Pith. sign in

REVIEW 5 major objections 9 minor 36 references

CX-Mind: A Pioneering Multimodal Large Language Model for Interleaved Reasoning in Chest X-ray via Curriculum-Guided Reinforcement Learning

T0 review · 5 major / 9 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 A chest X-ray model that alternates visible think–answer steps, trained with curriculum RL and verifiable process rewards, claims a 25.1% average gain over comparable CXR-specific models and top recall@1 on 14 real-world diseases.

desk verdict A serious, well-engineered CXR reasoning model whose headline gains need patient-disjoint re-evaluation before I trust the magnitudes. read the letter →

arxiv 2508.03733 v1 pith:2HBLB5G5 submitted 2025-07-31 cs.LG cs.AIcs.CLcs.CV

Wenjie Li , Yujie Zhang , Haoran Sun , Yueqi Li , Fanrui Zhang , Mengzhe Xu , Victoria Borja Clausich , Sade Mellin
show 20 more authors
Renhao Yang Chenrun Wang Jethro Zih-Shuo Wang Shiyi Yao Gen Li Yidong Xu Hanyu Wang Yilin Huang Angela Lin Wang Chen Shi Yin Zhang Jianan Guo Luqi Yang Renxuan Li Yang Xu Jiawei Liu Yao Zhang Lei Liu Carlos Gutiérrez SanRomán Lei Wang
This is my paper
classification cs.LGcs.AIcs.CLcs.CV
keywords chestX-raymultimodallargelanguagemodelinterleavedreasoningcurriculumreinforcementlearningverifiableprocessrewardsradiologyreportgenerationmedicalvisualquestionansweringspatiotemporalalignment
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

Chest X-ray diagnosis is usually delivered as a one-shot judgement: the model sees the image and emits an answer, sometimes after a long chain of thought that the user cannot inspect. This paper tries to establish that a different paradigm works better—an interleaved 'think–answer' format in which the model alternates between visible reasoning fragments and public sub-answers, so each step can be checked and interrupted. To make that format learnable, the paper builds a large instruction dataset (CX-Set, over 2.6 million samples) and trains in four stages, ending with reinforcement learning whose rewards are computed on the reasoning process itself, not just on the final answer. If the paper is right, interpretability and accuracy do not have to trade off in medical imaging: the reported average gain over comparable chest X-ray-specific models is 25.1%, with top performance on a real-world 14-disease dataset and higher clinician preference ratings.

What carries the argument

The load-bearing mechanism is CuRL-VPR—curriculum-based reinforcement learning with verifiable process rewards—applied to an interleaved 'think–answer' sequence. Each response is a chain of paired fragments, a <think> block followed by an <answer> block, so every intermediate conclusion is machine-parsable and human-readable. The reinforcement-learning backbone is Group Relative Policy Optimization (GRPO), which samples several candidate responses and normalizes rewards within the group to compute policy advantages. The new ingredient is a conditional process reward: the model receives extra credit for intermediate thinking and answer steps only when the format is valid, the final answer is correct, and the current batch accuracy exceeds the exponential moving average of previous batches, with thinking scored by BLEU-1/ROUGE-L similarity to real report text and intermediate answers scored all-or-none. This gating is what lets the paper claim process supervision without a pretrained reward model and without rewarding every step unconditionally.

What would settle it

Re-run the in-domain evaluations with strict patient-disjoint splits, grouping all studies in the public chest X-ray training databases by patient identifier before sampling the 2,000 test cases per task; if the average advantage over chest X-ray-specific baselines falls below the claimed 25.1%, or the open-ended disease identification gap on either in-domain dataset disappears, the generalization claim is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that interleaved reasoning can be made the native output format of a medical multimodal model, and that this format, once trained through a curriculum, is what delivers both accuracy and interpretability. The model produces alternating <think> and <answer> fragments: for close-ended questions it examines each option in turn and then concludes; for open-ended questions it lists differential diagnoses, confirms or rules out each one using image evidence, and then states the final diagnosis. Training proceeds through a medical text warm-up, large-scale instruction tuning on CX-Set, cold-start supervised fine-tuning on 42,828 report-supervised interleaved reasoning chains, and Group Relative Policy Optimization that moves from closed-ended to open-ended tasks. Rewards are rule-based and conditional: intermediate thinking steps get a lexical-similarity bonus against ground-truth report text, and intermediate answers get an all-or-none bonus, but only when the output format is correct, the final answer is correct, and the current batch's accuracy exceeds the running average of earlier batches. The paper reports that this system outperforms general and medical multimodal baselines across visual understanding, text generation, and spatiotemporal alignment (matching images to reports, distinguishing views, tracking progression, and localizing findings), with an average improvement of 25.1% over comparable chest X-ray-specific models, and that on the real-world Rui-CXR dataset its mean recall@1 across 14 diseases substantially surpasses the second-best result.

