Pith. sign in

REVIEW 2 major objections 5 minor 22 references

Evaluating the Diagnostic Robustness of Vision-Language Models Under Visual and Textual Perturbations

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read High accuracy masks reliability failures in medical vision-language models: flipping slice order or label order changes up to 67.8% of diagnoses, and removing lesion evidence still produces a diagnosis in 76.1% of cases.

desk verdict A robust empirical study with a load-bearing missing check: the negative-control overcommitment claim needs accuracy of forced diagnoses to rule out incomplete masks. read the letter →

arxiv 2608.04885 v1 pith:QU7AWHWV submitted 2026-08-05 cs.CV cs.CL

Ali Khoramfar , Mohammad Javad Dousti , Alireza Mohamadian , Heshaam Faili This is my paper
classification cs.CVcs.CL
keywords vision-languagemodelsdiagnosticrobustnessbrainMRIglioblastomavsmetastasispresentation-orderstabilitylabel-swapbiasnegative-controlablationabstentionbehavior
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

High accuracy on medical benchmarks can hide that a vision-language model's diagnosis is sensitive to how the same clinical evidence is presented. On a histopathologically confirmed brain MRI cohort distinguishing glioblastoma (GBM) from metastasis (MET), this paper shows that reversing or shuffling the slice order, moving lesion slices to the front or back, or swapping the order of the two target labels in the prompt all shift predictions, with flip rates up to 48.9% for visual reordering and 67.8% for label swapping. It further shows that when all expert-annotated lesion slices are removed, some models still issue a categorical GBM or MET diagnosis in up to 76.1% of cases instead of abstaining. The paper's conclusion is that static aggregate accuracy should be complemented by stability-based metrics and negative-control abstention tests before clinical deployment.

What carries the argument

The central object is a perturbation-based robustness evaluation pipeline that compares each model's diagnosis on the baseline anatomical slice sequence and standard prompt against deliberately altered presentations: visual perturbations (Reverse Order, Random Shuffle, ROI First, ROI Last) and textual perturbations (Label Swap, Lexical Rephrasing), plus a targeted evidence ablation that removes all expert-annotated lesion slices. The pipeline uses virtual indexing, which strips file metadata and tags each slice only as 'Image N', so models cannot exploit filenames, and temperature-0 generation to remove sampling noise. The core metric is the Flip Rate, the percentage of subjects where a model changes its categorical diagnosis under perturbation, alongside the Abstention Rate for the ablation condition. This machinery isolates the contribution of presentation and framing to diagnostic decisions while holding clinical evidence fixed.

What would settle it

A reader could settle the negative-control claim by showing the ablated, non-lesion slices to a board-certified neuroradiologist and asking for a forced GBM-versus-MET choice; if the radiologist discriminates above chance, the remaining slices carry diagnostic signal and the 76.1% forced-diagnosis rate is not overcommitment. Alternatively, a simple classifier trained on only the non-lesion slices that achieves above-chance held-out accuracy would falsify the claim that evidence has been fully removed.

Watch

Extended reading notes

Core claim

The paper establishes that aggregate accuracy overestimates the clinical reliability of current VLMs on a challenging GBM-versus-MET differential diagnosis. Under evidence-preserving perturbations, where the visual content and clinical meaning are unchanged, predictions flip at rates far above what a robust model should exhibit, and the largest instabilities come from the medically fine-tuned MedGemma-27B, which changes its diagnosis in 67.8% of T2 cases when the order of the two target labels in the prompt is inverted and in 48.9% of T2 cases when the slice sequence is simply reversed. In a negative control, removing all slices that intersect expert consensus annotations of the enhancing tumor core and peritumoral edema leaves a forced categorical diagnosis in 76.1% of T1CE cases for MedGemma-27B and 58.9% for Gemini-2.5-Pro. The authors interpret these results as demonstrating a decoupling between visual parsing and diagnostic reasoning: self-reported lesion indices remain consistent across perturbations while the final diagnosis changes, showing that high accuracy on canonical inputs does not imply grounded, stable clinical decision-making.

Load-bearing premise

The claim that forced diagnoses after ablation are overcommitment rests on the assumption that the two neuroradiologists' consensus pixel-level masks capture all clinically usable GBM-versus-MET evidence, so that removing every mask-positive slice truly removes all diagnostic signal; if subtle infiltrative tumor, edema, or mass-effect signs remain on non-lesion slices, the expected-UNSURE baseline is not valid and the overcommitment measure weakens.

