pith. sign in

arxiv: 2605.20469 · v1 · pith:XN7VEAOQnew · submitted 2026-05-19 · 💻 cs.CV

HalluCXR: Benchmarking and Mitigating Hallucinations in Medical Vision-Language Models for Chest Radiograph Interpretation

Haoyu Wang , Zitong Li This is my paper

Pith reviewed 2026-05-21 06:34 UTC · model grok-4.3

classification 💻 cs.CV
keywords hallucinationsvision-language modelschest radiographsmedical image interpretationbenchmarkensemble methodspatient safety
0
0 comments X

The pith

Vision-language models hallucinate in 62 to 82 percent of chest radiograph interpretations.

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

The paper introduces HalluCXR, a benchmark that measures how often vision-language models generate factually incorrect findings when reading chest X-rays. It tests six different models on hundreds of real radiographs and three types of questions, revealing that hallucinations appear in most outputs and that some of these errors are clinically dangerous. This matters for patient safety because these models are moving toward clinical use where made-up findings could lead to wrong diagnoses or treatments. The work also spots recurring patterns in the errors and shows that combining several models can cut down on fabrications.

Core claim

By evaluating six architecturally diverse vision-language models on 856 stratified MIMIC-CXR chest radiographs across three query types for 15,408 total outputs, the paper establishes that 61.9 to 82.3 percent of responses contain hallucinations and up to 80.2 percent include clinically dangerous errors. An eight-category taxonomy with severity ratings and a two-layer detection pipeline, validated on 250 human annotations, reveals three patterns: normal radiographs attract the most severe hallucinations, common findings are over-fabricated while rare ones are missed, and longer responses predict higher hallucination risk with AUC up to 0.908. Ensembles of models reduce fabrication by up to

What carries the argument

The HalluCXR benchmark with its eight-category hallucination taxonomy that includes clinical severity ratings and a two-layer detection pipeline validated against human annotations.

Load-bearing premise

The eight-category taxonomy and two-layer detection pipeline validated on 250 human annotations fully capture all clinically relevant hallucinations in the model outputs.

What would settle it

A larger-scale human review by radiologists on additional model outputs using a different or expanded error classification to check whether the 61.9-82.3 percent hallucination rates and severity distributions remain consistent.

Figures

Figures reproduced from arXiv: 2605.20469 by Haoyu Wang, Zitong Li.

Figure 1
Figure 1. Figure 1: Per-label fabrication and omission counts across all models. (a) Common findings (lung opacity, pleural effusion) are most frequently fabricated; (b) rare findings (enlarged cardiomediastinum, fracture) are most often omitted [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Confidence calibration vs. hallucination rate. Numbers above bars indicate sample counts per bin. Higher bars at high confidence indicate “confidently wrong” predictions. Qwen2-VL shows maximal miscalibration (ECE=0.725). BiomedCLIP Gemini 3 Flash GPT-5.1 InternVL2 LLaVA-NeXT Qwen2-VL BiomedCLIP Gemini 3 Flash GPT-5.1 InternVL2 LLaVA-NeXT Qwen2-VL 1.00 0.44 0.58 0.40 0.56 0.44 0.44 1.00 0.54 0.50 0.82 0.11… view at source ↗
Figure 3
Figure 3. Figure 3: Cross-model hallucination correlation (phi coefficients, ϕ). Correlations range from 0.11 to 0.82; Qwen2-VL shows the most distinctive pattern (ϕ≤0.45 with all other models). cator using ROC analysis ( [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Response length distribution for clean vs. hallucinated outputs (annotations show median values). Hallucinated responses are consistently longer across all models. truth or model internals. 4.7. Ensemble Mitigation Despite moderate average correlation, the wide spread in pairwise ϕ values (Section 4.5) suggests that ensembles can still exploit complementary errors. We evaluate three strate￾gies at the find… view at source ↗
Figure 5
Figure 5. Figure 5: Length-based hallucination risk scoring. (a) ROC curves show per-model AUCs from 0.712 to 0.908. (b) Hallucination rates by length quintile; rates generally increase with length (Q1 elevated by short BiomedCLIP outputs). bit quantization vs. full-precision commercial APIs, though gaps are not consistently in one direction. Data are from de-identified MIMIC-CXR-JPG [22] under PhysioNet credentialed access; … view at source ↗
read the original abstract