Load-bearing premise

The reported in-domain gains assume that the randomly selected test samples contain no patient who also appears in the training set; the main public chest X-ray database contains multiple studies per patient, and the paper does not state that the split was done at the patient level.

Editorial extensions

If this is right

  • If the reported gains hold, chest X-ray AI can be made inspectable without sacrificing accuracy: clinicians can read each sub-answer and intervene when a reasoning step goes wrong.
  • The reward design shows that meaningful process supervision can come from real radiology reports via lexical similarity, removing the need for a separately trained reward model in medical reinforcement learning.
  • The curriculum ordering—closed-ended tasks before open-ended diagnosis—provides a general recipe for reinforcement learning in sparse-reward medical domains, where stable reward signals are otherwise difficult to design.
  • Because the same interleaved format covers visual understanding, text generation, and spatiotemporal alignment, the paradigm unifies many chest X-ray tasks under one interaction protocol rather than requiring task-specific output heads.
  • The real-world evaluation suggests the approach transfers beyond academic benchmarks, while the paper's own results note that accuracy still falls on rarer disease categories such as lung lesion and pleural other, marking where further work is needed.

Reading between the lines

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

  • A direct extension the paper does not run: apply the same curriculum RL with report-grounded process rewards to other imaging modalities such as CT, MRI, or ultrasound; nothing in the mechanism is chest-X-ray-specific except the data it consumes.
  • The exponential-moving-average gate makes the reward signal itself curriculum-like—early batches get no process bonus until the model is already improving—so this could be studied as a general stabilizer for reinforcement learning in other sparse-reward multimodal domains.
  • The interleaved format enables a human-in-the-loop protocol the paper only gestures at: a clinician who flags a wrong sub-answer could trigger regeneration of just the remaining reasoning chain, a testable product-level extension of the claimed interpretability benefit.
  • The main validity check to run next is at the data level: re-split the public training database with strict patient-level grouping, because the paper does not state that its random test samples are patient-disjoint, and same-patient leakage would inflate the reported in-domain margins.
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

5 major / 9 minor

Summary. The paper introduces CX-Mind, a multimodal large language model for chest X-ray interpretation that generates interleaved 'think–answer' reasoning chains and is trained with a curriculum-guided reinforcement learning scheme (CuRL-VPR). The authors construct a large instruction-tuning dataset, CX-Set, with 708,473 images and 2,619,148 samples, and use GPT-4o to generate 42,828 interleaved reasoning traces supervised by clinical reports. Training proceeds in four stages: text-only medical SFT, large-scale visual instruction tuning, cold-start SFT on interleaved reasoning data, and GRPO-based RL with rule-based format, final-answer, and conditional process rewards. The model is evaluated on in-domain MIMIC-CXR and CheXpert tasks, an out-of-domain OpenI benchmark, a private Rui-CXR clinical dataset, and a multi-center expert study, reporting an average improvement of 25.1% over comparable chest X-ray-specific models.

