Abstract

Vision-Language models (VLMs) reliability in medical diagnosis is challenged by trust-undermining hallucinations. Existing hallucination detection approaches mainly focus on identifying factual inconsistencies between generated text and reference data. While some studies analyze where models attend in images, they seldom verify whether such attention truly reflects the visual evidence supporting the generated text. To address this gap, we propose Counter-Evidence Verification (CoEV), a training-free plug-and-play framework that detects and corrects hallucinations through evidence-based factual consistency verification. CoEV performs bidirectional verification between textual assertions and visual evidence, testing whether each statement is supported by its corresponding evidence region, and assigns each statement into a four-quadrant diagnostic map capturing combinations of text factuality and visual grounding. CoEV detects hallucinated content and serves as a post hoc refinement tool, correcting hallucinations without retraining. Extensive experiments on four medical datasets show that CoEV combats hallucinations in VLMs.For hallucination detection, CoEV consistently outperforms existing methods, improving average PR-AUC and ROC-AUC by 3.0% and 3.9% absolute points respectively, with notable gains of up to 18.5% in specific VQA scenarios. For hallucination correction, it improves Micro-F1 by up to 12.5%, reduces hallucination rates by over 11.9% on medical report generation, and also boosts medical VQA accuracy. These results show that CoEV enables reliable detection and correction of hallucinations, providing clinicians with dependable, evidence-based cues for diagnosis. Code is avaliable at https://github.com/Zhounan1222/CoEV.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0281_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhoNan_Hallucination_MICCAI2026,
        author = { Zhou, Nan AND Zou, Ke AND Liu, Meng AND He, Linchao AND Zhu, Jiaqi AND Zhang, Yi AND Chen, Hu AND Fu, Huazhu},
        title = { { Hallucination Detection and Correction in Medical VLMs via Counter-Evidence Verification } },
        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

    This paper introduces Counter-Evidence Verification (CoEV), a training-free, plug-and-play framework for both detecting and correcting hallucinations in medical Vision-Language Models (VLMs). The core innovation is a counterfactual intervention approach: for each clinical claim extracted from VLM output (via a semantic parser, CheXbert), CoEV localizes the corresponding visual evidence region (via a grounding model, Maira-2), then masks that region and re-queries the VLM. If a claim persists despite masking its supporting visual evidence, it is deemed linguistically driven rather than visually grounded. Claims are mapped onto a Four-Quadrant Diagnostic Map combining textual factual correctness (f) and visual grounding attribution (z) into four interpretable categories: Q1 (correct + grounded), Q2 (incorrect + grounded), Q3 (incorrect + ungrounded), Q4 (correct + ungrounded). Only Q1 claims are considered non-hallucinated. CoEV further enables hallucination correction by prompting the VLM to rewrite flagged sentences using validated visual evidence. Evaluation on four datasets (MIMIC-CXR, IU-Xray, VQA-RAD, MIMIC-VQA) across four VLM backbones demonstrates improvements in both detection (PR-AUC +3.0%, ROC-AUC +3.9%) and correction (F1-5 +7.01%, VQA accuracy +13.04%) over existing baselines.

  • 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- Novel counterfactual intervention mechanism : The idea of masking visual evidence regions and re-querying the VLM to test the causal dependency of claims on visual input is genuinely creative and conceptually appealing. Unlike purely similarity-based or uncertainty-based approaches (RadFlag, SE), CoEV provides a causal lens on hallucination testing whether removing visual evidence changes the model’s output. This goes beyond correlation to probe actual visual grounding.

    2- Interpretable four-quadrant diagnostic taxonomy : The Q1-Q4 mapping along the factual correctness x visual grounding axes (Eq. 6) is a valuable conceptual contribution. It distinguishes between language-driven bias (Q2), visual misinterpretation (Q3), and total failure (Q4), which is clinically meaningful. Clinicians can understand why a claim was flagged, not just that it was flagged a significant advantage for trust and adoption.

    3- Unified detection and correction framework : Unlike detection-only methods, CoEV provides actionable correction for both closed-ended Med-VQA (via verification, Eq. 7) and open-ended report generation (via sentence-level refinement, Eq. 8). The correction results are compelling, particularly the +27.16% accuracy gain for Qwen2-3B on MIMIC-VQA (Table 3).

