Pith. sign in

REVIEW 5 major objections 8 minor 31 references

The paper claims that hallucinations in medical visual question answering can be flagged deterministically by a single fixed score derived from the model's own token-level log-probabilities—no stochastic sampling, no external natural langua

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A fixed, sampling-free score — per-token log-probability variance times (1 + average |image-vs-text probability shift|) — detects medical-VQA hallucinations better than semantic-entropy baselines in 13 of 16 settings.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely cheap deterministic detector with a plausible but label-dependent empirical win; the mechanism story is internally inconsistent and the GREEN labels are unvalidated. the 5 major comments →

arxiv 2603.21693 v2 pith:GB32AITO submitted 2026-03-23 cs.AI

Deterministic Hallucination Detection in Medical VQA via Confidence-Evidence Bayesian Gain

Mohammad Asadi , Tahoura Nedaee , Jack W. O'Sullivan , Euan Ashley , Ehsan Adeli This is my paper
classification cs.AI
keywords hallucination detectionmedical visual question answeringuncertainty estimationdeterministicpointwise mutual informationtoken-level predictive variancelog-probability analysismultimodal large language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper claims that hallucinations in medical visual question answering can be detected deterministically from the model's own token-level log-probabilities, without stochastic sampling, external natural-language-inference models, or task-specific hyperparameters. Its central claim is that a single fixed score—the standard deviation of per-token log-probabilities multiplied by one plus the normalized absolute log-probability shift caused by the image—separates responses that contradict the image from grounded ones. Across four medical vision-language models and three VQA benchmarks, the score achieves the best AUC in 13 of 16 settings and improves over the previous sampling-based state of the art by roughly 8 AUC points on average. If correct, this means the hallucination signal is already present in one greedy generation and can be read out in three forward passes, making real-time clinical screening practical. The paper's evidence is framed around agreement with an automated reference-based grader's labels, not direct human judgment.

Core claim

CEBaG is the claim that a hallucinated medical VQA response leaves a distinctive trace in the target model's own output distribution: per-token confidence is inconsistent (high standard deviation σ over token log-probabilities) and the image barely moves those probabilities. The method computes the evidence gain G as the difference between the response's log-probability with the image and without it, interprets it as a pointwise mutual information between response and image, and normalizes by length to get E=|G|/L. The final score σ·(1+E) combines uncertainty and visual sensitivity without any tuning. On the paper's 16 model–benchmark settings, this score beats the previous state of the art

What carries the argument

The central object is the CEBaG score, σ·(1+|G|/L), where σ is the standard deviation of the response's per-token log-probabilities in the multimodal pass, G is the sequence-level evidence gain log P(r|x_image,x_q) − log P(r|x_q), and L is the response length. G is shown to equal the pointwise mutual information between response and image conditioned on the question, so |G|/L measures mean per-token visual influence; σ measures confidence inconsistency. The multiplicative coupling (1+E) amplifies σ exactly when the image shifts predictions strongly without resolving uncertainty, which the paper identifies as the signature of a struggling, potentially hallucinating model.

Load-bearing premise

The load-bearing premise is that the automated reference-based grader's labels mark genuine image contradiction and that the text-only scoring pass isolates visual evidence cleanly rather than introducing a length or positional artifact.