Editorial extensions

If this is right

  • Medical VLM deployment should include stability checks: a diagnosis that changes under slice reversal, shuffle, or label reordering cannot be trusted as a fixed patient-level answer.
  • Self-reported evidence localization is not a proxy for causal grounding; models can name the correct lesion slices while flipping the diagnosis, so verification needs behavioral consistency tests, not just rationales.
  • The negative-control ablation of removing all expert-annotated lesion slices should become a standard safety test, since a model that diagnoses without evidence is overcommitting rather than abstaining.
  • Modality matters: T2 sequences show larger flip rates and more flips into UNSURE, while T1CE flips are more often categorical, meaning reliability conclusions on one MRI sequence may not transfer to another.
  • Random Shuffle flip rates persist across multiple deterministic permutations, indicating that order sensitivity is structural rather than an artifact of a single shuffle.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 67.8% label-swap flip for MedGemma suggests a learned linguistic prior linking the class names to prototypical text; a direct probe would swap the target labels with invented neutral names to separate pure label-order bias from semantic priors.
  • The same perturbation engine could be applied to other volumetric modalities, such as CT or whole-slide pathology, and to longitudinal studies, where slice-order and framing sensitivity may compound across time.
  • The ROI First versus ROI Last asymmetry hints at a position-bias mechanism familiar from long-context language models; a continuous sweep of evidence position within the sequence would map the flip-rate curve.
  • If stability metrics were added to standard medical VLM benchmarks, models that currently report high accuracy could be re-ranked substantially, since flip rates vary from roughly 5% to 68% across models and modalities.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The manuscript evaluates the diagnostic robustness of four vision-language models (Gemini-2.5-Pro, GPT-5.4, Qwen3-VL-32B, MedGemma-27B) on a histopathology-confirmed GBM-versus-MET brain MRI cohort of 90 subjects. It measures prediction flip rates under evidence-preserving visual perturbations (reverse order, random shuffle, ROI first/last) and textual perturbations (label swap, lexical rephrasing), and it runs a negative control in which all expert-annotated lesion slices are removed, measuring how often models still issue a categorical diagnosis. The central claims are that models can show high baseline accuracy yet flip their diagnoses in up to 48.9% of cases under sequence reversal and 67.8% under label swapping, and that after lesion-slice removal they force a categorical diagnosis in up to 76.1% of cases, indicating diagnostic overcommitment that aggregate accuracy alone would not reveal.

Significance. The paper addresses an important and timely question: whether static accuracy on a canonical presentation adequately captures the clinical reliability of VLMs. The experimental design is generally careful, including temperature-0 sampling, virtual indexing to strip metadata, bootstrap confidence intervals, radiomics and random baselines, and a directionality decomposition of flips. If the headline results hold, the paper makes a useful empirical contribution by demonstrating that presentation-order and label-order changes can produce large diagnostic inconsistencies, and that negative-control testing can probe overcommitment. The flip-rate findings are likely to be robust; the negative-control finding is promising but rests on an assumption that the manuscript does not yet validate.

major comments (2)
  1. [Section 3.1 / Appendix A.2 / Table 3 / Section 4.4] The interpretation of the negative control as 'diagnostic overcommitment' assumes that the two neuroradiologists' consensus masks completely capture all GBM-versus-MET diagnostic evidence, so that after lesion-slice removal the remaining input contains no usable signal and the expected output is UNSURE. The manuscript never validates this premise at the level of the model behavior it is used to interpret. Specifically, Table 3 reports the rate of forced categorical diagnoses (e.g., 76.1% for MedGemma-27B on T1CE) but does not report whether those forced diagnoses are correct, at chance, or systematically wrong. If they are at chance, the result supports the overcommitment reading; if they are above chance, the masks are incomplete and the negative-control framing collapses. This is load-bearing for contribution 3 and for the abstract's prominent 76.1% figure. I recommend adding the accuracy of forced diagnoses after ablation (ideally with a chance-level comparison), and, if possible, an independent verification that non-lesion slices contain no diagnostic signal (e.g., a human-reader no-evidence control or an anatomical analysis of the forced-diagnosis errors).
  2. [Section 3.5 / Table 2 / Table 4] The definition of flip rate is ambiguous regarding subjects whose baseline prediction is UNSURE. Section 3.5 defines flip rate as the percentage of subjects where a model's decision diverges from its baseline prediction, which would include transitions from UNSURE to GBM/MET as flips. However, Table 2's header describes 'altered its baseline categorical diagnosis,' and the decomposition in Table 4 only categorizes Correct-to-Wrong, Wrong-to-Correct, and to-UNSURE, omitting the UNSURE-to-categorical transition. The manuscript does not state whether baseline UNSURE cases are included in the denominator, and this matters for models with substantial baseline abstention rates (e.g., Qwen3-VL-32B T2 at 28.9% and MedGemma-27B T2 at 34.4%). Please clarify the exact denominator and provide a complete transition matrix, or explicitly exclude baseline UNSURE cases from the flip-rate computation and adjust the wording in Section 3.5 accordingly.
minor comments (5)
  1. [Section 4.5 / Table 9] The phrase 'remarkably consistent' overstates what Table 9 shows; for example, MedGemma-27B T1CE IoU drops from 0.430 at baseline to 0.409 under Reverse Order, and other entries vary similarly. Consider saying 'relatively stable' and possibly adding a statistical comparison of IoU across conditions.
  2. [Appendix C.3 / Table 6] Appendix C.3 states that two out of 180 evaluation sequences exceeded MedGemma-27B's operational limits, but Table 6 shows N=88 only for T1CE while T2 has N=90. Please specify which modalities the two excluded sequences belong to, so the exclusion is fully transparent.
  3. [Table 10] Table 10's header is difficult to parse; the bracketed subcategories (e.g., 'Left/Right Swap / Lost Confidence / Shift to Bilateral') would be clearer with a separate legend or footnote.
  4. [Section 3.5 / Table 2] The term 'categorical diagnosis' is used in different senses: sometimes it means GBM/MET as opposed to UNSURE, and sometimes it means a complete diagnostic label. Fix the terminology to avoid confusion about what counts as a flip.
  5. [Appendix C.2] The exact prompts for the ROI First, ROI Last, and Targeted Evidence Ablation configurations are not shown; only the baseline and the two textual perturbations are included. Please state whether the system prompt is identical for these conditions or provide the exact prompts used.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical measurement study with independently defined metrics and no self-citations; no circular derivation found.