Vision-language models (VLMs) are increasingly used for medical image interpretation, yet they frequently hallucinate, generating clinically plausible but factually incorrect findings that pose direct patient safety risks. We introduce HalluCXR, a benchmark evaluating six architecturally diverse VLMs across 856 stratified MIMIC-CXR chest radiographs and three query types, yielding 15,408 model evaluations. An eight-category hallucination taxonomy with clinical severity ratings and a two-layer detection pipeline are validated against 250 human annotations (auto-detection F1=0.959; LLM judge F1=0.907). We find that 61.9--82.3% of outputs contain hallucinations, with clinically dangerous errors in up to 80.2%. Three key patterns emerge: normal radiographs paradoxically attract the most severe hallucinations, common findings are systematically over-fabricated while rare findings go under-detected, and response length alone predicts hallucination risk (AUC up to 0.908). A six-model ensemble reduces fabrication by up to 84.8% at the cost of increased omission; a three-model subset retains comparable performance at half the cost. These results establish that hallucination auditing, verbosity-based risk monitoring, and ensemble-based safety layers are prerequisites for clinical deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces HalluCXR, a benchmark for hallucinations in medical vision-language models interpreting chest radiographs. It evaluates six architecturally diverse VLMs on 856 stratified MIMIC-CXR images across three query types, producing 15,408 outputs. An eight-category hallucination taxonomy with severity ratings and a two-layer detection pipeline (automated + LLM judge) are introduced and validated on 250 human annotations (F1 scores 0.959 and 0.907). Key results include hallucination prevalence of 61.9–82.3% with clinically dangerous errors up to 80.2%, patterns such as more severe hallucinations on normal radiographs, over-fabrication of common findings, under-detection of rare ones, and response length as a predictor (AUC up to 0.908). Mitigation via a six-model ensemble reduces fabrication by up to 84.8%, with a three-model subset offering comparable performance at lower cost.

Significance. If the detection pipeline and taxonomy generalize beyond the validation set, the work provides a valuable large-scale empirical benchmark highlighting patient-safety risks in clinical VLM deployment. Strengths include the scale of evaluation (15,408 outputs), use of public MIMIC-CXR data, human annotation validation, and practical mitigation results (ensemble reductions). The identification of actionable patterns such as verbosity-based risk monitoring adds immediate utility for auditing systems.

major comments (1)
  1. [Methods / Results (validation subsection)] Validation of the detection pipeline (described in the methods and results sections): The eight-category taxonomy and two-layer detector are validated on only 250 human annotations out of 15,408 total outputs, with reported F1 scores of 0.959 (auto) and 0.907 (LLM judge). This small stratified sample provides limited evidence that the taxonomy captures all clinically relevant hallucination types or that the detector is free of systematic biases across models, query types, or radiograph findings (e.g., normal vs. abnormal films). If coverage or calibration does not generalize, the headline prevalence rates (61.9–82.3%) and severity claims (up to 80.2% dangerous) cannot be reliably extrapolated.
minor comments (2)
  1. [Dataset construction] The stratification details for the 856 radiographs (e.g., exact distribution of findings, normal vs. abnormal balance) are referenced but could be expanded with a supplementary table for full reproducibility.
  2. [Experimental setup] Clarify how the three query types were chosen and whether they cover the full range of clinical use cases for chest radiograph interpretation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. We address the major comment point by point below, acknowledging limitations where appropriate and outlining specific revisions to strengthen the work.

read point-by-point responses

  1. Referee: [Methods / Results (validation subsection)] Validation of the detection pipeline (described in the methods and results sections): The eight-category taxonomy and two-layer detector are validated on only 250 human annotations out of 15,408 total outputs, with reported F1 scores of 0.959 (auto) and 0.907 (LLM judge). This small stratified sample provides limited evidence that the taxonomy captures all clinically relevant hallucination types or that the detector is free of systematic biases across models, query types, or radiograph findings (e.g., normal vs. abnormal films). If coverage or calibration does not generalize, the headline prevalence rates (61.9–82.3%) and severity claims (up to 80.2% dangerous) cannot be reliably extrapolated.

