List of Papers Browse by Subject Areas Author List
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 answer-level hallucination detection method for short-form medical VQA 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.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6556_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/masadi-99/CEBaG
Link to the Dataset(s)
VQA-RAD dataset (radiology VQA; clinician-generated questions on CT, MRI, and X-ray images): https://osf.io/89kps/
SLAKE dataset (bilingual English–Chinese semantically-labeled medical VQA): https://www.med-vqa.com/slake/
PathVQA dataset (pathology-image VQA): https://github.com/UCSD-AI4H/PathVQA
BibTex
@InProceedings{AsaMoh_Deterministic_MICCAI2026,
author = { Asadi, Mohammad AND Nedaee, Tahoura AND O’Sullivan, Jack W. AND Ashley, Euan AND Adeli, Ehsan},
title = { { Deterministic Hallucination Detection in Medical VQA via Confidence-Evidence Bayesian Gain } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes CEBaG, a deterministic hallucination detection method for medical VQA using token-level variance and visual evidence gain without external models.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper’s major strengths are its computational efficiency (only 3 forward passes vs. 20+ generations for baselines).
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The formula σ⋅(1+E) of CEBaG is a heuristic design and lacks strict probabilistic or information-theoretic basis. Section 2.2 claims that this is “posterior uncertainty weighted by visual information gain”, but the multiplication combination method has no theoretical derivation support. Why not use addition or a more complex function form? The ablation experiments show (Table 3), the simple σ alone can achieve 67.0% AUC, while the complete CEBaG only improves to 67.9%, indicating that the gain mainly comes from the variance term, and the multiplicative effect of the evidence term is negligible.
2.The fatal limitation of “white box” requirements. On page 8 of the paper, it is acknowledged that “the limitation lies in its reliance on access to model log-probabilities, restricting usage to white-box models”. This is a significant drawback in the actual clinical environment. Many commercial medical AI systems (such as Google’s Med-PaLM M series) only provide API access and do not expose token probabilities.
3.The experimental setups actually only involved 3 small-scale datasets (VQA-RAD had only 451 samples, and PathVQA had 1000 samples). This sample size is insufficient for statistical significance testing (the paper did not report confidence intervals or significance tests), and the paper lacks cross-language and cross-modal (such as dermatology, endoscopy) validations.
4.The paper did not conduct a thorough analysis of the failure scenarios of CEBaG. From Table 1, on the SLAKE dataset of MedGemma-1.5, CEBaG (with an AUC of 57.5%) was even lower than the random level (assuming 50% as the baseline), almost on par with SE (57.7%). The authors did not explain why the method completely failed under this specific setting, which is crucial for understanding the boundaries of the method.
5.The two answers in Fig. 1 show a significant difference in quality. The grounded example involves specific anatomical localization (“right temporal lobe”). The hallucinated example is merely a simple existence judgment (“spleen is present”). This suggests that the method may be effective only for detailed descriptive answers, but has insufficient ability to distinguish between short yes/no answers (which are very common in clinical VQA). However, the paper does not discuss the impact of answer length or type on the performance.
6.The caption of Fig. 1 (b) is Incomplete and only the upper part is revealed. For the answer in Fig. 1 (b), after “yes” there is a comma, but in the horizontal axis, a period is shown.