full rationale

This is an empirical measurement study, not a derivation, and no load-bearing step reduces to its own inputs by the paper's own equations or by self-citation. The primary metrics — Flip Rate and Abstention/Forced-Diagnosis Rate (Section 3.5) — are defined independently of the conclusions: a flip is a divergence from the model's own baseline prediction, and the negative-control rate is a raw frequency of categorical outputs under ablation; no parameter is fitted, selected, or tuned to produce the headline numbers (48.9%, 67.8%, 76.1%). The perturbations are evidence-preserving by construction (identical slice images, reordered or re-prompted), so the finding that predictions change under content-identical inputs is a genuine measurement of the models, not an artifact of circular definitions. The 'diagnostic overcommitment' claim in Section 4.4 does rely on an external premise — that the two neuroradiologists' consensus masks (Section 3.1, Appendix A.2) capture all GBM-versus-MET evidence, making UNSURE the clinically expected output after ablation — but this premise is a substantive domain assumption about the world, not a definitional equivalence, and it does not originate from the models or from the authors' prior work. The paper never verifies mask completeness (Table 3 reports only forced-diagnosis rates, not whether those forced diagnoses are correct), and the Limitations section omits this threat; that is a validity and correctness risk, not circularity. There are no self-citations: every reference is external to the authors (model technical reports by the providers, prior perturbation-consistency studies, clinical literature), no uniqueness theorem is imported from the authors' own work, and no ansatz is smuggled in via citation. The contribution is an application of established consistency-testing methodology (cited to Rosenfeld et al. 2025, Chou et al. 2025) to volumetric medical VLMs with a novel negative-control variant; this is an application, not a renaming of a known result presented as a derivation. No circularity-raising step can be quoted, so the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central empirical results stand on four domain assumptions: histopathology labels are correct, expert masks fully capture lesion evidence, the expected response after ablation is UNSURE, and temperature-0 API sampling is effectively deterministic. There are no fitted free parameters and no invented entities; the paper's contributions are measurements, not derived quantities.

assumptions (4)
  • domain assumption Histopathology-confirmed subject labels are correct ground truth for GBM versus MET.
    Section 3.1: labels derived from post-surgical laboratory pathology reports are treated as definitive; any label noise would affect baseline accuracy and flip-rate interpretation.
  • domain assumption The two neuroradiologists' consensus masks completely identify all slices containing task-critical lesion evidence (enhancing core and peritumoral edema).
    Section 3.1 and Appendix A.2: masks decide which slices are removed in the negative control; incomplete masks would leave diagnostic evidence in the ablated input.
  • ad hoc to paper After removal of all annotated lesion slices, the clinically correct response is UNSURE.
    Section 3.4: this normative expectation underpins the overcommitment claim; it is stated but not independently validated by human expert reads on ablated volumes.
  • domain assumption API evaluation at temperature 0 yields effectively deterministic outputs.
    Section 3.3 and C.1: the study treats temperature 0 as eliminating sampling variability; proprietary APIs may still have nondeterminism.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating the Diagnostic Robustness of Vision-Language Models Under Visual and Textual Perturbations." pith.science (2026). https://pith.science/paper/QU7AWHWV