What would settle it

Collect clinician judgments of whether each response contradicts the image on the pathology and bilingual benchmark subsets, and compare the CEBaG score's AUC against those human labels; a collapse to chance, or a correlation explained by response length alone, would refute the claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One greedy generation plus two scoring passes replaces 10–20 stochastic generations and an external NLI model, cutting the compute per question by about an order of magnitude.
  • The score requires no hyperparameter tuning; the same fixed formula applies across the encoder–decoder and decoder-only architectures tested.
  • Detection is deterministic, so reported AUCs are exact for a given model's greedy answer rather than averages over stochastic sampling runs.
  • Because the evidence gain is computed as a pointwise mutual information, the same framework can be used to flag which tokens are weakly grounded, not just whole responses.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The evidence-gain quantity G is formally a pointwise mutual information; a natural extension is to inspect per-token G values to localize the ungrounded span of a response, turning detection into a token-level correction signal.
  • For decoder-only architectures, the text-only pass removes the image token from the prompt, which alters sequence length and positional encodings; part of the measured evidence gain may therefore be structural rather than purely visual—a testable artifact.
  • The evaluation labels hallucinations by an automated reference-based grader's threshold; a stronger test would be against human clinician judgments of image contradiction, especially on pathology and bilingual data where that grader was not designed to operate.
  • The white-box dependence on log-probabilities is an acknowledged limitation; perturbing outputs to estimate variance is a plausible black-box adaptation, though the evidence-gain term would need a proxy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The paper proposes CEBaG, a deterministic hallucination detector for medical VQA. The method computes two signals from a single greedy generation and two teacher-forced scoring passes: token-level predictive variance σ (standard deviation of per-token log-probabilities) and evidence magnitude E = |G|/L, where G is the difference between sequence log-probabilities with and without the image. The final score is σ·(1+E). The authors claim this parameter-free score outperforms Semantic Entropy (SE), Vision-Amplified Semantic Entropy (VASE), and RadFlag, achieving the best AUC in 13 of 16 settings across four medical MLLMs and three VQA datasets, with an average AUC improvement of +8.2 over VASE. Ablations show that σ alone contributes the bulk of the signal, with E providing a small additional gain.

Significance. If the empirical claims were fully substantiated, this would be a practically valuable contribution: hallucination detection without stochastic sampling, external NLI models, or tunable hyperparameters would be much cheaper and easier to deploy. The central finding that simple token-level log-probability variance outperforms semantic entropy and vision-amplified entropy on these benchmarks is surprising and worth investigating further. The method is transparent, requires only three forward passes, and the score is fully deterministic, which are genuine strengths. However, the evaluation's reliance on an unvalidated automatic labeler (GREEN), the absence of any uncertainty quantification, and the confounding of the text-only baseline in LLaVA-style models currently prevent the strong conclusions from being accepted as stated.

