List of Papers Browse by Subject Areas Author List
Abstract
Vision-language models achieve increasing accuracy in medical imaging tasks, but accuracy alone cannot reveal whether models rely on the correct visual evidence. In clinical settings, decisions must be justified by evidence, not merely correct outcomes. Correct answers built on wrong evidence undermine clinical trust. We introduce the Evidence Attribution Score (EAS), a behavior-based metric for evaluating visual grounding in model decisions. It compares the images a model actually depends on, measured through leave-one-out ablation, with the images it claims to use, measured through repeated prompting. Unlike accuracy, EAS reveals whether a model genuinely relied on the relevant images when arriving at a correct answer, a distinction that is critical for clinical trust. Evaluating eight vision-language models across multi-image medical questions, we find that 40 to 79 percent of correct answers are confabulated, meaning the model identified the right answer without using the right visual evidence. Three independent causal validations support that EAS captures genuine visual grounding. Chain-of-thought prompting dramatically increases EAS in open-source models while leaving closed-source models unchanged, revealing a systematic behavioral divide between model families that is invisible to accuracy alone. Our results show that accuracy-only evaluation can systematically overestimate reliability, and that EAS highlights scenarios where deployment may require additional safeguards.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3813_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/chidwipak/eas-medical-vlm
Link to the Dataset(s)
https://huggingface.co/datasets/SuhaoYu1020/MedFrameQA
BibTex
@InProceedings{KupChi_Right_MICCAI2026,
author = { Kuppani, Chidwipak AND Halavar, Bheemappa},
title = { { Right Answer, Wrong Evidence: Revealing Visual Confabulation in Multi-Image Medical VLMs with the Evidence Attribution Score } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper’s main contribution is a black-box auditing framework for a clinically important failure mode in multi-image medical VLMs: producing a correct answer while relying on inappropriate visual evidence, which the authors term visual confabulation. Concretely, the paper constructs a Behavioral Evidence Set (BES) using leave-one-out image ablations, a Claimed Evidence Set (CES) using repeated prompting, and summarizes their overlap with the Evidence Attribution Score (EAS). The empirical contribution is a comparative study across eight open- and closed-source VLMs on 200 MedFrameQA questions, showing that answer accuracy and evidence-grounding can diverge substantially, and that chain-of-thought prompting markedly increases EAS in open-source models while leaving closed-source models largely unchanged.
- 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) It addresses an important and clinically relevant problem. The paper goes beyond standard answer accuracy and asks whether a model used the appropriate visual evidence, which is especially important for longitudinal or multi-image medical reasoning tasks. This is a meaningful gap in current evaluation practice.
2) The proposed protocol is practical and model-agnostic. A notable strength is that the method only requires input-output access. That makes it applicable not only to open-source models, but also to closed-source APIs where gradients, attention maps, or internal representations are unavailable. This gives the work clear practical relevance for real-world auditing.
3) The most interesting novelty is the combination of behavioral and self-reported evidence. The mathematically novel part of EAS itself is limited, but the paper’s more original idea is to compare what the model behaviorally depends on (via ablation) against what it says it used (via repeated evidence prompts) at the image level in a multi-image medical setting. That combination is interesting because it targets a failure mode that is difficult to observe through accuracy alone.
4) The empirical results are nontrivial and potentially important. The paper reports high rates of low-EAS behavior even among correct answers, weak correspondence between answer accuracy and EAS, and a strong chain-of-thought effect for open-source models. Even if one is not fully convinced by the metric, these findings are interesting and could motivate follow-up work.
5) The evaluation is broader than a single-point metric report. The authors do not stop at reporting EAS; they also include evidence-swap, perturbation-fragility, and Shapley-based analyses. These do not fully resolve all concerns, but they do show an effort to validate the method more carefully than many explainability papers.