Significance. If the reported results hold, the paper makes a useful contribution to medical MLLM research. Its strengths include a very large curated instruction dataset, a clearly specified four-stage curriculum, a rule-based process reward that avoids pretrained reward models, extensive ablations isolating training stages, data composition, and reward definitions, and external validation on a private hospital dataset with expert assessment. The interleaved reasoning format is a reasonable and practically motivated design for clinician-facing interpretability. The main threats to the paper's central claims are evaluation leakage in the in-domain split and a partially circular process-reward construction; these are fixable but currently load-bearing.

major comments (5)
  1. [Section 4.1, Table 2, Figure 5] The in-domain evaluation split is not documented as patient-disjoint. Section 4.1 states that 2,000 samples per task category were randomly selected as the test set from MIMIC-CXR and CheXpert, with the remainder used for training. Both datasets contain multiple studies per patient, and the temporal tasks (image-text matching, disease progression) are constructed by pairing studies of the same patient over time. A random sample-by-sample split can therefore place the same patient's images in both training and test, inflating the in-domain accuracy numbers that underlie the 25.1% aggregate claim and the reported spatiotemporal gains. The authors should re-run the split at the patient level, report the patient-ID disjointness explicitly, and recompute all in-domain metrics, or provide an analysis demonstrating that leakage does not affect the reported margins.
  2. [Section 3.4.2, Eq. (12)] The process reward r_think is computed as a BLEU-1/ROUGE-L similarity between the generated thought fragment and a reference reasoning text tau_i^(k) that was generated by GPT-4o from the same clinical reports that define the final-answer labels. This means the claimed 'verifiable process supervision' is partly a similarity score to model-generated text rather than a direct measure of alignment with clinical ground truth. The authors should clarify the provenance of the reference reasoning traces, and should provide a human-evaluation or report-grounded measure showing that higher r_think corresponds to more accurate and clinically faithful reasoning rather than merely to closer imitation of GPT-4o's wording.
  3. [Section 4.2, Table 2, Figure 5] The metric for open-ended disease identification is ambiguous. The text says accuracy is computed with a threshold of Jaccard > 0.5, but Table 2 reports CheXpert open-disease accuracy values around 11.5–13.0 while Jaccard values are around 43–47, and Figure 5 uses precision/Jaccard for open-ended tasks. The main text's claims of '3.1% and 17.2%' improvements on open-ended tasks are not clearly tied to a single well-defined metric. The authors should define the exact accuracy computation, report both metrics consistently in all tables and figures, and re-derive the headline improvement numbers under a pre-specified metric.
  4. [Tables 2–3, Figure 5] No confidence intervals or significance tests are reported for any of the main comparisons. Many of the improvements over strong ablations are small: for example, Table 3 shows that CX-Mind exceeds 'Base (Think)' by only 1.3 points on average, and several individual rows differ by less than one point. With 2,000 test samples per task, the authors should report bootstrap confidence intervals or paired significance tests, at least for the aggregate 25.1% claim and for the differences between CX-Mind and the strongest baselines, so that 'significantly outperforms' is supported statistically.
  5. [Section 4.7.3, Figure 10] The expert evaluation is presented as confirming clinical utility, but no inter-rater reliability statistic is reported. The text states that approximately 10% of cases were re-sampled to check intra-rater consistency, yet no kappa, ICC, or agreement value is given. Without this, the expert scores could be driven by a single rater's preferences, and the claim that interleaved reasoning is preferred by clinicians across multiple centers is not quantitatively supported.
minor comments (9)
  1. [Section 1] There is a duplicated phrase 'consistently consistently' in the first paragraph of the introduction.
  2. [Section 4.2] The heading 'Evaluation Metircs' contains a typo; it should read 'Evaluation Metrics'.
  3. [Sections 3.3.2 and 3.4.1] Equation (4) is used twice: once for the SFT loss in Section 3.3.2 and again for the interleaved reasoning format in Section 3.4.1. Please renumber the equations to avoid ambiguity.
  4. [Eqs. (12)–(14), Algorithm 1] Several free parameters appear without sensitivity analysis: lambda in Eq. (14), alpha in Eq. (12), gamma in Eq. (13), the EMA condition in Eq. (11), the 120-token report threshold in Section 3.4.2, and the Jaccard/IoU thresholds in Section 4.2. The authors should state how these values were chosen and, ideally, provide a small sensitivity study.
  5. [Section 4.4] Baseline descriptions lack version and configuration details, such as the access date for closed models and the exact checkpoint for open models like CheXagent and ChestX-Reasoner. This limits reproducibility of the comparisons.
  6. [Figure 7] The caption states that circle size in panel (a) represents model parameter scale, but the parameter values are not shown on the plot; please add a legend or numerical labels.
  7. [Section 3.3.1] The list of 23 datasets includes MIMIC-III, which is a text-only database, while other entries are image or multimodal; please clarify which subsets of these datasets were used and how they were combined, and add license or version information.
  8. [Section 4.7.2] The text uses 'top-1 diagnostic accuracy' while the abstract and Figure 9 refer to 'recall@1'; please make the terminology consistent and clarify whether the reported number is recall at k=1.
  9. [Data Availability] The data availability statement says preprocess code will be available on request, while the abstract and GitHub link promise release of the source code. Please specify what will be released, when, and under what license.

