Pith. sign in

REVIEW 4 major objections 5 minor 71 references

Mr3D-VL: A generalist vision language foundation model for Multiparametric 3D Magnetic Resonance Imaging

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A 4B vision-language model built for volumetric mpMRI outperforms 2D-slicing baselines on brain MRI report generation and question answering.

desk verdict A plausible architecture with a couple of genuinely new pieces, but the evaluation protocol is circular enough that the headline SOTA claim does not hold. read the letter →

arxiv 2608.12689 v1 pith:CQD6PHAT submitted 2026-08-13 cs.CV cs.AI

Zhi Qiao , Xintong Wu , Yichu He , Feng Shi This is my paper · ORCID
classification cs.CVcs.AI
keywords multiparametricMRI3Dvision-languagemodelmedicalreportgenerationvisualquestionansweringrotarypositionembeddingself-supervisedpretrainingbrainmultimodallearning
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

The paper sets out to show that a vision-language model designed for native volumetric processing can serve as a generalist assistant for multiparametric brain MRI, handling several imaging sequences at once and answering clinical questions in natural language. It claims that Mr3D-VL, a 4-billion-parameter model, surpasses existing general-purpose and medical vision-language models on mpMRI report generation and visual question answering, with BERTScore 0.856, open-ended QA accuracy 0.713, and multiple-choice accuracy 0.912. The result would matter because current models either flatten 3D volumes into 2D slices, losing spatial relationships and inflating token counts, or handle only a single modality, whereas real neuro-oncology decisions require combining sequences such as T1, T2, FLAIR, and DWI. If the claim holds, hospitals could query a model about tumor location, signal behavior, and cross-modality findings and receive interpretable, report-quality text.

What carries the argument

The load-bearing object is the 4D rotational positional embedding (4D-RoPE), which extends rotary position encoding from the usual time-height-width axes to include depth, so that a 3D volume keeps its native geometry inside the LLM instead of being flattened into slice sequences. Around this sit two further mechanisms: a shared modality-agnostic 3D encoder (a MedNext-style CNN pretrained with DINOv2-style self-supervision) that lets one set of weights handle every MRI sequence without modality-specific branches, and a multi-resolution vision token injection strategy, inspired by DeepStack, that draws low-, medium-, and high-resolution feature maps from the encoder and feeds them into the projection ViT and then into multiple early layers of the language model. Together these give the model positional awareness in volumetric space, cross-modal feature sharing, and simultaneous access to global anatomy and fine lesion detail.

What would settle it

Re-run the head-to-head with baselines fine-tuned on the same SFT data and score a held-out set of real clinical mpMRI reports by blinded radiologists; if the fine-tuned baselines match Mr3D-VL or the expert accuracy on open-ended questions falls well below 0.713, the central claim collapses.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a comparatively small model can dominate much larger and domain-specific predecessors on 3D mpMRI language tasks when the architecture respects volumetric structure. Mr3D-VL shares one modality-agnostic 3D encoder across all MRI sequences, pretrains it with self-supervised DINOv2-style learning, positions tokens in a four-dimensional coordinate system of time, depth, width, and height via rotational embeddings, and injects low-, medium-, and high-resolution visual features into the LLM decoder. Across the paper's benchmarks the model reports a BERTScore of 0.856, METEOR 0.496, and CIDEr 0.655 for report generation, BLEU-4 0.525 and BERTScore 0.794 for open-ended QA, and accuracy of 0.713 and 0.912 on open-ended and multiple-choice questions, all stated as surpassing the tested 4B/7B/30B rivals while using far fewer FLOPs and less memory.

Load-bearing premise

The load-bearing premise is that the evaluation measures real mpMRI capability: test cases come from the same LLM-generated dataset used for training, the rival models are not fine-tuned on that data, and open-ended answers are judged by an LLM rather than clinicians; if that protocol is not a fair measure, the reported superiority does not follow.

Editorial extensions

If this is right

  • If the reported numbers are right, a 4B model can generate clinically structured radiology reports from multiple 3D MRI sequences, making deployment on modest hospital hardware feasible.
  • Native 4D positioning removes the need to slice volumes, which directly cuts token counts and GPU memory, and the paper reports roughly 96% lower FLOPs than a 4B generalist under five-modality input.
  • Because a single shared encoder handles all sequences, new MRI modalities can be added without per-modality branches; the paper presents this as a path to a broader mpMRI foundation model.
  • Stable BERTScore and METEOR as the number of input modalities grows suggests native multimodal training avoids the semantic degradation the paper observes in medical baselines.
  • If the approach generalizes beyond brain MRI, the same architecture could be retrained for prostate, liver, and spine mpMRI, which the paper explicitly lists as next steps.