    4- Comprehensive experimental evaluation : The paper evaluates across 4 datasets spanning two task types (MRG and VQA), 4 VLM backbones of varying scales and training paradigms (general-purpose: InternVL3-2B, Qwen2-3B; medical-tuned: Lingshu-7B, LLaVA-Med), with multiple complementary metrics. This breadth provides solid evidence of generalizability across architectures.

    5- Training-free, plug-and-play design : CoEV requires no fine-tuning or additional labeled data, making it practically deployable as a post-hoc layer on any existing VLM pipeline. This is a major practical advantage for clinical adoption.

    6- Well-structured ablation study : Table 4 cleanly demonstrates that both diagnostic axes (textual and visual) contribute complementary signal, with the combined model achieving the best F1-Score (0.6112). The progressive improvement pattern is convincing.

  • 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- Simplistic masking strategy may introduce artifacts : The Evidence Intervention Module (Eq. 4) masks regions with black pixels. This is a crude intervention that could introduce distribution shift: VLMs may react to the unnatural black patch itself (as an out-of-distribution artifact) rather than to the genuine absence of visual evidence. More principled interventions (e. g. , inpainting with plausible tissue textures, Gaussian blurring, or learned perturbations) should be discussed. The authors should at minimum analyze whether the black mask itself triggers false behavioral changes in the VLM, potentially inflating the false positive rate for hallucination detection.

    2- Under-specified claim persistence function : The function phi(c_i, T’_i) in Eq. 5, which determines whether a claim persists in the re-generated output after masking, is critically under-defined. Is this exact string matching? Semantic similarity? CheXbert re-extraction? This is the linchpin of the entire CECE module, and different implementations would yield vastly different results. Without this specification, the method cannot be reproduced.

    3- Self-correction circularity concern : The correction mechanism (Eq. 8) relies on the same VLM that generated the hallucination to rewrite the flagged sentence. If the VLM has a systematic bias (e. g. , always reporting cardiomegaly), re-prompting it with “corrected” evidence may not overcome this bias. The paper does not discuss when self-correction fails or analyze the correction failure modes. A control experiment using a different VLM for correction would strengthen the claim.

    4- Insufficient statistical rigor : No confidence intervals, standard deviations, or significance tests are reported for any result across Tables 1-4.Given that the method involves stochastic components (VLM generation variability, grounding model uncertainty), the reported improvements may not be statistically meaningful. For instance, the gains on IU-Xray for LLaVA-Med detection (Table 1) are modest, and without uncertainty estimates, it is impossible to assess reliability.

    5- Ablation study limited to a single dataset : Table 4 presents ablation results only on MIMIC-CXR. The relative contribution of textual vs. visual axes may differ on datasets with different characteristics (e. g. , VQA-RAD with shorter answers vs. MIMIC-CXR with full reports). Extending the ablation to at least one VQA dataset would strengthen the analysis.

    6- Ground-truth dependency on GREEN model : The entire evaluation relies on the GREEN model [18] to generate hallucination ground-truth labels. GREEN’s own accuracy and potential systematic biases are not discussed. If GREEN shares biases with the VLMs being evaluated, the ground-truth labels may systematically undercount or overcount certain hallucination types. No validation against human radiologist annotations is provided.

    7- Limited to chest X-ray domain : All four evaluation datasets involve chest radiography. The method’s reliance on CheXbert (chest X-ray specific) and Maira-2 (radiology grounding) raises questions about generalizability to other modalities (CT, MRI, pathology, dermatology). The paper does not discuss what adaptations would be needed.

    8- Computational overhead not reported : CoEV requires running the VLM at least twice per claim (original + masked), plus CheXbert extraction and Maira-2 grounding for each claim. For a report with multiple claims, this could multiply inference cost significantly. No runtime analysis or GPU memory requirements are provided.

  • Please rate the clarity and organization of this paper

    Good

  • 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.

    (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?

    CoEV addresses the clinically important problem of hallucination detection and correction in medical VLMs with a genuinely creative approach: counterfactual visual evidence intervention. The core idea masking evidence regions and re-querying the VLM to test causal visual dependency – is more principled than existing similarity-based or uncertainty-based methods, providing a causal rather than correlational assessment of hallucination. The four-quadrant diagnostic taxonomy (Q1-Q4) adds interpretability that is valuable for clinical deployment. The unified detection + correction framework is practically useful, and the experimental evaluation is comprehensive (4 datasets, 4 backbones, both detection and correction tasks).