Circularity Check

1 steps flagged · score 4.0 of 10

Process reward's 'ground-truth' reasoning is GPT-4o text derived from the same reports that define the answer labels, so the verifiable process supervision is partially self-referential; diagnostic results remain independently evaluated.

  1. self definitional [Section 3.4.2 and Section 3.5.2, Eq. (12)]
    "Leveraging the advanced language and vision capabilities of ChatGPT-4o, we generated interleaved reasoning processes for various task types. ... we utilized the authentic radiological findings extracted earlier as supervisory signals to minimize hallucinations. ... For the thinking reward r_think, the ground-truth reasoning text is drawn verbatim from authoritative medical reports. ... rthink = α·BLEU1(ˆτ_i^(k), τ_i^(k)) + (1−α)·ROUGE L(ˆτ_i^(k), τ_i^(k))."

    The reference reasoning text τ_i^(k) used in Eq. (12) is not an independent gold chain: it is GPT-4o output generated from the same clinical reports that supply the final-answer labels in Eqs. (8)-(9). The paper calls this text 'ground-truth reasoning ... drawn verbatim from authoritative medical reports,' but earlier in Section 3.4.2 it states that ChatGPT-4o generated the interleaved reasoning processes. Therefore the 'verifiable process reward' is partly a lexical-overlap score against a model-generated paraphrase of the label source, not an external verification of reasoning.

full rationale

The central diagnostic and text-generation results are measured against real report-derived labels and an external Rui-CXR dataset with expert evaluation, so those headline numbers are not forced by the process-reward loop. However, the paper's signature contribution—rule-based verifiable process rewards (CuRL-VPR)—contains a circular component: the 'ground-truth' reasoning chain is produced by GPT-4o from the same reports that define the disease labels, and Eq. (12) rewards lexical similarity to that model-generated text while Section 3.5.2 describes it as verbatim from authoritative medical reports. This closed loop affects the claimed process-supervision mechanism, but not the independent final-answer evaluation. The possible patient-level leakage in the Section 4.1 random split is a correctness and external-validity risk rather than a circularity, so it is not scored here.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper rests on standard ML training assumptions plus several domain-specific premises about the validity of LLM-generated reasoning supervision and the cleanliness of evaluation splits. The most consequential free parameters are the hand-chosen reward weights and the EMA gate for process rewards, which are not subjected to sensitivity analysis.