- 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 theoretical novelty of the Evidence Alignment Score (EAS) is relatively constrained. Mathematically, EAS is defined as the Jaccard overlap between the Behaviorally Estimated evidence Set (BES) and the Conceptually self-reported evidence Set (CES). In this regard, the primary contribution lies in the operationalization of the BES/CES pipeline rather than the introduction of a fundamentally new metric. This work situates itself within the existing body of literature on faithfulness, hallucinations, and attribution, sharing commonalities with frameworks such as CHAIR (2018), POPE (2023), and FaithScore (2023), as well as classic explainability methods like Grad-CAM and SHAP.
2) EAS does not directly evaluate whether the model utilized clinically appropriate evidence. The metric exclusively quantifies the agreement between two model-derived proxies: behavioral estimation and self-reported attribution. Consequently, a model could rely on a spurious or non-clinical image feature and subsequently report that same image, resulting in a high EAS despite fundamentally flawed reasoning. Without ground-truth annotations for relevant images or clinician-led adjudication, the paper fails to establish a definitive link between high EAS and the use of clinically correct evidence.
3) The BES serves as a somewhat coarse proxy for evidence utilization. By relying on a single-image leave-one-out ablation—counting only those images whose removal alters the discrete final answer—the BES may overlook sub-threshold effects, distributed evidence, or synergistic dependencies between multiple images. This limitation is particularly critical in multi-image medical reasoning, where diagnostic evidence is often encoded across several frames rather than localized within a single, indispensable image.
4) The CES is heuristic-driven and highly sensitive to explanation styles. The determination of CES depends on repeated prompting and a 2-of-3 majority rule, both of which are heuristic in nature. A model’s reporting behavior, prompt sensitivity, or specific instruction-following style may influence the CES more than its actual visual attribution. For instance, a verbose model might be unfairly penalized by the Jaccard denominator, while a more concise model could receive a disproportionately high score.
5) The interpretation of Chain-of-Thought (CoT) results requires further validation. The authors posit that CoT gains indicate enhanced visual engagement in open-source models. While plausible, this conclusion is not uniquely established; since CES is itself derived from prompts, the observed improvements might reflect changes in explicit reporting behavior rather than a fundamental shift in the model’s internal visual reasoning processes.
6) Cross-model comparisons lack a strict ``apples-to-apples’’ foundation. Mean EAS is calculated only for correctly answered questions, meaning different models are evaluated on disparate subsets of the benchmark. This approach introduces potential item-selection bias. A matched analysis, restricted to a common set of questions answered correctly by all compared models, would provide a more rigorous basis for cross-model conclusions.
7) The selection of specific thresholds lacks robust justification. The 2/3 majority rule for CES and the reliability threshold ($\tau=0.3$) appear largely heuristic. Although the manuscript suggests that conclusions remain qualitatively consistent across various thresholds, it lacks a strong clinical or task-specific calibration to demonstrate that these values correspond to meaningful levels of diagnostic reliability.
8) The scope of external validation and benchmark analysis is limited. The study relies on 200 questions from a single benchmark and lacks clinician-labeled evidence for relevance, a text-only baseline, or a thorough analysis of benchmark bias. Given the paper’s emphasis on “text priors,” it remains unclear to what extent low-EAS behavior stems from model-specific confabulation versus inherent answerability from the question text or other dataset artifacts.
- 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?
My recommendation is Weak Reject. The paper addresses an important problem, is clearly written, and contains genuinely interesting empirical observations, especially the dissociation between answer accuracy and evidence-grounding behavior. I also appreciate the attempt to evaluate both open- and closed-source models in a unified way.
However, I am not yet convinced that the paper’s central metric supports the strongest claims made in the manuscript. The scoring function itself is not particularly novel, and its usefulness depends almost entirely on BES and CES being valid proxies for evidence use. At present, those proxies remain heuristic, and the paper does not provide external validation showing that EAS tracks clinically correct evidence rather than agreement between two model-derived views of evidence. In addition, cross-model comparisons are correctness-conditioned, the key thresholds are only weakly justified, and the benchmark-level analysis is still limited. For these reasons, I view the work as promising and potentially impactful, but not yet sufficiently mature in its current form.