major comments (5)
  1. [Section 3, 'Ground truth and metrics'] The ground-truth labels are generated entirely by the GREEN model: a response is labeled hallucinated iff its GREEN score is below 1.0. GREEN was designed for radiology-report evaluation; the paper provides no evidence that its judgments are valid for pathology images (PathVQA) or for the bilingual SLAKE dataset. Every AUC/AUG in Table 1, including the 13-of-16 headline and the +8.2 average improvement over VASE, is a measure of agreement with GREEN, not with human clinical judgment. If GREEN's errors correlate with surface properties (e.g., length, phrasing, token confidence), they could systematically favor CEBaG's signals. The authors should validate GREEN against human annotations on a subset of these datasets, or at minimum report agreement statistics and error analysis. This is load-bearing for the central claim.
  2. [Table 1 and §3.1] Only point estimates of AUC/AUG are reported, with no confidence intervals, bootstrapping, or significance tests. Several settings are near-ties where the claimed superiority is not established: MedGemma-4b VQA-RAD Open (CEBaG 61.0 vs SE 61.9), MedGemma-1.5 SLAKE (CEBaG 57.5 vs VASE 59.5), and LLaVA-Med VQA-RAD Open (CEBaG 65.8 vs VASE 68.4). Since CEBaG is deterministic only for a fixed generated answer, the randomness in model generation still induces variability in the downstream detection score. The authors should provide bootstrap CIs or paired tests to show that differences are not explainable by sampling noise.
  3. [Section 3, Implementation details] For LLaVA-style models (LLaVA-Med, HuatuoGPT), the text-only pass removes the <image> token from the prompt. This changes the token sequence length and positional encodings for all subsequent text tokens. Consequently, G in Eq. (3) captures not only the absence of visual information but also a structural positional shift. Part of the measured 'evidence magnitude' may therefore be an artifact of the text-only construction rather than visual evidence. A cleaner control would keep positions fixed (e.g., replace the image embedding with a zero vector or a learned blank patch) or otherwise verify that the effect is driven by visual content.
  4. [Section 2.2, Eq. (6)] The functional form σ·(1+E) is introduced with intuitive motivation but is not derived. The paper calls the method 'hyperparameter-free,' yet the choice of the standard deviation, the absolute value, and the multiplicative combination are design decisions that are not theoretically justified. The ablation (Table 3) shows that a tuned variant CEBaGλ achieves 71.9% average AUC versus 67.9% for the fixed formula, indicating that the fixed form leaves about 4 AUC points on the table. The authors should either provide a principled derivation or frame the formula more modestly as an empirical design; the claim 'no task-specific hyperparameters' is technically true only after committing to this arbitrary form.
  5. [Section 3.1, threshold stability] The manuscript states: 'Furthermore, our method maintained stable detection performance across GREEN thresholds between 0.4 and 0.8, confirming its insensitivity to specific ground-truth boundaries.' No table, figure, or data supporting this claim is provided. This is an empirical assertion that should either be shown or removed. It is directly relevant to the concern that the choice of the 1.0 threshold may be arbitrary.
minor comments (8)
  1. [Section 2.1, Eq. (4)] The statement 'G is directly proportional to log P(x_v | r, x_q)' is imprecise. In Eq. (4), G = log P(x_v | r, x_q) - log P(x_v | x_q), so G is equal to that quantity plus a constant with respect to r. 'Proportional' should be 'equal up to an additive constant' or 'monotonically related.'
  2. [Title and throughout] The method is called 'Bayesian Gain,' but the paper explicitly disclaims Bayesian inference over parameters. The quantity G is a pointwise mutual information. Consider using a less misleading name in the title and text, or adding a remark explaining why the Bayesian terminology is appropriate.
  3. [Abstract vs. GitHub link] The abstract says 'The code will be made available upon acceptance,' while the abstract also contains the GitHub URL https://github.com/masadi-99/CEBaG. Please clarify whether the code is currently available.
  4. [Table 3] The table shows only 8 of 16 settings in detail, then reports averages over all 16. The selection criterion for the displayed rows should be stated (e.g., representative diversity across models and datasets).
  5. [Section 3, Baselines] RadFlag is described as checking consistency with radiology reports. For SLAKE and PathVQA, it is unclear what report is used. Please specify how RadFlag is adapted to datasets without a structured radiology report.
  6. [Figure 1] The caption states that the grounded answer has G=38.5 and the hallucinated answer G=1.6. It would be helpful to state whether these are actual examples from the evaluation and how the y-axis is normalized.
  7. [References] Reference [24] (Saab et al.) appears in the bibliography but is not cited in the text. Check the reference list for consistency.
  8. [Section 3, Implementation details] The paper says greedy decoding with temperature T=0.1. Greedy decoding typically implies T=0; if T=0.1 is used, it is not strictly greedy. Clarify the decoding scheme.

Circularity Check

0 steps flagged

No significant circularity: CEBaG's score is a fixed, parameter-free formula; evaluation is against external GREEN labels, not its own construction.

full rationale

CEBaG's derivation chain is self-contained and contains no circular reduction. Equations (1)-(4) define the evidence gain G as log P(r|x_v,x_q) - log P(r|x_q), and the Bayes-rule rewrite in Eq. (4) is standard pointwise-mutual-information algebra; G is not defined in terms of the hallucination label. E=|G|/L is a length-normalized monotone function of G, and the final score sigma*(1+E) in Eq. (6) is a fixed, hyperparameter-free heuristic. No parameter is fitted to the GREEN labels and then renamed a prediction; the reported AUCs are empirical comparisons against the external GREEN-based labels adopted from VASE [18]. There are no self-citations or imported uniqueness theorems. The unvalidated transfer of GREEN to PathVQA/SLAKE and the unsupported threshold-stability claim are validity/reporting concerns, not circularity. Likewise, the sign of E in Eq. (6) is in tension with the stated 'weak sensitivity' observation, but that is a correctness/sign issue, not an Eq-X=Eq-Y reduction. The paper's own stated limitation (white-box log-probability access) also does not create a circular step.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

CEBaG introduces no new latent variable, model, mediator, or mechanism — it is a deterministic function of the target model's own log-probabilities, so the invented-entities ledger is empty. What the central claim rests on instead is a set of domain assumptions: labeler trustworthiness (GREEN), the Bayes-consistent interpretation of two separate forward passes, the empirical reality of the variance signature, and the cleanliness of the text-only pass. The only 'free' design element is the hand-chosen functional form and sign of the combination, which appears calibrated to the evaluation data.