    The main factors supporting acceptance are:

    • Conceptual novelty: The counterfactual masking approach is a distinct and well-motivated mechanism that clearly differentiates CoEV from prior work.
    • Strong correction results : The VQA accuracy gains (avg. +13.04%) and report quality improvements are substantial and clinically meaningful.
    • Breadth of evaluation : The cross-dataset and cross-architecture evaluation demonstrates robustness.

    The main concerns are:

    • Masking artifact risk : Black-pixel masking is a crude intervention that could confound the causal interpretation.
    • Under-specification of phi : The claim persistence function is critically under-defined, hindering reproducibility.
    • Statistical rigor : No confidence intervals or significance tests.
    • Self-correction circularity : Using the same VLM for correction is methodologically questionable.
  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper proposes Counter-Evidence Verification (CoEV), a training-free and plug-and-play framework for detecting and correcting hallucinations in medical vision–language models (VLMs). The key idea is to perform claim-level counterfactual verification by masking the visual regions associated with each generated clinical claim and testing whether the claim persists. This enables a direct assessment of whether a claim is grounded in visual evidence.

    The method further integrates a four-quadrant diagnostic mapping that jointly considers factual correctness and visual grounding, providing an interpretable taxonomy of hallucination types. Extensive experiments across multiple datasets (e.g., MIMIC-CXR, IU-Xray, VQA-RAD, MIMIC-VQA) and tasks (report generation and Med-VQA) demonstrate that CoEV improves both hallucination detection and downstream task performance.

    Additionally, the framework supports post-hoc correction, including answer verification for VQA and sentence-level rewriting for reports, highlighting its practical utility in clinical AI systems.

  • 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) Novel and intuitive formulation of claim-level counterfactual verification The paper introduces a simple yet effective idea: verifying each claim by masking its associated visual evidence and observing whether the claim persists. Compared to prior global perturbation approaches, this localized, per-claim intervention better captures causal dependence between visual input and generated text.

    (2) Interpretable diagnostic framework The proposed four-quadrant taxonomy (factuality × grounding) provides a clear and clinically meaningful decomposition of hallucination types, enabling fine-grained analysis rather than binary detection. This improves interpretability and aligns well with clinical needs.

    (3) Training-free and model-agnostic design The method operates at inference time without retraining the base VLM, making it practical for real-world deployment where retraining large medical models is often infeasible.

    (4) Strong empirical coverage The evaluation spans multiple datasets, tasks, and model types (general and medical VLMs), demonstrating robustness and general applicability. Improvements in detection metrics (AUC) and downstream performance are generally consistent.

    (5) Practical correction mechanisms Beyond detection, the framework supports post-hoc correction (verification and rewriting), which is valuable for clinical workflows where actionable outputs are required.

  • 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) Key components are under-specified, affecting reproducibility Several critical modules are not described in sufficient detail, including claim extraction, evidence localization, factuality assessment, and persistence estimation. In particular, the paper suggests that evidence regions are localized by a “radiology expert,” which is not scalable in practice. If an external grounding model is used instead, this introduces hidden dependencies that weaken the claim of being training-free and may lead to error propagation. Clear disclosure of these components is necessary for reproducibility.

    (2) Potential evaluator–label leakage and circularity The use of the same or closely related tools for both label construction and factuality evaluation raises concerns about evaluation bias and circular reasoning. This setup may inflate the reported performance improvements and should be carefully justified or mitigated.

    (3) Limitations of masking-based counterfactuals The masking strategy may introduce distribution shifts or artifacts, and may not adequately handle cases where visual evidence is global, diffuse, or distributed across multiple regions (e.g., cardiomegaly or interstitial patterns). The paper does not provide a sensitivity analysis on mask size, shape, or alternative masking strategies (such as blur or inpainting), making it unclear how robust the method is.

    (4) Lack of computational efficiency analysis The approach requires multiple additional model inferences for each claim, which could result in significant computational overhead. The paper does not report runtime or scalability, especially for long reports containing many claims, making it difficult to assess practical feasibility.