    Authors: We appreciate the referee's concern regarding the scale of the human validation. The 250 annotations were obtained via a stratified sampling strategy explicitly designed to cover all six models, the three query types, and a balanced distribution of normal versus abnormal radiographs, maximizing representativeness given the substantial cost and time required for expert medical annotation. The high F1 scores reflect strong agreement with human judgments on this sample. We nevertheless agree that the sample size constrains strong claims of full generalization across every possible subgroup and that undetected systematic biases remain possible. In the revised manuscript we will expand the human validation set to 750 annotations total, with targeted oversampling of rare findings and per-model breakdowns. We will add a new subsection reporting detector performance stratified by model, query type, and finding prevalence, and we will include an expanded limitations paragraph explicitly discussing the risk of extrapolation and the value of future community validation on the released annotations. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical rates derived from external annotations and public data

full rationale

The paper is a standard empirical benchmark study. Hallucination prevalence (61.9–82.3 %) and severity claims are obtained by running the authors' eight-category taxonomy and two-layer detector over 15,408 model outputs on stratified MIMIC-CXR radiographs; the detector itself is validated on an independent set of 250 human annotations (F1 scores reported). No equations, fitted parameters, or self-citations are used to derive the headline statistics; the taxonomy is defined once and then applied, with coverage checked against external labels rather than by construction. The analysis therefore contains no self-definitional, fitted-input, or self-citation-load-bearing steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The work rests on the assumption that MIMIC-CXR labels and the chosen query templates are representative of clinical use; no free parameters are fitted to produce the headline rates.

axioms (1)
  • domain assumption MIMIC-CXR provides reliable ground-truth findings for hallucination labeling
    Used as the reference for all 15,408 evaluations
invented entities (1)
  • Eight-category hallucination taxonomy with severity ratings no independent evidence
    purpose: To classify and score model errors
    New classification scheme introduced for this benchmark

pith-pipeline@v0.9.0 · 5760 in / 1253 out tokens · 32324 ms · 2026-05-21T06:34:04.060946+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 5 internal anchors

  1. [1]

    Gemini 3 flash model card,

    Google DeepMind, “Gemini 3 flash model card,” tech. rep., Google, 2025. 1, 2

    work page 2025

  2. [2]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge,et al., “Qwen2-VL: Enhancing vision- language model’s perception of the world at any resolution,” arXiv preprint arXiv:2409.12191, 2024. 1, 2, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

    Z. Chen, W. Wang, Y . Cao, Y . Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liu,et al., “Expanding performance bound- aries of open-source multimodal models with model, data, and test-time scaling,”arXiv preprint arXiv:2412.05271, 2024. 1, 2, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    LLaV A-NeXT: Improved reasoning, OCR, and world knowledge

    H. Liu, C. Li, Y . Li, B. Li, Y . Zhang, S. Shen, and Y . J. Lee, “LLaV A-NeXT: Improved reasoning, OCR, and world knowledge.” https://llava-vl.github.io/ blog/2024-01-30-llava-next/, 2024. 1, 3

    work page 2024

  5. [5]

    LLaV A-Med: Training a large language-and-vision assistant for biomedicine in one day,

    C. Li, C. Wong, S. Zhang, N. Usuyama, H. Liu, J. Yang, T. Naumann, H. Poon, and J. Gao, “LLaV A-Med: Training a large language-and-vision assistant for biomedicine in one day,” inProceedings of NeurIPS, 2023. 1, 2

    work page 2023

  6. [6]

    V.; Valanarasu, J

    Z. Chen, M. Varma, J. Xu, M. Paschali, D. Van Veen, A. John- ston, A. Youssef, L. Blankemeier, J.-B. Delbrouck, J. P. Cohen,et al., “A vision-language foundation model to en- hance efficiency of chest X-ray interpretation,”arXiv preprint arXiv:2401.12208, 2024. 1, 2

    work page arXiv 2024

  7. [7]

    Med- Flamingo: A multimodal medical few-shot learner,

    M. Moor, Q. Huang, S. Wu, M. Yasunaga, Y . Dalmia, J. Leskovec, C. Zakka, E. P. Reis, and P. Rajpurkar, “Med- Flamingo: A multimodal medical few-shot learner,”Proceed- ings of Machine Learning Research (PMLR), vol. 225, 2023. 1, 2

    work page 2023

  8. [8]

    The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision)

    Z. Yang, L. Li, K. Lin, J. Wang, C.-C. Lin, Z. Liu, and L. Wang, “The dawn of LMMs: Preliminary explorations with GPT-4V(ision),”arXiv preprint arXiv:2309.17421, 2024. 1, 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    Med-Halt: Medical domain hallucination test for large language models,

    A. Pal, L. K. Umapathi, and M. Sankarasubbu, “Med-Halt: Medical domain hallucination test for large language models,” inProceedings of EMNLP, 2023. 1, 2

    work page 2023

  10. [10]

    MedVH: To- wards systematic evaluation of hallucination for large vi- sion language models in the medical context,