free parameters (3)
  • Functional form and sign of the CEBaG score (σ·(1+E), positive sign on E) = fixed formula; sign +E chosen to align with labels
    The abstract and Fig. 1 say hallucinations have weak visual sensitivity (low E), which would predict E suppressing the score; yet Eq. 6 multiplies by (1+E) and E-only achieves 59% AUC in the high-E-is-hallucinated direction. The form and sign were evidently selected with knowledge of these benchmarks; no derivation fixes them.
  • Decoding temperature for answer generation = T = 0.1
    The generated response r is produced at T=0.1 (§3, Implementation details). A fixed config choice, harmless, but the detector's inputs depend on it.
  • GREEN label threshold = GREEN score < 1.0 ⇒ hallucinated
    Inherited from VASE's protocol; determines every ground-truth label. The stability claim (0.4–0.8) is stated but not shown, and the gap between 0.8 and 1.0 is unexplained.
axioms (5)
  • domain assumption Bayes-consistency: the model's with-image and without-image sequence distributions are treated as conditionals of one coherent joint distribution, so G = logP(x_v|r,x_q) − logP(x_v|x_q) (Eq. 4).
    For a real transformer the two conditioning contexts are separate forward passes and need not be consistent marginals; the PMI interpretation of G is heuristic, though the score G itself is just a defined log-prob difference.
  • domain assumption GREEN scores below 1.0 correctly identify hallucinations on VQA-RAD, SLAKE, and PathVQA.
    GREEN was designed for radiology report evaluation (§3, Ground truth and metrics); no validation of its labels is provided for pathology (PathVQA) or bilingual (SLAKE) data. All reported AUCs measure agreement with GREEN labels.
  • domain assumption Hallucinated responses exhibit high per-token log-probability variance ('confidence spikes' on function words).
    The central observation motivating σ is asserted in §2.2 and supported only indirectly by the σ-only ablation (67% average AUC); no distributional statistics (e.g., σ histograms by class) are shown.
  • domain assumption Removing the <image> token yields a clean text-only baseline for LLaVA-style models.
    For LLaVA-Med and HuatuoGPT, token removal changes sequence length and positional encodings; part of G may reflect structural template shifts rather than visual evidence (§3, Implementation details).
  • standard math Standard chain rule of probability and arithmetic (Eqs. 1–3).
    Unobjectionable background manipulation.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Deterministic Hallucination Detection in Medical VQA via Confidence-Evidence Bayesian Gain." pith.science (2026). https://pith.science/paper/GB32AITO

@misc{pith2026260321693,
  author       = {Pith},
  title        = {Pith review of: Deterministic Hallucination Detection in Medical VQA via Confidence-Evidence Bayesian Gain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GB32AITO}},
  note         = {Machine review of arXiv:2603.21693}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multimodal large language models (MLLMs) have shown strong potential for medical Visual Question Answering (VQA), yet they remain prone to hallucinations, defined as generating responses that contradict the input image, posing serious risks in clinical settings. Current hallucination detection methods, such as Semantic Entropy (SE) and Vision-Amplified Semantic Entropy (VASE), require 10 to 20 stochastic generations per sample together with an external natural language inference model for semantic clustering, making them computationally expensive and difficult to deploy in practice. We observe that hallucinated responses exhibit a distinctive signature directly in the model's own log-probabilities: inconsistent token-level confidence and weak sensitivity to visual evidence. Based on this observation, we propose Confidence-Evidence Bayesian Gain (CEBaG), a deterministic hallucination detection method that requires no stochastic sampling, no external models, and no task-specific hyperparameters. CEBaG combines two complementary signals: token-level predictive variance, which captures inconsistent confidence across response tokens, and evidence magnitude, which measures how much the image shifts per-token predictions relative to text-only inference. Evaluated across four medical MLLMs and three VQA benchmarks (16 experimental settings), CEBaG achieves the highest AUC in 13 of 16 settings and improves over VASE by 8 AUC points on average, while being fully deterministic and self-contained. https://github.com/masadi-99/CEBaG

Figures

Figures reproduced from arXiv: 2603.21693 by Ehsan Adeli, Euan Ashley, Jack W. O'Sullivan, Mohammad Asadi, Tahoura Nedaee.