- 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
Review #2
- Please describe the contribution of the paper
The paper introduces the Evidence Attribution Score (EAS), a simple behavior-based metric to detect “visual confabulation” in multi-image medical vision-language models. It measures the overlap between images a model actually depends on (via leave-one-out ablation to build the Behavioral Evidence Set, BES) and images the model claims to use (via repeated prompting to build the Claimed Evidence Set, CES). Applied to 200 questions from MedFrameQA across eight VLMs, it shows that 40–79% of correct answers rely on wrong or no visual evidence. It also highlights a clear difference: chain-of-thought prompting strongly improves EAS in open-source models but barely affects closed-source ones. The work argues that accuracy alone is misleading for clinical reliability and proposes EAS as a complementary metric.
- 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 core idea is timely and valuable. Many people are now discussing how VLMs can give correct answers for the wrong reasons, especially in medicine where justification matters. Defining visual confabulation clearly and separating it from regular hallucination is helpful. 2.The EAS formulation is straightforward yet insightful. Using leave-one-out ablation for behavioral dependence is model-agnostic and works for both open- and closed-source models, which is practical. The three causal validations (evidence swap, perturbation fragility, and Shapley agreement) give decent support that EAS tracks real grounding rather than just correlation. 3.The chain-of-thought finding is striking and potentially useful: it suggests open-source models have latent visual capabilities that default prompting fails to engage. This behavioral divide between model families is something accuracy metrics completely miss. The paper is clearly written, the figures are helpful, and the discussion on limitations is honest.
- 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 main limitation is the small evaluation scale. The authors use only 200 questions from MedFrameQA, even though the full benchmark has over 2,800 pairs. While they argue that the computational cost (thousands of forward passes) makes larger tests difficult, 200 examples feel thin when drawing broad conclusions about systematic differences between all open-source and closed-source models, or about confabulation rates in general. More questions, or at least a clearer justification for why this subset is representative, would strengthen the claims. 2.Another concern is the Claimed Evidence Set (CES). Asking the model three times “which images are important” and taking a 2/3 majority is a reasonable heuristic, but it still relies on the model’s own verbal reporting. It may not reliably reflect what the model actually attends to during inference. The authors could have strengthened this by comparing CES against internal signals (e.g., attention weights or activation patterns in open-source models) or by benchmarking EAS against other grounding metrics from prior work on VLM faithfulness. Without such cross-validation, it is harder to be fully convinced that low EAS always means true visual confabulation rather than limitations in how CES is elicited. 3.The paper could also discuss more explicitly how well leave-one-out captures distributed or redundant visual information across images, even though they provide Shapley results as support.
- 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
Overall I found the topic engaging and the proposed EAS a useful step toward more trustworthy evaluation of medical VLMs. The distinction between behavioral dependence and claimed evidence is worth exploring further. Addressing the scale of the study and providing more direct validation of the CES component (perhaps on open models where internal signals are accessible) would make the conclusions more robust. Minor suggestion: consider reporting results on a larger random subset if computationally feasible, or at least include an ablation on subset size.
- 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?
I recommend weak accept mainly because the problem is interesting and relevant to the MICCAI community, and the EAS idea offers a clean, actionable way to look beyond accuracy in multi-image medical VQA. The CoT results add a nice behavioral insight that could influence how people prompt models in practice. That said, the relatively small test set (200 questions) limits how strongly we can interpret the broad claims about open- vs closed-source models and overall confabulation prevalence. The CES elicitation also feels a bit indirect. These are fixable issues, and I think the paper sits right around the acceptance borderline. A solid rebuttal addressing the sample size concern or adding supporting analyses would likely push it higher for 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 authors provided a targeted rebuttal that addressed my primary concerns, though I remain cautious about the limitations of their evaluation scale.
Their statistical defense of the 200 question subset is mathematically sound. The tight bootstrap confidence intervals and strong p values indicate that the sample size is adequate to support their specific empirical claims. However, it is undeniable that 200 questions serve as a bare minimum proof of concept. While sufficient for a methodological contribution, the community will eventually need larger scale validation to confirm whether these confabulation rates generalize across broader medical tasks.