free parameters (6)
  • lambda (Eq. 14) = 0.2
    Weights the format reward versus the final-answer reward. Chosen by hand; no sensitivity analysis is reported.
  • alpha (Eq. 12) = 0.3
    Balances BLEU-1 and ROUGE-L in the think reward. Chosen by hand.
  • gamma (Eq. 13) = 0.2
    Fixed bonus for perfectly matching all intermediate answer fragments. Chosen by hand.
  • EMA condition (Eq. 11) = not specified
    The process reward is granted only when the mini-batch accuracy exceeds the exponential moving average of previous batches; the EMA initial value and update rate are not given, and this condition gates all process rewards.
  • Report token threshold = 120 tokens
    Section 3.4.2 keeps reports with 'Findings' sections exceeding 120 tokens for CoT generation. This cutoff affects which reports are used for cold-start data.
  • Evaluation thresholds = Jaccard > 0.5, IoU > 0.5
    Section 4.2 defines open-ended disease identification accuracy with a Jaccard threshold and localization accuracy with an IoU threshold; these choices affect the reported accuracy numbers.
assumptions (5)
  • domain assumption BLEU-1 and ROUGE-L similarity to GPT-4o-generated reference reasoning is a valid supervisory signal for reasoning quality.
    Eq. (12) uses lexical overlap as the think reward; this presumes that matching the reference text produced by GPT-4o improves clinical reasoning rather than merely encouraging paraphrasing.
  • domain assumption GPT-4o-generated interleaved reasoning chains, supervised by clinical reports, are high-quality ground truth for cold-start and process rewards.
    Section 3.4.2 constructs 42,828 CoT examples using GPT-4o guided by real reports; the paper treats these as reliable supervision without independent verification.
  • ad hoc to paper Randomly splitting MIMIC-CXR and CheXpert samples into train and test does not leak patient information across the split.
    Section 4.1 describes random test-set selection without patient grouping; MIMIC-CXR contains multiple studies per patient, so this assumption is load-bearing for the in-domain results.
  • domain assumption Automated keyword matching on Rui-CXR reports yields accurate disease labels for the external validation set.
    Section 4.7.1 uses keyword matching followed by manual filtering to label 4,031 images; label noise would affect all models equally but could still bias the comparison.
  • standard math The GRPO framework and its KL-divergence regularization apply as standard.
    Section 3.5.1 presents the GRPO objective without proof; the paper relies on established RL theory from [3].

how reviews work

0 comments
Cite this review

Pith. "Pith review of CX-Mind: A Pioneering Multimodal Large Language Model for Interleaved Reasoning in Chest X-ray via Curriculum-Guided Reinforcement Learning." pith.science (2026). https://pith.science/paper/2HBLB5G5

@misc{pith2026250803733,
  author       = {Pith},
  title        = {Pith review of: CX-Mind: A Pioneering Multimodal Large Language Model for Interleaved Reasoning in Chest X-ray via Curriculum-Guided Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2HBLB5G5}},
  note         = {Machine review of arXiv:2508.03733}
}
read the original abstract

Chest X-ray (CXR) imaging is one of the most widely used diagnostic modalities in clinical practice, encompassing a broad spectrum of diagnostic tasks. Recent advancements have seen the extensive application of reasoning-based multimodal large language models (MLLMs) in medical imaging to enhance diagnostic efficiency and interpretability. However, existing multimodal models predominantly rely on "one-time" diagnostic approaches, lacking verifiable supervision of the reasoning process. This leads to challenges in multi-task CXR diagnosis, including lengthy reasoning, sparse rewards, and frequent hallucinations. To address these issues, we propose CX-Mind, the first generative model to achieve interleaved "think-answer" reasoning for CXR tasks, driven by curriculum-based reinforcement learning and verifiable process rewards (CuRL-VPR). Specifically, we constructed an instruction-tuning dataset, CX-Set, comprising 708,473 images and 2,619,148 samples, and generated 42,828 high-quality interleaved reasoning data points supervised by clinical reports. Optimization was conducted in two stages under the Group Relative Policy Optimization framework: initially stabilizing basic reasoning with closed-domain tasks, followed by transfer to open-domain diagnostics, incorporating rule-based conditional process rewards to bypass the need for pretrained reward models. Extensive experimental results demonstrate that CX-Mind significantly outperforms existing medical and general-domain MLLMs in visual understanding, text generation, and spatiotemporal alignment, achieving an average performance improvement of 25.1% over comparable CXR-specific models. On real-world clinical dataset (Rui-CXR), CX-Mind achieves a mean recall@1 across 14 diseases that substantially surpasses the second-best results, with multi-center expert evaluations further confirming its clinical utility across multiple dimensions.

Figures

Figures reproduced from arXiv: 2508.03733 by the authors.