7.Figure 1(b) shows a “textbook-like” grounded answer (G = 38.5) and a “textbook-like” hallucination (G = 1.6). This extreme case comparison is misleading. In fact, according to the ablation experiments in Table 3, the AUC of “E only” is only 59.0%, which is much lower than 67.0% of “σ only”, indicating that the G values of a large number of samples are in a blurry area (intermediate values), rather than this idealized polarization. In the lower right corner of Fig. 1, there is an “asterisk” indicating that it typically represents “AI-generated” in academic publication standards. The use of AI-generated charts in medical AI papers is ironic: a paper claiming to be able to detect AI hallucinations may have its own charts that are themselves the result of AI hallucinations.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(1) Strong Reject — must be rejected due to major flaws
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
(1) Theoretical weakness—the CEBaG formula is heuristic with no rigorous derivation, and ablation shows the evidence term adds only 0.9% AUC over variance alone. (2) Misleading figures—Fig.1 cherry-picks extreme cases without distribution plots, with suspected AI-generated visualization lacking transparency.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal fails to address the paper’s fundamental flaws. First, the formulation σ·(1+E) is exposed as post-hoc rationalization: the tuned variant CEBaGλ outperforms it by +4.0 AUC (Table 3), proving the fixed formula is substantially suboptimal, not theoretically optimal. Second, the evidence term E is individually weak (59.0% AUC alone, near-random) and only marginally helps in long answers, yet clinical VQA is dominated by short responses, which means E’s practical contribution negligible. Third, the white-box claim is misleading; unlike VASE, which operates on black-box text outputs, CEBaG requires internal per-token log-probabilities inaccessible via standard APIs, severely limiting deployment. Fourth, the paper offers zero human clinician validation and relies on circular GREEN auto-labeling, while completely ignoring hallucination-type stratification (CEBaG fundamentally cannot detect knowledge), only hallucinations where visual and textual priors concur on a medically wrong answer, the most dangerous failure mode in practice. Combined with cherry-picked AI-generated figures, the work does not meet MICCAI standards for methodological rigor or clinical relevance.
Review #2
- Please describe the contribution of the paper
This paper presents CEBaG, a hyperparameter-free and training-free method for hallucination detection in medical MLLMs. CEBaG harnesses mutual information between the response and the image as evidence gain, complemented by token-level predictive variance during the decoding stage, to precisely evaluate posterior uncertainty and detect hallucinations. Extensive experiments on three medical VQA benchmarks and four medical MLLMs demonstrate its effectiveness.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The computation of evidence gain by evaluating the difference between the posterior and prior using Bayes’ rule is an interesting approach. Simplifying the evidence gain as the mutual information between the response and the image is both intuitive and reasonable. 2.The entire process of obtaining the CEBaG score requires only three forward passes, without the need for additional model training or external knowledge as guidance, making it extremely efficient for hallucination detection. 3.The comprehensive experiments conducted across four base models, three benchmarks, and four comparison methods provide strong and convincing validation of the proposed CEBaG’s performance.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Since CEBaG detects hallucinations from two perspectives: response uncertainty and visual evidence. A more detailed discussion, along with case studies, is needed to explore the contributions and interplay of these two components. For instance, under what circumstances does uncertainty in the response play a critical role, and in which cases does visual evidence become dominant? 2.CEBaG_λλ in Section 3.3 is not sufficiently explained in the methodology section. Adding a scalar weight to the method results in significant performance gains, which raises the question of whether more adaptive designs could further improve performance, rather than adhering to a strictly hyperparameter-free approach. 3.Since CEBaG requires three forward passes, it is unclear whether sampling parameters, such as the sampling policy (e.g., Top-k/p) or temperature, could influence the final results in hallucination detection. Evaluating the sensitivity of the method to these parameters would enhance the study’s robustness. 4.According to the ablation study in Table 3, response uncertainty appears to have a much larger impact on performance than per-token evidence gains. The authors should provide a more detailed discussion on these findings and elaborate on the implications of this observation. Minor comments: The image quality of Fig. 1 needs improvement. The bottom-right portion of Fig. 1 is incomplete and should be revised to display all content clearly.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed CEBaG score offers an efficient solution for detecting hallucinations in medical MLLMs without relying on external resources. The theoretical framework used to derive the CEBaG score is well-constructed. However, further discussion on the practical utility of the two components in CEBaG would strengthen the work. Including case studies with specific examples of uncertainty values and visual gains would provide deeper insights into their individual roles and contributions.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Although the paper proposes a novel framework for hallucination detection, the use of AI tools for figure generation should be treated with caution, especially when the figure involves numerical results, as in Fig. 1.AI-generated figures may mislead readers and undermine the reliability of the paper.