Regarding the CES heuristic, their clarification that the Jaccard similarity operates purely on discrete image IDs rather than text tokens resolves the verbosity bias concern. The most robust technical point in the rebuttal is the evidence that CoT prompting structurally alters the Behavioral Evidence Set. Proving that the actual behavioral dependency shifts, rather than just the text output, is a highly valuable finding for prompt engineering.
While Reviewer 1 is correct that the lack of clinician adjudicated ground truth is a severe limitation, I accept the authors’ concession that EAS is strictly a behavioral proxy. Developing rigorous, annotation free evaluation protocols to test evidence utilization and visual reasoning is practically necessary for auditing modern multi-image VLMs. The proposed black box framework is genuinely novel and pushes the boundary of how we evaluate medical AI beyond simple accuracy. I am raising my score to Accept, conditional on the authors strictly avoiding any overclaims regarding clinical correctness in the final manuscript.
Review #3
- Please describe the contribution of the paper
The paper studies whether multi-image medical vision-language models arrive at correct answers using the correct visual evidence, rather than relying on shortcuts such as textual priors. It claims to introduce the Evidence Attribution Score (EAS), which compares an ablation-based Behavioral Evidence Set (BES) with a prompt-based Claimed Evidence Set (CES). It uses their overlap to quantify alignment between actual and claimed evidence. The paper evaluates this framework on a multi-image medical VQA benchmark across several open- and closed-source models, reports high rates of “visual confabulation” among correct answers, and presents additional analyses including chain-of-thought prompting and several causal-style validation experiments.
- 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 empirical effort is substantial. The paper evaluates several model families, compares open- and closed-source systems, studies the effect of image count, and includes additional validation experiments beyond the main metric.
2.The chain-of-thought result is interesting. The observation that CoT substantially increases EAS in open-source models while leaving closed-source models largely unchanged is potentially useful for future work on inference-time safeguards and model auditing.
3.The paper addresses an important and practically relevant problem. In medical multi-image settings, answer accuracy alone is not sufficient if the model does not rely on the appropriate visual evidence. This is a worthwhile evaluation question for clinically motivated AI systems.
- 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.My primary concern is not that the paper applies ideas to medical imaging. That is entirely reasonable for MICCAI. My concern is that the paper presents the broader methodological idea as more novel than it is, while failing to properly acknowledge closely related prior work from ML/CV/xAI and VQA faithfulness. In particular, the paper does not cite (or talk about) important works on models being right for the wrong reasons, shortcut learning, faithful explanations, and faithful/plausible grounding, including Right for the Right Reasons [R1], Unmasking Clever Hans Predictors [R2], ERASER [R3], Taking a HINT [R4], and Measuring Faithful and Plausible Visual Grounding in VQA [R5]. These are not peripheral citations. They are directly relevant to the paper’s central framing. As a result, the paper’s contribution is not positioned accurately enough relative to prior work. The contribution would be much more convincing if framed as an application and adaptation of these ideas to the setting of multi-image medical VLMs, rather than as if the broader conceptual gap were largely unaddressed. In its current form, the paper risks presenting a community’s existing ideas as if they were substantially new in this manuscript.
2.There are also several critical citation errors or misuses, which weaken the credibility of the technical framing: (a) The paper gives the wrong arXiv identifier for MedVH; the cited identifier does not match the MedVH paper [R6]. (b) The paper states that its threshold choice is analogous to slight-to-fair agreement in Cohen’s kappa, but cites Pearl’s Causality for this claim. That is not an appropriate citation for a kappa-interpretation statement. If the paper wants to use that language, it should cite a standard agreement-interpretation reference such as [R7].
References: [R1] Ross, A.S., Hughes, M.C. and Doshi-Velez, F., 2017, August. Right for the right reasons: training differentiable models by constraining their explanations. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (pp. 2662-2670).