Figure 1
Figure 1. CX-Mind: An MLLM for interleaved reasoning in Chest X-ray. (a) Task Definition and Data Collection: A broad suite of chest X-ray interpretation tasks, with more than 30 single-model and multimodal public datasets collected. (b) Visual Instruction Tuning: Fine-tuning with large-scale instruction data. (c) RL with Verifiable Answer Rewards: Reinforcement learning driven solely by outcome (answer) rewards. (d) CuRL-VPR… view at source ↗
Figure 2
Figure 2. CX-Mind training pipeline. (a) Enhancing Foundational Medical Capabilities: We collect public text corpora and filter them with LLM-based scoring to obtain the CX-Text, then warm up the language-model component. (b) Injecting Domain-Specific Knowledge: We curate and construct over two million QA pairs to cultivate professional X-ray interpretation capability. (c) Cultivating the Interleaved Reasoning Pattern: Based … view at source ↗
Figure 3
Figure 3. Illustration of the interleaved reasoning process required for open-ended questions. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Prompts following specific thinking-answer formats for different types of questions. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Comparative study of CX-Mind and various baseline models in visual understanding capabilities. (a) [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Comparative study of CX-Mind and various baseline models in text generation capabilities. (a) [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Comparative study of CX-Mind and various baseline models in spatiotemporal alignment capabilities. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Ablation study on the impact of training data for our trained models, including [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Evaluation of different models on Rui-CXR. (a) Diagnostic performance comparison of different models on 14 diseases. (b) Comparison of text generation performance of different models on Rui-CXR. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]
Figure 10
Figure 10. Figure 10: Expert evaluation of CX-Mind and baselines by multi-center clinicians. (a) Physician scoring workflow interface: the system randomly presents a de-identified chest radiograph and an anonymized model output. Following unified guidelines, physicians annotate the image a…
Figure 11
Figure 11. Figure 11: Comparative examples of diagnostic responses from GPT-4o, CheXagent and CX-Mind. (a) Open-ended diagnostic question results. (b) Close-ended diagnostic question results. queries. By interleaving concise thinking and answering steps, CX-Mind mirrors the diagnos￾tic wor…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 14 canonical work pages

  1. [27]

    R. Xie, D. Qiu, D. Gopinath, D. Lin, Y. Sun, C. Wang, S. Potdar, B. Dhingra, In- terleaved reasoning for large language models via reinforcement learning, arXiv preprint arXiv:2505.19640 (2025)

    arXiv 2025

  2. [16]

    S. Rui, K. Chen, W. Ma, X. Wang, Improving medical reasoning with curriculum-aware reinforcement learning, arXiv preprint arXiv:2505.19213 (2025)

    arXiv 2025

  3. [3]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al., Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, arXiv preprint arXiv:2501.12948 (2025)

    arXiv 2025

  4. [1]

    Jaech, A

    A. Jaech, A. Kalai, A. Lerer, A. Richardson, A. El-Kishky, A. Low, A. Helyar, A. Madry, A. Beutel, A. Carney, et al., Openai o1 system card, arXiv preprint arXiv:2412.16720 (2024)

    arXiv 2024

  5. [2]

    Doshi, Gemini 2.5: Our most intelligent models are getting even better, Recuperado el 15 (2025)

    T. Doshi, Gemini 2.5: Our most intelligent models are getting even better, Recuperado el 15 (2025)

    work page 2025

  6. [4]

    Y. Wang, S. Wu, Y. Zhang, S. Yan, Z. Liu, J. Luo, H. Fei, Multimodal chain-of-thought reasoning: A comprehensive survey, arXiv preprint arXiv:2503.12605 (2025)

    arXiv 2025

  7. [5]

    Holzinger, M

    A. Holzinger, M. Dehmer, F. Emmert-Streib, R. Cucchiara, I. Augenstein, J. Del Ser, W. Samek, I. Jurisica, N. Díaz-Rodríguez, Information fusion as an integrative cross- cutting enabler to achieve robust, explainable, and trustworthy medical artificial intelli- gence, Information Fusion 79 (2022) 263–278

    work page 2022

  8. [6]

    S. Ali, T. Abuhmed, S. El-Sappagh, K. Muhammad, J. M. Alonso-Moral, R. Confalonieri, R. Guidotti, J. Del Ser, N. Díaz-Rodríguez, F. Herrera, Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy artificial intelligence, Infor- mation fusion 99 (2023) 101805

    work page 2023

Show all 36 references
  1. [7]

