning

    Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. InProceed- ings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035–4045, 2018

  60. [68]

    Evidential deep learning to quantify classification uncertainty.Advances in neural infor- mation processing systems, 31, 2018

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty.Advances in neural infor- mation processing systems, 31, 2018

  61. [69]

    Princeton university press, 1976

    Glenn Shafer.A Mathematical Theory of Evidence. Princeton university press, 1976

  62. [70]

    Talking about large language models.Communica- tions of the ACM, 67(2):68–79, 2024

    Murray Shanahan. Talking about large language models.Communica- tions of the ACM, 67(2):68–79, 2024

  63. [71]

    Belief functions: the disjunctive rule of combination and the generalized bayesian theorem.International Journal of approximate reasoning, 9(1):1–35, 1993

    Philippe Smets. Belief functions: the disjunctive rule of combination and the generalized bayesian theorem.International Journal of approximate reasoning, 9(1):1–35, 1993

  64. [72]

    Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. InProceedings of...

  65. [73]

    Eyes wide shut? exploring the visual shortcomings of multimodal llms

    Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568–9578, 2024

  66. [74]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023. 36

  67. [75]

    Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Alma- hairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288, 2023

  68. [76]

    Attention is all you need.Advances in Neural Information Processing Systems, 2017

    A Vaswani. Attention is all you need.Advances in Neural Information Processing Systems, 2017

  69. [77]

    Rui Wang, Tengkun Yang, Ci Liang, Mengying Wang, and Yusheng Ci. Reliable autonomous driving environment perception: Uncertainty quantification of semantic segmentation.Journal of Transportation En- gineering, Part A: Systems, 151(3):04024117, 2025

  70. [78]

    Self-consistency improveschainofthoughtreasoninginlanguagemodels

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sha- ran Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improveschainofthoughtreasoninginlanguagemodels. InThe Eleventh International Conference on Learning Representations

  71. [79]

    Evaluating and analyzing relationship halluci- nations in large vision-language models

    Mingrui Wu, Jiayi Ji, Oucheng Huang, Jiale Li, Yuhang Wu, Xiaoshuai Sun, and Rongrong Ji. Evaluating and analyzing relationship halluci- nations in large vision-language models. InProceedings of the 41st In- ternational Conference on Machine Learning, volume 235, pages 53553– ...

  72. [80]

    On hallucination and predictive uncertainty in conditional language generation

    Yijun Xiao and William Yang Wang. On hallucination and predictive uncertainty in conditional language generation. InProceedings of the 16th Conference of the European Chapter of the Association for Com- putational Linguistics: Main Volume, 2021

  73. [81]

    Uncertainty quantifica- tion with pre-trained language models: A large-scale empirical analysis

    Yuxin Xiao, Paul Pu Liang, Umang Bhatt, Willie Neiswanger, Ruslan Salakhutdinov, and Louis-Philippe Morency. Uncertainty quantifica- tion with pre-trained language models: A large-scale empirical analysis. InFindings of the Association for Computational Linguistics: EMNLP 2022...

  74. [82]

    Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. InThe Twelfth International Conference on Learning Representations, 2024. 37

  75. [83]

    Deep evidential fusion network for medical image classification.International Journal of Approximate Reasoning, 150:188–198, 2022

    Shaoxun Xu, Yufei Chen, Chao Ma, and Xiaodong Yue. Deep evidential fusion network for medical image classification.International Journal of Approximate Reasoning, 150:188–198, 2022

  76. [84]

    mplug-owl3: Towards long image-sequence understanding in multi-modal large language models

    Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024

  77. [85]

    mplug-owl2: Revolutionizing multi- modal large language model with modality collaboration

    Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi- modal large language model with modality collaboration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition,...

  78. [86]

    R-tuning: Instructing large language models to say ‘i don’t know’

    Hanning Zhang, Shizhe Diao, Yong Lin, Yi Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. R-tuning: Instructing large language models to say ‘i don’t know’. InProceedings of the 2024 Conference of the North American Chapter of the Association for Com- putat...

  79. [87]

    A survey of controllable text generation using transformer-based pre- trained language models.ACM Computing Surveys, 56(3):1–37, 2023

    Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. A survey of controllable text generation using transformer-based pre- trained language models.ACM Computing Surveys, 56(3):1–37, 2023

  80. [88]

    Siren’s song in the ai ocean: a survey on hallucination in large language models

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. Siren’s song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 38

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.