Review #3
- Please describe the contribution of the paper
This paper proposes Confidence-Evidence Bayesian Gain (CEBaG), a deterministic method for hallucination detection in medical Visual Question Answering (VQA). CEBaG identifies hallucinations by combining token-level predictive variance, which captures inconsistent model confidence, with evidence magnitude, which measures visual grounding via the log-probability shift between multimodal and text-only inference. Unlike previous sampling-based approaches, CEBaG requires no external models or stochastic generations. The method is evaluated across four medical multimodal large language models (MLLMs) and three public benchmarks: VQA-RAD, SLAKE, and PathVQA.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The method requires no training.
- The method is less computationally demanding than existing ones.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.I have two concerns regarding Eqn. (1). 1a) “This formulation can be interpreted as posterior uncertainty weighted by visual information gain.” How does the formulation handle cases with low σ but high positive G? 1b) Following the question above, the relative contribution between σ and E is indeed an important factor (hyper-parameter) to consider. Otherwise, cases with low σ but high positive G would be incorrectly identified as hallucinations. By the way, what is the magnitude comparison between these two variables?
2.Table 1: Although the proposed method can only be applied to open-source models, it is still important and necessary to compare against cutting-edge proprietary models.
3.”Furthermore, our method maintained stable detection performance across GREEN thresholds between 0.4 and 0.8, …” This claim is not supported by any evidence.
4.Table 2: There are no direct efficiency metrics, e.g., FLOPS, latency, or memory use, reported.
5.Table 3: What hyper-parameter(s) are swept over in CEBaGλ?
6.In Introduction: “The resulting score, σ · (1 + |G|/L), is parameter-free, …” The formulation is introduced later. It is confusing to place it here without any explanation. Similarly, G is used before it is introduced in Section 2.1. 7.The metrics AUC and AUG are not explained.
8.Fig. 1(b): How is the per-token evidence gain, for each token, computed?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
- It is better to move Fig. 1 to earlier places, instead of experiments. In addition, the sub-caption of Fig. 1(b) is incomplete.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The formulation of CEBaG is straightforward and intuitive, and demonstrates gains in experiments. However, the potential flaws in the simplistic formulation, as well as the missing experimental contents, concern me.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addressed some of my concerns. I keep my original, positive rating.
Author Feedback
We thank the reviewers and AC, and address all the points below.
Meta/R1.1/R3.1: The formulation is derived to be minimal, principled, and hyperparameter-free. (i) σ is the standard token-level predictive variance widely used in LM uncertainty estimation. (ii) Sec. 2.1 derives G from Bayes’ rule as the pointwise mutual information between response r and image x_v conditional on x_q, and E=|G|/L is its length-normalized magnitude. Both are independently grounded. (iii) The multiplicative combination is a deliberate design satisfying three requirements: f(0,E)=0 (a confident answer with high visual sensitivity is non-hallucinated, per R3 Q1a), monotonicity in both arguments, and freedom from per-dataset normalization. σ(1+E) is the simplest function meeting all three. Additive σ+αE (R1.1) violates f(0,E)=0, requires tuning α, and flags confident image-grounded answers as hallucinations. The parameter-free score is therefore principled, not arbitrary.
Meta/R1.1/R2.4: Per Table 3, CEBaG beats σ-only in 10/16 settings, with gains up to +4.7 in the long-answer dominated datasets. The E term is specifically helpful in longer answers with a small number of image-sensitive tokens (e.g. cases with linguistically sound but hallucinated observations, which would otherwise hide the uncertainty of important tokens behind confident grammatical tokens). The +0.9 average understates E’s contribution by dilution with the majority short-answer questions (VQA-RAD closed, SLAKE).
R1.4: This is a model-dataset property, not a CEBaG-specific failure: AvgProb=57.4, SE=57.7, VASE=59.5, RadFlag=57.2, CEBaG=57.5 - all five methods within 2.5 AUC. MedGemma-1.5 produces uniformly confident outputs on SLAKE, compressing the σ signal all uncertainty methods rely on. CEBaG remains best in 13/16 settings overall.