Figure 1
Figure 1. Figure 1: Overview of CEBaG for hallucination detection in medical VQA. (a) The CE￾BaG pipeline on a concrete example: the model generates answer r, which is scored under two conditions, with and without the image, yielding two signals: token-level predictive variance σ and evidence magnitude E=|G|/L. CEBaG combines these as σ · (1+E). (b) Per-token evidence gain for a grounded answer (top; G=38.5, image strongly in… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 12 linked inside Pith

  1. [1]

    In: ICLR (2024)

    Chen, C., Liu, K., Chen, Z., Gu, Y., Wu, Y., Tao, M., Fu, Z., Ye, J.: INSIDE: LLMs’ internal states retain the power of hallucination detection. In: ICLR (2024)

    2024

  2. [2]

    arXiv:2406.19280 (2024)

    Chen, J., Guo, D., Li, M., Lin, R., Shan, C., Zhao, Y., Shen, T., Wan, X.: HuatuoGPT-Vision, towards injecting medical visual knowledge into multimodal LLMs at scale. arXiv:2406.19280 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    In: EMNLP

    Cohen, R., Hamri, M., Geva, M., Globerson, A.: LM vs LM: Detecting factual errors via cross examination. In: EMNLP. pp. 12621–12640 (2023)

    2023

  4. [4]

    Nature630(8017), 625–630 (2024)

    Farquhar, S., Kossen, J., Kuhn, L., Gal, Y.: Detecting hallucinations in large lan- guage models using semantic entropy. Nature630(8017), 625–630 (2024)

    2024

  5. [5]

    In: CVPR

    Favero, A., Zancato, L., Trager, M., Choudhary, S., Perera, P., Achille, A., Swami- nathan, A., Soatto, S.: Multi-modal hallucination control by visual information grounding. In: CVPR. pp. 14303–14312 (2024)

    2024

  6. [6]

    In: AAAI

    Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. In: AAAI. vol. 38, pp. 18135–18143 (2024)

    2024

  7. [7]

    In: ICLR (2020)

    He, P., Liu, X., Gao, J., Chen, W.: DeBERTa: Decoding-enhanced BERT with disentangled attention. In: ICLR (2020)

    2020

  8. [8]

    arXiv:2003.10286 (2020)

    He, X., Zhang, Y., Mou, L., Xing, E., Xie, P.: PathVQA: 30000+ questions for medical visual question answering. arXiv:2003.10286 (2020)

    Pith/arXiv arXiv 2003

  9. [9]

    ACM Trans

    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 Trans. Inf. Syst. (2023)

    2023

  10. [10]

    In: CVPR

    Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. In: CVPR. pp. 27036–27046 (2024)

    2024

  11. [11]

    arXiv:2207.05221 (2022)

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., et al.: Language models (mostly) know what they know. arXiv:2207.05221 (2022)

    Pith/arXiv arXiv 2022

  12. [12]

    arXiv:2406.15927 (2024)

    Kossen, J., Han, J., Razzak, M., Schut, L., Malik, S., Gal, Y.: Semantic en- tropy probes: Robust and cheap hallucination detection in LLMs. arXiv:2406.15927 (2024)

    Pith/arXiv arXiv 2024

  13. [13]

    In: CVPR (2024) 10 M

    Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In: CVPR (2024) 10 M. Asadi et al

    2024

  14. [14]

    In: ACL Findings

    Manevich, A., Tsarfaty, R.: Mitigating hallucinations in large vision-language mod- els via language-contrastive decoding. In: ACL Findings. pp. 6008–6022 (2024)

    2024

  15. [15]

    Scientific Data 5(1), 1–10 (2018)

    Lau, J.J., Gayen, S., Ben Abacha, A., Demner-Fushman, D.: A dataset of clinically generated visual questions and answers about radiology images. Scientific Data 5(1), 1–10 (2018)

    2018

  16. [16]

    NeurIPS36(2024)

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: LLaVA-Med: Training a large language-and-vision assistant for biomedicine in one day. NeurIPS36(2024)