    Z. Gu, C. Yin, F. Liu, and P. Zhang, “MedVH: To- wards systematic evaluation of hallucination for large vi- sion language models in the medical context,”arXiv preprint arXiv:2407.02730, 2024. 1, 2

    work page arXiv 2024

  11. [11]

    Object hallucination in image captioning,

    A. Rohrbach, L. A. Hendricks, K. Burns, T. Darrell, and K. Saenko, “Object hallucination in image captioning,” in Proceedings of EMNLP, 2018. 1, 2

    work page 2018

  12. [12]

    Evaluating object hallucination in large vision-language mod- els,

    Y . Li, Y . Du, K. Zhou, J. Wang, W. X. Zhao, and J.-R. Wen, “Evaluating object hallucination in large vision-language mod- els,” inProceedings of EMNLP, 2023. 2

    work page 2023

  13. [13]

    FaithScore: Fine-grained evaluations of hallucinations in large vision-language models,

    L. Jing, R. Li, Y . Chen, and X. Du, “FaithScore: Fine-grained evaluations of hallucinations in large vision-language models,” inFindings of the Association for Computational Linguistics: EMNLP 2024, pp. 5042–5063, Association for Computational Linguistics, 2024. 1, 2

    work page 2024

  14. [14]

    BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs

    S. Zhang, Y . Xu, N. Usuyama, J. Bagga, R. Tinn, S. Preston, R. Rao, M. Wei, N. Valluri, C. Wong,et al., “BiomedCLIP: A multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs,”arXiv preprint arXiv:2303.00915, 2023. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    PubMedCLIP: How much does CLIP benefit visual question answering in the medical domain?,

    S. Eslami, G. de Melo, and C. Meinel, “PubMedCLIP: How much does CLIP benefit visual question answering in the medical domain?,”Findings of EACL, 2023. 2

    work page 2023

  16. [16]

    To- wards generalist foundation model for radiology by lever- aging web-scale 2D&3D medical data,

    C. Wu, X. Zhang, Y . Zhang, Y . Wang, and W. Xie, “To- wards generalist foundation model for radiology by lever- aging web-scale 2D&3D medical data,”arXiv preprint arXiv:2308.02463, 2023. 2

    work page arXiv 2023

  17. [17]

    Detecting and Evaluating Medical Hallucinations in Large Vision Language Models

    J. Chen, D. Yang, T. Wu, Y . Jiang, X. Hou, M. Li, S. Wang, D. Xiao, K. Li, and L. Zhang, “Med-HallMark: Detecting and evaluating medical hallucinations in large vision language models,”arXiv preprint arXiv:2406.10185, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    FactCheXcker: Mitigating measurement hallucinations in chest X-ray report generation models,

    A. Heiman, X. Zhang, E. Chen, S. E. Kim, and P. Rajpurkar, “FactCheXcker: Mitigating measurement hallucinations in chest X-ray report generation models,” inProceedings of CVPR, 2025. 2

    work page 2025

  19. [19]

    A survey on hallucination in large language models: Principles, taxon- omy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, and T. Liu, “A survey on hallucination in large language models: Principles, taxon- omy, challenges, and open questions,”ACM Transactions on Information Systems, 2025. 2

    work page 2025

  20. [20]

    Judg- ing LLM-as-a-judge with MT-Bench and Chatbot Arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing,et al., “Judg- ing LLM-as-a-judge with MT-Bench and Chatbot Arena,” in Proceedings of NeurIPS, 2023. 2

    work page 2023

  21. [21]

    Chatbot arena: An open platform for evaluating LLMs by human preference,

    W.-L. Chiang, L. Zheng, Y . Sheng, A. N. Angelopoulos, T. Li, D. Li, H. Zhang, B. Zhu, M. Jordan, J. E. Gonzalez, and I. Stoica, “Chatbot arena: An open platform for evaluating LLMs by human preference,” inProceedings of ICML, 2024. 2

    work page 2024

  22. [22]

    MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs,

    A. E. W. Johnson, T. J. Pollard, N. R. Greenbaum, M. P. Lun- gren, C.-y. Deng, Y . Peng, Z. Lu, R. G. Mark, S. J. Berkowitz, and S. Horng, “MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs,”Scientific Data, vol. 6, no. 1, p. 317, 2019. 2, 8

    work page 2019

  23. [23]

    CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison,

    J. Irvin, P. Rajpurkar, M. Ko, Y . Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya,et al., “CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison,” inProceedings of the AAAI Conference on Artificial Intelligence, 2019. 2

    work page 2019