@misc{pith2026260804885,
  author       = {Pith},
  title        = {Pith review of: Evaluating the Diagnostic Robustness of Vision-Language Models Under Visual and Textual Perturbations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QU7AWHWV}},
  note         = {Machine review of arXiv:2608.04885}
}
read the original abstract

Standard accuracy metrics for VLMs often mask significant reliability failures in sensitive domains. In this work, we utilize a histopathology-validated brain MRI dataset to systematically assess the diagnostic robustness of four VLM families under evidence-preserving perturbations. By reordering anatomical slices and swapping target label positions, we evaluate whether models maintain consistent predictions when clinical evidence remains invariant. Our results reveal significant vulnerabilities in presentation-order stability, with models exhibiting prediction flips in up to 48.9% of cases under simple sequence reversals. We further identify a textual selection bias, where label reordering triggers inconsistent diagnoses in up to 67.8% of cases despite identical visual inputs. Negative-control tests further reveal diagnostic overcommitment: models generate categorical diagnoses in up to 76.1% of cases after expert-annotated lesion slices are removed. These results demonstrate that high accuracy can overestimate clinical reliability, masking sensitivity to sequential presentation and textual framing that is not captured by aggregate accuracy. Our findings highlight the necessity of stability-based metrics for the deployment of VLMs in safety-critical clinical applications. Our evaluation data and code will be made public upon acceptance.

Figures

Figures reproduced from arXiv: 2608.04885 by the authors.

Figure 1
Figure 1. Overview of the evaluation framework, illustrating visual perturbations, textual perturbations, and targeted [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 11 canonical work pages

  1. [1]

    doi:10.1109/TPAMI.2024.3369699 , shorttitle =

    Zhang, Jingyi and Huang, Jiaxing and Jin, Sheng and Lu, Shijian , date =. doi:10.1109/TPAMI.2024.3369699 , shorttitle =

    work page arXiv 2024

  2. [2]

    , urldate =

    Wu, Jiayang and Gan, Wensheng and Chen, Zefeng and Wan, Shicheng and Yu, Philip S. , urldate =. 2023. doi:10.1109/BigData59044.2023.10386743 , shorttitle =

    arXiv 2023

  3. [3]

    National Science Review , volume =

    Yin, Shukang and Fu, Chaoyou and Zhao, Sirui and Li, Ke and Sun, Xing and Xu, Tong and Chen, Enhong , title =. National Science Review , volume =. 2024 , month =. doi:10.1093/nsr/nwae403 , url =

    doi:10.1093/nsr/nwae403 2024

  4. [4]

    Artificial Intelligence Review , volume=

    Few-shot adaptation of multi-modal foundation models: A survey , author=. Artificial Intelligence Review , volume=. 2024 , publisher=

    work page 2024

  5. [5]

    Steiner and Can Kirmizibayrak and Rory Pilgrim and Daniel Golden and Lin Yang , year=

    Andrew Sellergren and Sahar Kazemzadeh and Tiam Jaroensri and Atilla Kiraly and Madeleine Traverse and Timo Kohlberger and Shawn Xu and Fayaz Jamil and Cían Hughes and Charles Lau and Justin Chen and Fereshteh Mahvar and Liron Yatziv and Tiffany Chen and Bram Sterling and Stefanie Anna Baby and Susanna Maria Baby and Jeremy Lai and Samuel Schmidgall and L...

    arXiv

  6. [6]

    doi:10.3389/frai.2024.1430984 , journaltitle =

    Frontiers in Artificial Intelligence , author =. doi:10.3389/frai.2024.1430984 , journaltitle =

    work page arXiv 2024

  7. [7]

    2503.10694 , archivePrefix=

    Ahmed Alaa and Thomas Hartvigsen and Niloufar Golchini and Shiladitya Dutta and Frances Dean and Inioluwa Deborah Raji and Travis Zack , year=. 2503.10694 , archivePrefix=

    arXiv

  8. [8]

    Nature Machine Intelligence , volume=

    Visual cognition in multimodal large language models , author=. Nature Machine Intelligence , volume=. 2025 , publisher=

    work page 2025

Show all 22 references
  1. [9]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , url =

    Michalkiewicz, Mateusz and Bai, Sheena and Baktashmotlagh, Mahsa and Jampani, Varun and Balakrishnan, Guha , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , url =. 2025 , pages =

    2025

  2. [10]