[R2] Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W. and Müller, K.R., 2019.Unmasking Clever Hans predictors and assessing what machines really learn. Nature Communications, 10(1), p.1096. [R3] DeYoung, J., Jain, S., Rajani, N.F., Lehman, E., Xiong, C., Socher, R. and Wallace, B.C., 2020, July. ERASER: A benchmark to evaluate rationalized NLP models. In Proceedings of the 58th annual meeting of the Association for Computational Linguistics (pp. 4443-4458).
[R4] Selvaraju, R.R., Lee, S., Shen, Y., Jin, H., Ghosh, S., Heck, L., Batra, D. and Parikh, D., 2019.Taking a hint: Leveraging explanations to make vision and language models more grounded. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 2591-2600).
[R5] Reich, D., Putze, F. and Schultz, T., 2023, December. Measuring faithful and plausible visual grounding in VQA. In Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 3129-3144).
[R6] Gu, Z., Chen, J., Liu, F., Yin, C. and Zhang, P., 2026.MedVH: Toward systematic evaluation of hallucination for large vision language models in the medical context. Advanced Intelligent Systems, 8(1), p.2500255. [R7] Landis, J Richard and Gary G. Koch. “The measurement of observer agreement for categorical data.” Biometrics 33 1 (1977): 159-74 .
- 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
The paper would be significantly stronger if it were reframed more modestly and more accurately. In particular, I encourage positioning the work as an application and adaptation of evidence-faithfulness ideas to the multi-image medical VLM setting, rather than as if the broader methodological idea were largely missing from prior work. A careful revision of the related work and several key references is necessary.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I find the empirical study interesting and the application worthwhile, but the paper currently does not meet the standard I would expect for scholarly positioning. The related work and citation practice need substantial correction.
- 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.
After reading the rebuttal, I maintain my recommendation to reject. The rebuttal acknowledges the missing related work and citation issues, but these are only promised for the camera-ready version, and I cannot assess whether the paper will be properly reframed without seeing the revised discussion and related-work text. More importantly, my concern is not only missing citations, but that the paper’s central claims remain stronger than what the metric supports. EAS is still an agreement score between two model-derived proxies, not evidence that the model used clinically appropriate visual information. BES remains a coarse leave-one-out proxy, CES remains prompt-dependent, the thresholding choices remain heuristic, and the evaluation is still limited to 200 questions without clinician-labeled evidence or a stronger benchmark-bias analysis. The rebuttal improves the framing in principle, but it does not sufficiently resolve the core validity and overclaiming concerns. Therefore, I do not think the paper is ready for acceptance in its current form.
Author Feedback
We thank the Reviewers and ACs for recognizing our clinical relevance, practical protocol, and impactful CoT findings. Major concerns are addressed below.
1) Novelty & Literature (Reviewers 1 & 3): Our main contribution is the BES/CES framework and the resulting Evidence Attribution Score (EAS), which together provide a practical black-box auditing protocol for testing whether multi-image medical VLMs arrive at correct answers for the right visual reasons when gradient access is unavailable. We thank Reviewer 3 for highlighting relevant prior work. While [R1] and [R3] rely on gradient access or token-level rationales, and [R4]-[R5] focus on single-image VQA grounding, our setting is annotation-free and uses image-level evidence across multiple medical images. We agree that [R1]-[R5] provide important conceptual foundations and will cite and discuss them in the Introduction and Related Work. We will also correct the MedVH citation [R6] and replace the Pearl citation with Landis & Koch (1977) [R7] in the camera-ready version.
2) Evaluation Scale (Reviewers 1 & 2): Due to computational cost, we sampled a balanced subset of 200 MedFrameQA questions (50 each for 2-5 images). Unlike standard VQA benchmarking, each EAS evaluation requires repeated CES prompting alongside leave-one-out causal ablations across all images. Evaluating the full benchmark across 8 models requires >200,000 forward passes and ~700,000 API calls/prompts, making exhaustive auditing prohibitive. Our stratified subset (Sec 4) yielded tight 95% bootstrap CIs (width 0.06-0.12) and achieved p<0.001 for all primary findings, providing sufficient statistical precision.