    Z. Chen, M. Varma, J. Xu, M. Paschali, D. Van Veen, A. Johnston, A. Youssef, L. Blanke- meier, C. Bluethgen, S. Altmayer, et al., A vision-language foundation model to enhance efficiency of chest x-ray interpretation, arXiv e-prints (2024) arXiv–2401

    2024

  2. [8]

    Fallahpour, J

    A. Fallahpour, J. Ma, A. Munim, H. Lyu, B. Wang, Medrax: Medical reasoning agent for chest x-ray, arXiv preprint arXiv:2502.02673 (2025)

    2025 arXiv

  3. [9]

    Y.Lai, J.Zhong, M.Li, S.Zhao, X.Yang, Med-r1: Reinforcementlearningforgeneralizable medical reasoning in vision-language models, arXiv preprint arXiv:2503.13939 (2025)

    2025

  4. [10]

    J. Pan, C. Liu, J. Wu, F. Liu, J. Zhu, H. B. Li, C. Chen, C. Ouyang, D. Rueckert, Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning, arXiv preprint arXiv:2502.19634 (2025)

    2025 arXiv

  5. [11]

    Sharma, Cxr-agent: Vision-language models for chest x-ray interpretation with uncer- tainty aware radiology reporting, arXiv preprint arXiv:2407.08811 (2024)

    N. Sharma, Cxr-agent: Vision-language models for chest x-ray interpretation with uncer- tainty aware radiology reporting, arXiv preprint arXiv:2407.08811 (2024)

    2024 arXiv

  6. [12]

    A. S. Albahri, A. M. Duhaim, M. A. Fadhel, A. Alnoor, N. S. Baqer, L. Alzubaidi, O. S. Albahri, A. H. Alamoodi, J. Bai, A. Salhi, et al., A systematic review of trustworthy and explainable artificial intelligence in healthcare: Assessment of quality, bias risk, and data fusion...

    2023

  7. [13]

    D. Xu, Y. Chen, Z. Chai, Y. Xiao, Y. Yan, W. Ding, H. Wang, Z. Jin, W. Jiao, W. Yue, et al., Knowledge fusion in deep learning-based medical vision-language models: A review, Information Fusion (2025) 103455

    2025

  8. [14]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al., Deepseek-v3 technical report, arXiv preprint arXiv:2412.19437 (2024). 32

    2024 arXiv

  9. [15]

    Hurst, A

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Weli- hinda, A. Hayes, A. Radford, et al., Gpt-4o system card, arXiv preprint arXiv:2410.21276 (2024)

    2024 arXiv

  10. [17]

    Bluethgen, P

    C. Bluethgen, P. Chambon, J.-B. Delbrouck, R. van der Sluijs, M. Połacin, J. M. Zam- brano Chaves, T. M. Abraham, S. Purohit, C. P. Langlotz, A. S. Chaudhari, A vision– language foundation model for the generation of realistic chest x-ray images, Nature Biomedical Engineering ...

    2024

  11. [18]

    D. Ma, J. Pang, M. B. Gotway, J. Liang, A fully open ai foundation model applied to chest radiography, Nature (2025) 1–11

    2025

  12. [19]

    J. Yao, X. Wang, Y. Song, H. Zhao, J. Ma, Y. Chen, W. Liu, B. Wang, Eva-x: A foun- dation model for general chest x-ray analysis with self-supervised learning, arXiv preprint arXiv:2405.05237 (2024)

    2024 arXiv

  13. [20]

    Z. Chen, M. Varma, J.-B. Delbrouck, M. Paschali, L. Blankemeier, D. Van Veen, J. M. J. Valanarasu, A. Youssef, J. P. Cohen, E. P. Reis, et al., Chexagent: Towards a foundation model for chest x-ray interpretation, arXiv preprint arXiv:2401.12208 (2024)

    2024 arXiv

  14. [21]

    R. W. Lee, K. H. Lee, J. S. Yun, M. S. Kim, H. S. Choi, Comparative analysis of m4cxr, an llm-based chest x-ray report generation model, and chatgpt in radiological interpretation, Journal of Clinical Medicine 13 (23) (2024) 7057