    (5) Limited statistical rigor The paper does not report confidence intervals, statistical significance tests, or variability across runs. Some reported improvements are unusually large and would benefit from stronger statistical validation.

    (6) Incomplete comparison with related work The method is closely related to prior work on visual counterfactual testing and verification, but these connections are not sufficiently discussed. Additional comparisons with relevant approaches would help better position the contribution and clarify its advantages.

  • 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

    This work addresses an important problem in medical AI, and the idea of claim-level counterfactual verification is both intuitive and promising. To further strengthen the paper, I encourage the authors to:

    Provide detailed descriptions of all key components (claim extraction, grounding, factuality scoring, persistence computation), including model choices and hyperparameters. Clarify the role of external models and annotations, and refine the “training-free” claim accordingly. Add statistical analysis (confidence intervals, significance testing) to support the reported improvements. Report runtime and computational cost, especially in comparison to existing baselines. Include robustness studies on masking strategies and challenging cases (e.g., diffuse findings). Expand comparisons to include related verification and counterfactual methods.

    Overall, the work is promising and has strong potential impact if these issues are addressed.

  • 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 paper proposes a novel and intuitive framework for hallucination detection based on claim-level counterfactual verification, with strong interpretability and practical relevance to clinical AI. The empirical evaluation is broad and demonstrates consistent improvements across tasks and datasets.

    However, the work has several limitations that prevent it from being a clear accept. In particular, key components are under-specified, raising concerns about reproducibility and hidden dependencies. The use of external evaluators may introduce bias, and the lack of statistical rigor and runtime analysis weakens the empirical claims. Additionally, robustness to masking strategies and comparisons with closely related methods are insufficient.

    Overall, the paper presents a promising and potentially impactful idea, but requires additional methodological transparency and stronger evaluation to fully meet top-tier standards.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We sincerely thank the reviewers for their constructive feedback and recognition of CoEV as a novel, training-free framework for medical hallucination detection and correction. Reviewer #1’s comments reference “GLoVe” and “global-local verification,” which are unrelated to our submission; therefore, we do not provide a response to Reviewer#1.We provide point-to-point responses to the other comments, which will be integrated into the final version of the paper.

[Q] Key Components (R#2, R#3). [A] To ensure reproducibility, we clarify the implementation details as follows: (1) Claim Extraction: CheXbert as the semantic parser P(·) extracts positive disease labels (Cpos) from text. (2) Evidence Localization: The pretrained radiology grounding model MAIRA-2 generates bounding boxes using images and disease descriptions as input. (3) Persistence Estimation (ϕ): ϕ(c_i, T’i) is implemented by feeding the report T’i into the semantic parser P(·) to verify if the claim (c_i) persists as a positive finding despite the absence of visual evidence. (4) Factuality Assessment: factual correctness is determined based on CheXbert disease labels extracted from the generated report and the reference report. claims are considered factual if their labels are semantically consistent.

[Q] Intervention Strategy & Artifacts (R#2, R#3). [A] (1) Masking Strategy: Black-pixel masking ensures rigorous “information zeroing”, which ensures visual evidence is genuinely inaccessible. We leave the exploration of alternative interventions for future work and add a corresponding discussion in the limitations section. (2) As illustrated in Fig. 2 in our paper, masking Region A primarily suppresses Claim A while leaving unrelated findings largely unaffected. This localized behavioral change suggests that the VLM response is mainly influenced by evidence removal rather than by a global OOD artifact effect.

[Q] Statistical Rigor (R#2, R#3). [A] We acknowledge this omission. We have conducted repeated-run analysis and statistical significance testing, and the improvements over the strongest baselines remain statistically significant across datasets (p < 0.05). Furthermore, the consistent gains across 4 datasets and 4 VLM backbones suggest the improvements are robust rather than due to random variance.

[Q] Computational Efficiency (R#2). [A] CoEV is a training-free, plug-and-play solution, avoiding the high costs of fine-tuning. CoEV averages 3.47s per image (vs. 1.78s for VASE) on a NVIDIA L20 (48GB). Given the clinical importance of hallucination detection, we believe this additional latency is a reasonable trade-off for improved reliability and interpretability.

[Q] Clarification on Circularity (R#2). [A] We clarify that the correction mechanism is model-agnostic and does not rely on the original VLM. In our implementation, Qwen is used as the correction model, taking only textual evidence and flagged claims as input without reusing the original image.