3) Clinical Ground Truth & BES Limitations (Reviewers 1 & 2): EAS is a behavioral proxy, not a clinician-adjudicated ground-truth metric (Sec 6). While high EAS does not guarantee clinical correctness, low EAS on a correct answer conservatively signals potential confabulation and text-prior reliance. Reviewer 1 asks if low EAS reflects benchmark-level text answerability. If low EAS were purely a dataset artifact, we would expect similar BES=∅ rates across models. Instead, substantial model-specific variation (e.g., open-source vs. closed-source) supports that EAS captures genuine visual dependence differences. While leave-one-out ablation may miss synergistic dependencies, our Shapley analysis demonstrated moderate-to-strong agreement (0.393) with LOO BES, supporting LOO as a computationally feasible lower bound.
4) CES Heuristics & Verbosity Bias (Reviewers 1 & 2): The EAS threshold (τ) is a tunable hyperparameter. While findings are robust across τ∈[0.1,0.5], we intentionally set a lenient baseline (τ=0.3) to avoid artificially inflating failure rates from redundant slices. Under even higher evaluated thresholds (e.g., τ=0.7), confabulation predictably soars to 70-90%. Yet, our low baseline still exposes very high rates of confabulation. During pilots, 5 and 7 CES prompts yielded no major shift in decisions. Since 3 prompts already cost 48k queries, scaling further wastes compute without improving results. Jaccard sets contain discrete image IDs (e.g., {Img1, Img3}), not text tokens. A verbose model claiming Img1 in a paragraph has a CES size of 1, identical to a concise model.
5) CoT Gains & Cross-Model Comparisons (Reviewer 1): The strongest evidence that CoT changes visual behavior, not just explanation style, is that BES itself changes. BES is measured entirely through ablation and does not depend on explanation text. Under default prompting, open-source models frequently exhibit text-prior dependency (BES=∅). Under CoT, these cases drop substantially across models, meaning image removal changes the answer more often. This supports the interpretation that CoT increases genuine visual dependence rather than merely changing reporting style. Finally, on the strict intersection of correctly answered questions across all models, rankings and conclusions remain consistent with the main text.
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.
As both the merits (relevance, model-agnostic design, interesting empirical findings) and limitations (failure to acknowledge/compare existing works in the general ML domain, lack of justification for key experimental design choices) stand out, as mentioned by reviewers, the authors are invited to submit a rebuttal.
- 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 received mixed reactions after the rebuttal. Reviewers agree on the relevance of the question and the reasonable clarification regarding the sample size, while having remaining concerns over the over-simplified approach (the overlap between leave-one-out results and self-reported importance) when positioned against the more complex inconsistency/hallucination detection contexts. Taking the reviews, responses, and follow-up comments together, it is reasonable to say that the proposed approach (more like for behavioral consistency under simplified settings) does not fully support the claimed contributions toward addressing the complex mirage + hallucination problem.
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.
Personally I like this paper. It makes a highly important conceptual contribution, separates “getting the answer right” from “using the right evidence,” which is directly relevant for clinical trust. The finding that 40–79% of correct answers are confabulated is striking. I recommend the acceptance of this work.
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 is accepted as a borderline but meaningful contribution. The reviewers generally agreed that the work addresses an important and underexplored problem in evaluating evidence grounding and visual confabulation in multimodal medical AI, particularly for closed-source multi-image VLMs. The rebuttal successfully addressed several major concerns, including additional analyses on CoT impact, statistical clarification for the small dataset setting, and improved transparency regarding the evaluation protocol. At the same time, some limitations remain. In particular, the current EAS/BES/CES metrics should be framed more carefully as heuristic proxies for evidence alignment rather than direct measures of clinical validity, and the paper should avoid overstating causal or clinical conclusions. The final version should also incorporate the missing explainability and faithfulness-related citations raised during review and further clarify the limitations of the proposed evaluation framework. Despite these remaining concerns, the benchmark and auditing protocol provide a practical and novel evaluation perspective that is likely to be valuable for the community.