    2511.11206 , archivePrefix=

    Amir Rosenfeld and Neta Glazer and Ethan Fetaya , year=. 2511.11206 , archivePrefix=

  3. [11]

    MM -R ^3 : On (In-)Consistency of Vision-Language Models ( VLM s)

    Chou, Shih-Han and Chandhok, Shivam and Little, Jim and Sigal, Leonid. MM -R ^3 : On (In-)Consistency of Vision-Language Models ( VLM s). Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.246

    2025 doi

  4. [12]

    2506.05429 , archivePrefix=

    Ashwin Ramesh Babu and Sajad Mousavi and Vineet Gundecha and Sahand Ghorbanpour and Avisek Naug and Antonio Guillen and Ricardo Luna Gutierrez and Soumyendu Sarkar , year=. 2506.05429 , archivePrefix=

    arXiv

  5. [13]

    On the Risk of Misleading Reports: Diagnosing Textual Biases in Multimodal Clinical AI

    Restrepo, David and Ktena, Ira and Vakalopoulou, Maria and Christodoulidis, Stergios and Ferrante, Enzo. On the Risk of Misleading Reports: Diagnosing Textual Biases in Multimodal Clinical AI. AI for Clinical Applications. 2026

    2026

  6. [14]

    Adams and Keno K

    Johannes Moll and Markus Graf and Tristan Lemke and Nicolas Lenhart and Daniel Truhn and Jean-Benoit Delbrouck and Jiazhen Pan and Daniel Rueckert and Lisa C. Adams and Keno K. Bressem , year=. 2510.11196 , archivePrefix=

  7. [15]

    Decoupling Visual Parsing and Diagnostic Reasoning for Vision--Language Models (GPT-4o and GPT-5): Analysis Using Thoracic Imaging Quiz Cases , author=

    Han, Dae Hee and Hwang, Eui Jin and Yoon, Soon Ho and Kim, Hyungjin and Lee, Taehee , title =. Decoupling Visual Parsing and Diagnostic Reasoning for Vision--Language Models (GPT-4o and GPT-5): Analysis Using Thoracic Imaging Quiz Cases , author=. American Journal of Roentgeno...

  8. [16]

    Aaditya Singh and Adam Fry and Adam Perelman and Adam Tart and Adi Ganesh and Ahmed El-Kishky and Aidan McLaughlin and Aiden Low and AJ Ostrow and Akhila Ananthram and Akshay Nathan and Alan Luo and Alec Helyar and Aleksander Madry and Aleksandr Efremov and Aleksandra Spyra an...

    arXiv

  9. [17]

    2507.06261 , archivePrefix=

    Comanici, Gheorghe and Bieber, Eric and Schaekermann, Mike and Pasupat, Ice and Sachdeva, Noveen and Dhillon, Inderjit and Blistein, Marcel and Ram, Ori and Zhang, Dan and Rosen, Evan and others , year=. 2507.06261 , archivePrefix=

    arXiv

  10. [18]

    2511.21631 , archivePrefix=

    Shuai Bai and Yuxuan Cai and Ruizhe Chen and Keqin Chen and Xionghui Chen and Zesen Cheng and Lianghao Deng and Wei Ding and Chang Gao and Chunjiang Ge and Wenbin Ge and Zhifang Guo and Qidong Huang and Jie Huang and Fei Huang and Binyuan Hui and Shutong Jiang and Zhaohai Li a...

    arXiv

  11. [19]

    2026 , month=

    Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2026 , month=. doi:10.1609/aaai.v40i44.41142 , number=

    2026 doi

  12. [20]

    2602.22955 , archivePrefix=

    Feng Guo and Jiaxiang Liu and Yang Li and Qianqian Shi and Mingkun Xu , year=. 2602.22955 , archivePrefix=

  13. [21]

    and Lupo, J.M

    Cha, S. and Lupo, J.M. and Chen, M.-H. and Lamborn, K.R. and McDermott, M.W. and Berger, M.S. and Nelson, S.J. and Dillon, W.P. , title =. 2007 , doi =. https://www.ajnr.org/content/28/6/1078.full.pdf , journal =

    2007

  14. [22]

    doi:10.52202/079017-4455 , editor =

    Xia, Peng and Chen, Ze and Tian, Juanxi and Gong, Yangrui and Hou, Ruibo and Xu, Yue and Wu, Zhenbang and Fan, Zhiyuan and Zhou, Yiyang and Zhu, Kangyu and Zheng, Wenhao and Wang, Zhaoyang and Wang, Xiao and Zhang, Xuchao and Bansal, Chetan and Niethammer, Marc and Huang, Junz...

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.