    2024

  15. [22]

    J. Chen, Z. Cai, K. Ji, X. Wang, W. Liu, R. Wang, J. Hou, B. Wang, Huatuogpt-o1, towards medical complex reasoning with llms, arXiv preprint arXiv:2412.18925 (2024)

    2024 arXiv

  16. [23]

    B. Wang, H. Zhao, H. Zhou, L. Song, M. Xu, W. Cheng, X. Zeng, Y. Zhang, Y. Huo, Z. Wang, et al., Baichuan-m1: Pushing the medical capability of large language models, arXiv preprint arXiv:2502.12671 (2025)

    2025 arXiv

  17. [24]

    Zhang, P

    W. Zhang, P. Zhang, J. Guo, T. Cheng, J. Chen, S. Zhang, Z. Zhang, Y. Yi, H. Bu, Patho- r1: A multimodal reinforcement learning-based pathology expert reasoner, arXiv preprint arXiv:2505.11404 (2025)

    2025

  18. [25]

    T. Lin, W. Zhang, S. Li, Y. Yuan, B. Yu, H. Li, W. He, H. Jiang, M. Li, X. Song, et al., Healthgpt: A medical large vision-language model for unifying comprehension and gener- ation via heterogeneous knowledge adaptation, arXiv preprint arXiv:2502.09838 (2025)

    2025 arXiv

  19. [26]

    Y. Luo, J. Zhang, S. Fan, K. Yang, M. Hong, Y. Wu, M. Qiao, Z. Nie, Biomedgpt: An open multimodal large language model for biomedicine, IEEE Journal of Biomedical and Health Informatics (2024)

    2024

  20. [28]

    Bengio, J

    Y. Bengio, J. Louradour, R. Collobert, J. Weston, Curriculum learning, in: Proceedings of the 26th annual international conference on machine learning, 2009, pp. 41–48

    2009

  21. [29]

    A. E. Johnson, T. J. Pollard, S. J. Berkowitz, N. R. Greenbaum, M. P. Lungren, C.-y. Deng, R. G. Mark, S. Horng, Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports, Scientific data 6 (1) (2019) 317. 33

    2019

  22. [30]

    Irvin, P

    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 uncer- tainty labels and expert comparison, in: Proceedings of the AAAI conference on artificial intelligen...

    2019

  23. [31]

    Demner-Fushman, M

    D. Demner-Fushman, M. D. Kohli, M. B. Rosenman, S. E. Shooshan, L. Rodriguez, S. An- tani, G. R. Thoma, C. J. McDonald, Preparing a collection of radiology examinations for distribution and retrieval, Journal of the American Medical Informatics Association 23 (2) (2016) 304–310

    2016

  24. [32]

    M. Liu, W. Hu, J. Ding, J. Xu, X. Li, L. Zhu, Z. Bai, X. Shi, B. Wang, H. Song, et al., Medbench: A comprehensive, standardized, and reliable benchmarking system for evalu- ating chinese medical large language models, Big Data Mining and Analytics 7 (4) (2024) 1116–1128

    2024

  25. [33]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al., Qwen2. 5-vl technical report, arXiv preprint arXiv:2502.13923 (2025)

    2025 arXiv

  26. [34]

    Z. Wu, X. Chen, Z. Pan, X. Liu, W. Liu, D. Dai, H. Gao, Y. Ma, C. Wu, B. Wang, et al., Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding, arXiv preprint arXiv:2412.10302 (2024)

    2024 arXiv

  27. [35]

    J. Chen, C. Gui, R. Ouyang, A. Gao, S. Chen, G. H. Chen, X. Wang, R. Zhang, Z. Cai, K. Ji, et al., Huatuogpt-vision, towards injecting medical visual knowledge into multimodal llms at scale, arXiv preprint arXiv:2406.19280 (2024)

    2024 arXiv

  28. [36]

    Z. Fan, C. Liang, C. Wu, Y. Zhang, Y. Wang, W. Xie, Chestx-reasoner: Advancing radi- ology foundation models with reasoning through step-by-step verification, arXiv preprint arXiv:2504.20930 (2025). 34

    2025 arXiv

Pith tools

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