acceptodds
Under review as a conference paper at ICLR 2027

EVIDENCE THAT CAN CONTRADICT: WHAT REPORT- AUDITING METRICS ACTUALLY MEASURE IN MEDICAL VLMS

Abstract

As vision-language models (VLMs) begin drafting radiology reports in clinical settings, their safety relies on audit signals that flag which statements require physician review. The community typically evaluates these signals by pooling all generated claims into a single ranking and reporting the AUROC. We show that this standard metric is dominated by a prior that requires no visual information at all. Pooled claim-level AUROC decomposes exactly into within-concept and between-concept comparisons; on MIMIC-CXR, to of the pairs it scores compare different concepts, where the ordering is largely settled by knowing which findings are generally stated correctly. We derive a prior ceiling - the maximum score achievable by any predictor constant within a (concept, polarity) group, computable directly from claim counts without a model or an image. Across five generators from M to B parameters, this ceiling exceeds the strongest visual audit signals, including a recently published black-box auditor, by to AUROC. On the standard CheXbert-14 evaluation, the between-condition share is to and the ceiling is to , so reported pooled AUROCs should be read against this ceiling, not . Because the ceiling is an oracle, we also rank claims by training base rates alone; the visual audit signals beat this realisable baseline, but by a surprisingly narrow margin. We propose a stratified metric that removes this prior, under which the base-rate predictor scores exactly , and show that it reorders current auditing methods.

Primary area:
datasets and benchmarks
Keywords:
evaluation metrics, benchmark design, medical vision-language models, radiology report generation, hallucination detection, dataset bias, shortcut learning, visual grounding, chest X-ray

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wZoYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper examines what pooled claim-level AUROC measures when evaluating auditors of radiology reports. It separates within-concept from between-concept comparisons and derives the maximum AUROC achievable by scores that are constant within specified groups. Across five generators, this oracle ceiling exceeds the tested auditors, while a predictor based on training base rates remains competitive. The paper proposes evaluation within groups and studies claim triage and regional dependence across several audit signals.

Critical strengths
  1. The paper addresses an important evaluation problem: strong rewards for group-level priors can obscure an auditor's ability to distinguish errors within a finding and polarity.

  2. The decomposition and ceiling provide clear, useful diagnostics for interpreting pooled scores.

  3. Mathematical proofs help to improve the theoretical contribution of this work.

Critical weaknesses
  1. Some claims about the source of predictive information go beyond the mathematical results. A 90 to 94% share of between-concept pairs does not mean that priors account for the same fraction of a method's achieved AUROC. Visual discrimination can also improve these comparisons.

  2. Stratification changes the evaluated claims and their weighting, alongside removing cross-group comparisons.

  3. The relationship between the proposed metric and practical triage needs clarification. The probe leads the stratified ranking on all five generators. Yet at 20% review, it catches fewer errors.

  4. Correctness is defined through agreement with parsed reference reports, treating unmentioned concepts as absent and excluding claims outside the concept bank.

  5. The citation of Self-Supervised Learning From Images With a Joint-Embedding Predictive Architecture is not correct. I-JEPA is accepted to CVPR 2023 rather than ICCV.

Questions for authors

Please refer to the Weakness Section.

Reviewer_hmimOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a problem with an evaluation method: after a medical vision-language model generates a radiology report, an "audit signal" is needed to flag which statements may be wrong and require physician review. The standard practice in the community is to pool all generated claims into a single ranking and report the pooled AUROC. The paper shows that this metric does not measure what its name suggests: 90%-94% of its weight falls on between-concept comparisons, and this kind of ordering can be settled using only the statistics of the training set, without looking at the image at all. The authors derive a "prior ceiling" — an upper bound on AUROC computed purely from claim counts, with no model and no image — and this ceiling exceeds every visual audit signal tested across five generators (61M to 4.3B parameters) by 0.11–0.22. Even RadFlag, a published auditor, scores below a text-only baseline that ranks claims by disease base rates. Finally, the paper uses a controlled experiment to show that high accuracy does not mean the model looks at the right place (two configurations reach almost identical AUROCs while differing five-fold in their dependence on the finding's region), and proposes a corrected metric stratified by (concept, polarity), under which the text-only baseline scores exactly 0.5 (chance level) and the method ordering is completely reversed.

Critical strengths
  1. The analysis gives the community a useful checking tool. The within/between decomposition of AUROC is not new mathematics, but the prior ceiling derived in the paper can be computed from claim counts alone. Anyone who knows the claim distribution of a published paper can check how much of its reported AUROC comes from the prior, without a model and without images. This is a key step from "pointing out the problem" to "providing a practical auditing tool."