[Q] Generalizability Beyond Chest X-ray (R#2). [A] CoEV is currently developed and evaluated on chest X-ray data. It is modality-agnostic in principle, relying on claim-level counterfactual verification rather than chest X-ray-specific modeling. Extending CoEV to other modalities is part of our future work and would require adapting parsing and grounding modules to modality-specific tools.

[Q] Evaluator-Label Independence (R#3). [A] We clarify that CoEV uses CheXbert for factuality assessment, while hallucination ground-truth labels are constructed using GREEN. Although these are different tools, some domain bias may still exist, which will be discussed in the revised manuscript. [Q] Comparison with related works (R#3). [A] Unlike VCD/SID’s global perturbations, we use evidence-guided local counterfactual interventions to explicitly test visual grounding. We will incorporate a more detailed discussion of these methodological distinctions in our updated manuscript.




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.

    [Note from PCs: review submitted by Reviewer #1 has now been removed from the system]

    I recommend inviting a rebuttal. I note that Reviewer #1 appears to have reviewed a different paper, as the review repeatedly refers to “GLoVe” and global-local verification rather than the submitted CoEV counter-evidence verification method. I therefore do not rely on Reviewer #1’s assessment in forming this recommendation.

    Based on the substantive reviews from Reviewers #2 and #3, the paper addresses an important clinical safety problem: detecting and correcting hallucinated claims in medical VLM outputs. The proposed claim-level counter-evidence verification strategy is promising, since masking visual evidence and re-querying the VLM provides an intuitive way to test whether a generated clinical statement is visually grounded. The four-quadrant diagnostic map and the extension from detection to correction are also useful and interpretable.

    However, several issues need clarification before a final decision. The rebuttal should specify key components, especially claim extraction, evidence localization, factuality scoring, and the claim-persistence function after masking. The authors should also address whether black-pixel masking introduces artifacts, discuss possible circularity in label construction and factuality evaluation, report statistical significance or confidence intervals, and provide runtime analysis. A convincing rebuttal on these points would substantially strengthen confidence in the contribution.

  • 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.

    I recommend acceptance. Reviewer #1’s review has been removed and was not considered in the final assessment. Based on the substantive reviews from Reviewers #2 and #3, the paper presents a promising and clinically relevant framework for hallucination detection and correction in medical VLMs.

    The main contribution, claim-level counter-evidence verification by masking visual evidence and checking whether the claim persists, is intuitive, novel, and practically meaningful. The four-quadrant diagnostic map also provides useful interpretability, and the evaluation covers multiple datasets, tasks, and VLM backbones.



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.

    Most of the clarity issues raised by the reviewers have been addressed in the rebuttal. The problem is timely, and the baselines are up to date. The authors should commit to releasing the data splits upon acceptance for reproducibility. Based on the content of this paper itself, I consider it sufficient for acceptance.

    However, my major concern remains the concurrent submission “GloVe”. The two papers do share a common motivation: detecting hallucinations in medical VLMs by checking whether generated claims are supported by visual evidence. The tasks also overlap, including report generation and medical VQA. The baselines and datasets overlap a lot as well (MIMIC-CXR). The methods are not identical for sure, CoEV uses visual region masking as a counter-evidence intervention, while GLoVe combines global semantic assessment with local grounding verification. But there is still methodological overlap, as both rely on the same claim extraction and visual grounding/localization.



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.

    I recommend acceptance. The paper proposes CoEV, a training free framework that detects and corrects hallucinations in medical VLMs through claim level counterfactual visual masking, organized via an interpretable four quadrant taxonomy of factuality and visual grounding. Both reviewers scored the paper 4 and recognized its conceptual novelty, clinical relevance, and broad evaluation across four datasets and four VLM backbones. The rebuttal concretely resolves the major concerns by specifying the previously under defined components (CheXbert for claim extraction and persistence, MAIRA-2 for grounding, CheXbert based factuality scoring), clarifying that correction uses Qwen rather than the original VLM, decoupling ground truth labels (GREEN) from factuality scoring (CheXbert), and reporting significance testing at p < 0.05 along with runtime numbers (3.47s vs 1.78s for VASE).



back to top