R1.5/1.6/1.7: Fig. 1, though based on real measurements (MedGemma/VQA-RAD idx=317, 405), is purely a schematic to help readers understand the concept of evidence gain. The empirical claims rest entirely on Tables 1-3 over 16 settings, not on this figure. The figure was stylized with assistance from a generative AI tool. We will replace it with a non-stylized version, fix the caption, and move it to Sec. 1 (also per R3).
R1.2/R3.2: We use “white-box” deliberately instead of “open-source”, meaning access to the log-probabilities. (1) Many frontier labs’ completions APIs (incl. OpenAI) support logprobs, so CEBaG is applicable in principle to commercial models, albeit with several API calls instead of the 3-forward-pass. (2) For medical VQA specifically, local hosting is the dominant deployment pattern under HIPAA/GDPR. The discontinuation of Med-PaLM M by Google in favor of open models such as MedGemma and MedGemma-1.5 further affirms this. (3) This is not a limitation only for CEBaG. All compared methods, including the state-of-the-art VASE (MICCAI 2025), share the same limitation.
R1.3: We strictly followed the VASE evaluation protocol for fair, direct comparison. The total N=2512 across three benchmarks matches prior MICCAI work on this task. This is an evaluation only work (no training), and therefore significance testing is well powered. The current datasets span different modalities such as general radiology (VQA-RAD) and pathology (PathVQA).
R3.1a/b: The multiplication specifically accounts for cases with low sigma but high positive G, by using the evidence gain to amplify the token-level uncertainty. In our settings, sigma is typically ~0.45+/-0.19 and E is typically ~0.71+/-0.73. R2.2/R3.5: CEBaG_λ uses min-max-normalized E and sweeps a single scalar weight lambda in [-1.5,+1.5], selecting the λ that maximizes AUC per (model,dataset). λ is the only hyperparameter, reported as an upper bound, not the proposed method.
R2.3/R3.4/R3.6-7: Scoring is teacher-forced/deterministic (Top-k/p N/A). CEBaG≈1.5x one generation vs. SE/VASE’s added 7B encoder. AUC/AUG and G will be defined explicitly in Sec. 3.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers acknowledge the efficiency and practical appeal of the proposed training-free hallucination detection method. However, significant concerns remain regarding the theoretical foundation of the proposed formulation, particularly the heuristic combination of uncertainty and evidence terms, as well as the limited empirical validation supporting the contribution of the evidence component. Authors need to carefully read the reviews and address the points, especially for R1.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper presents an efficient training-free method, but key concerns remain after rebuttal. R1 kept rejection, though R2/R3 are positive with the rebuttal. This is not an easy decision and the AI-assisted figure issue should be handled cautiously, but it further raises concerns about presentation reliability.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Reviewers have acknowledged the merits of the paper: relevance of problem, neat problem formulation, and efficient computation, while raised concerns on limited type of hallucinations that it can address (global MIRAGE and token prob std), the use of GenAI to generate figures/charts, and unstable performance (sometimes counterproductive) of E. In the camera-ready, the authors must 1.clearly narrow down the scope of the type of halluciantion it addresses and avoid selling it as a generic hallucination identification approach; 2.re-draw the figures and qualitative examples that reflect the authentic nature of the work as accurate as possible without the use of GenAI; 3.be more transparent about the parameter searching approach of the lambda and more clearly admit the stability limitations of their vanilla approach.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes an efficient, training-free hallucination detection method for medical VQA. The hallucination problem is timely, and the method is practical because it avoids stochastic sampling and external models in the loop.
The main concern comes from R1: the formulation is somewhat questionable, and the contribution of the evidence term is modest. Although this concern remains valid, the rebuttal provides a reasonable justification for the design, and CEBaG did outperform the variance-only baseline in most AUC settings. The formulation could be further investigated, and the paper provides a good opportunity for such discussion.
The paper could also be strengthened by adding clinical validation on a subset, rather than relying solely on GREEN-generated labels. The method is only validated on open-source models, which I think is minor limitation. Overall, the method is efficient and sufficiently validated for acceptance.