  2. The experimental design is solid. Five generators with very different architectures and scales are evaluated under a single protocol (no prompt leakage, greedy decoding); besides the baselines implemented by the authors themselves, the paper also runs the published RadFlag and carefully handles its coverage limitation. In addition, the decomposition identity is verified numerically to high precision, and the key conclusions are supported by multiple stratum floors and bootstrap confidence intervals.

  3. The idea of a controlled "instrument" has some methodological novelty. By changing the action space of a world model (CheXWorld) from spatial transforms to clinical statements, the authors can vary the model's dependence on anatomical regions while holding its accuracy fixed, which cleanly proves that "accuracy does not constrain where the model reads."

Critical weaknesses
  1. The core insight has limited originality. That "text priors can carry the metric of a visual task" was already established in the VQA literature: question-only models that never see the image score close to full systems [1,2], and the problem structure is identical. The paper's contribution is to make this phenomenon precise in the medical auditing setting, not a conceptual breakthrough; the novelty of its impact lies mainly in the application scenario.

  2. The external validity of the headline conclusion is limited. Except for RadFlag[3], all audit signals are baselines implemented by the authors themselves, and some stronger methods in the literature (e.g., ConRad and CURV) are excluded because their architectures cannot audit a report from a third-party model. Therefore, "visual signals are weaker than text-only priors" strictly holds only for "the signals the authors could run," and readers may easily over-generalize it to "all published auditing methods are ineffective."

  3. The evidence chain behind the findings is rather indirect. The finding that "the similarity signal does not look at the finding at all" (0.691 from the finding's own region vs 0.687 from a control region) is measured on a two-tower encoder trained by the authors, not on the official model of the concept-bottleneck method[4]; and the "five-fold difference in region dependence" is supported by only three random seeds, which the authors themselves admit only justifies saying "about five-fold." In addition, the concept bank excludes claims that fail to map into it, which biases the evaluation toward frequent conditions and compresses its coverage of real clinical errors.

References:

[1] Goyal, Yash, et al. "Making the v in vqa matter: Elevating the role of image understanding in visual question answering." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.

[2] Agrawal, Aishwarya, et al. "Don't just assume; look and answer: Overcoming priors for visual question answering." 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2018.

[3] Zhang, S., et al. "RadFlag: A Black-Box Hallucination Detection Method for Medical Vision Language Models. arXiv 2024." arXiv preprint arXiv:2411.00299.

[4] Han, Tianyu, et al. "CLEAR: an auditable foundation model for radiology grounded in clinical concepts." Nature Biomedical Engineering (2026): 1-16.

Questions for authors

Please see the weaknesses.

Reviewer_fVC9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper researches the evaluation of the quality of a radiology report generated by vision-language models. It investigates the existing way that pools all generated claims into a single ranking and reporting the AUROC and identifies its shortcomings. This paper then derives a prior ceiling that is a maximum score achieved by a predictor under a specific condition. In addition, this work proposes a stratified metric that addresses the shortcomings identified above. Experimental study is conducted to demonstrate the properties of the involved metrics.

Critical strengths
  1. Evaluating the quality of the radiology report generated by AI models is of importance. This paper gains new insights on an existing evaluation method and proposes a new solution.

  2. This paper makes efforts to conduct a systematic and thorough investigation by considering various related factors, which is appreciated.

  3. Some parts of this work such as abstract are clearly presented.

Critical weaknesses
  1. The major issue of this work is the clarity of the presentation. In many places, especially in the experimental part, this paper becomes hard to follow. This work involves many technical concepts and methods. However, they are not clearly defined or explained before used. This has significantly affected the quality of the investigation in this paper.

  2. The technical contribution of this work seems to be limited. If understood correctly, 1) the AUROC decomposition shown in Eq.(6) largely follows the literature (Kallus & Zhou (2019)); 2) the proposition 1 is straightforward under the Gaussian distribution assumption; 3) the clinical-action world model is from CheXWorld (Yue et al., 2025). This work shall clarify its technical contribution.

Questions for authors
  1. For the evaluation of radiology report generation, the literature has seen many NLG-related or clinical-efficacy-related criteria. How are they related to this work?

  2. To objectively assess whether an audit signal or evaluation metric is effective or better than others or not, a ground-truth is needed. However, this paper does not seem to discuss this issue. Is there any annotation collected from human experts (such as board-certified radiologists) as ground truth?

  3. The first paragraph on page 2 (Line 054-058) is hard to follow. Please clarity and improve.

  4. In the paragraph of "Training and Polarity Contrast" (Line 246), please explicitly explain what is trained and what the true statement is.

  5. The decomposition of AUROC in Eq.(6) shall be improved by explicitly defining each term and explaining how it is proved (in Appendix).

  6. The paragraph between Line 287 and Line 291 needs to be more clearly explained; The paragraph between Line 313 and Line 323 is hard to follow. Please clarify and improve it too. Ideally, the experimental part could be further proofread and improved in terms of clarity.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.