    2024

  17. [17]

    In: EMNLP

    Li, Q., Geng, J., Lyu, C., Zhu, D., Panov, M., Karray, F.: Reference-free hallu- cination detection for large vision-language models. In: EMNLP. pp. 4542–4551 (2024)

    2024

  18. [18]

    In: MICCAI (2025)

    Liao, Z., Hu, S., Zou, K., Fu, H., Zhen, L., Xia, Y.: Vision-amplified semantic en- tropy for hallucination detection in medical visual question answering. In: MICCAI (2025)

    2025

  19. [19]

    In: ISBI

    Liu, B., Zhan, L.M., Xu, L., Ma, L., Yang, Y., Wu, X.M.: SLAKE: A semantically- labeled knowledge-enhanced dataset for medical visual question answering. In: ISBI. pp. 1650–1654 (2021)

    2021

  20. [20]

    arXiv:2411.16778 (2024)

    Liu, B., Zou, K., Zhan, L., Lu, Z., Dong, X., Chen, Y., Xie, C., Cao, J., Wu, X.M., Fu, H.: GEMeX: A large-scale, groundable, and explainable medical VQA benchmark for chest X-ray diagnosis. arXiv:2411.16778 (2024)

    Pith/arXiv arXiv 2024

  21. [21]

    arXiv:2402.00253 (2024)

    Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv:2402.00253 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    In: EMNLP

    Min,S.,Krishna,K.,Lyu,X.,Lewis,M.,Yih,W.t.,Koh,P.,Iyyer,M.,Zettlemoyer, L., Hajishirzi, H.: FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In: EMNLP. pp. 12076–12100 (2023)

    2023

  23. [23]

    arXiv:2405.03595 (2024)

    Ostmeier, S., Xu, J., Chen, Z., Varma, M., Blankemeier, L., Bluethgen, C., Michal- son, A.E., Moseley, M., Langlotz, C., Chaudhari, A.S., et al.: GREEN: Generative radiology report evaluation and error notation. arXiv:2405.03595 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    arXiv:2404.18416 (2024)

    Saab, K., Tu, T., Weng, W.H., Tanno, R., Stutz, D., Wulczyn, E., Zhang, F., Strother, T., Park, C., et al.: Capabilities of Gemini models in medicine. arXiv:2404.18416 (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    arXiv:2507.05201 (2025)

    Sellergren, A., Kazemzadeh, S., Jaroensri, T., Kiraly, A., Traverse, M., Kohlberger, T., Xu, S., Jamil, F., Hughes, C., Lau, C., et al.: MedGemma technical report. arXiv:2507.05201 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    arXiv:2411.00299 (2024)

    Sambara, S., Zhang, S., Banerjee, O., Acosta, J., Fahrner, J., Rajpurkar, P.: Rad- Flag: A black-box hallucination detection method for medical vision language mod- els. arXiv:2411.00299 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Wang, X., Pan, J., Ding, L., Biemann, C.: Mitigating hallucinations in large vision- language models with instruction contrastive decoding. In: ACL. pp. 15840–15853 (2024)

    2024

  28. [28]

    arXiv:2405.08603 (2024)

    Xiao, H., Zhou, F., Liu, X., Liu, T., Li, Z., Liu, X., Huang, X.: A comprehensive surveyoflargelanguagemodelsandmultimodallargelanguagemodelsinmedicine. arXiv:2405.08603 (2024)

    Pith/arXiv arXiv 2024

  29. [29]

    Science China Inf

    Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. Science China Inf. Sci.67(12), 220105 (2024)

    2024

  30. [30]

    In: CVPR

    Yu, Q., Li, J., Wei, L., Pang, L., Ye, W., Qin, B., Tang, S., Tian, Q., Zhuang, Y.: HalluciDoctor: Mitigating hallucinatory toxicity in visual instruction data. In: CVPR. pp. 12944–12953 (2024) CEBaG for Medical VQA Hallucination Detection 11

    2024

  31. [31]

    arXiv:2411.11919 (2024)

    Zhang, R., Zhang, H., Zheng, Z.: VL-Uncertainty: Detecting hallucination in large vision-language model via uncertainty estimation. arXiv:2411.11919 (2024)

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.