Reading between the lines

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

  • A direct way to separate architectural gains from data advantages is to fine-tune the baselines on the same SFT set; if they close the gap, the difference is largely the dataset rather than 4D-RoPE.
  • The open-ended accuracy judge is another LLM (Qwen3.5-27B); a small-sample human radiologist rating study would show whether the 0.713 figure reflects expert agreement.
  • The test split is drawn from the same LLM-generated corpus as training, so a prospective evaluation on real clinical reports with expert ground truth would be the stronger demonstration of clinical utility.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces Mr3D-VL, a 4-billion-parameter vision-language foundation model for multiparametric 3D brain MRI. The system combines a shared MedNeXt 3D encoder pretrained with DINOv2-style self-supervision, a 4D rotary positional encoding scheme, and multi-resolution visual feature injection into a Qwen3-4B language model. Training uses a private dataset of 103,605 mpMRI cases (460,541 images), augmented by an LLM-based pipeline that rewrites radiology reports and synthesizes question-answer pairs and multiple-choice questions. The authors report results on report generation, open-ended VQA, and multiple-choice VQA, claiming state-of-the-art performance, e.g., BERTScore 0.856 for report generation, open-ended QA accuracy 0.713, and multiple-choice accuracy 0.912.

Significance. If the reported results were supported by a sound evaluation, the paper would make a useful contribution: it addresses a real gap in 3D mpMRI vision-language modeling, and the architectural components (shared modality-agnostic encoder, 4D-RoPE, multi-resolution feature injection, DINOv2 pretraining) are well motivated and individually ablated in the appendix. The scale of the private dataset and the efficiency analysis (Table 3) are also assets. However, the central claim that Mr3D-VL 'surpassed all competing models' is not established by the evidence: the test labels are generated by the same LLM pipeline used for training, baselines are evaluated in a disadvantaged protocol, and the accuracy numbers lack external or human validation and statistical confidence. The current evaluation is therefore insufficient to support the paper's headline conclusions.

major comments (4)
  1. [§5.1, Table 2] The evaluation is circular with respect to the claimed state-of-the-art results. The test data 'primarily originated from the fine-tuning dataset constructed during the post-training phase' (§5.1), and that dataset was produced by the same LLM-based generation pipeline that created the SFT training data (§4.2.1, Table 2). Even though the split is by patient, the test labels are machine-generated and drawn from the same distribution on which the model was fine-tuned. Consequently, the large margins in §5.3 (e.g., BERTScore 0.856 vs. 0.733 for the strongest zero-shot baseline) may reflect distribution familiarity rather than genuine 3D mpMRI capability. An external benchmark with expert-written reports and human-verified QA, or at least a held-out set whose labels were not produced by the training pipeline, is required before any superiority claim can be accepted.
  2. [§5.1, Tables 4–6] The baseline comparison is not fair. Lingshu and the Qwen3.5 series are evaluated by slicing 3D volumes into 2D image stacks, and HuluMed—which the paper itself states supports volumetric data—is also 'uniformly adopted' into the multi-image 2D-slicing protocol despite native 3D support. This disadvantages 3D-capable baselines. Moreover, no baseline is fine-tuned on the mpMRI SFT data, so the comparison is effectively zero-shot baselines versus a model fine-tuned on the test distribution. The claim in §5.3 that Mr3D-VL 'surpassed all competing models' is therefore not a controlled comparison; fair baselines should be fine-tuned on the same data and, for HuluMed, evaluated in its native 3D mode.
  3. [§5.3, Figure 4] The accuracy numbers are not statistically or clinically validated. Open-ended QA accuracy is judged by QWen3.5-27B acting as an 'external expert judge', but this model is itself one of the baselines in Table 4, and no human agreement, calibration, or error analysis is reported. Multiple-choice accuracy is derived by post-processing, but no confidence intervals, significance tests, or per-category breakdowns are provided for any metric in Tables 4 and 5. Without such validation, the claimed improvements (e.g., 0.713 open-ended accuracy and 0.912 multiple-choice accuracy) cannot be distinguished from noise or judge bias.
  4. [§4.2.1, §5.4 Future Work] The reliability of the LLM-generated labels is explicitly acknowledged as a limitation in the paper itself: Future Work calls for 'Automated Data Quality Control' to 'remove samples with semantic drift or hallucinations' and for a 'Post-Training Paradigm Shift' to incorporate human feedback. These statements indicate that the training and test labels have not been human-validated. Since the reported evaluation depends entirely on these labels, the manuscript should either provide evidence of label quality (e.g., human review of a sample, agreement statistics) or temper the SOTA claim accordingly.
minor comments (5)
  1. [§1, §3] There are numerous typographical and OCR-style artifacts throughout, including 'corss modality', 'ia', 'ns', and inconsistent terminology between 'M3RoPE' and 'M4RoPE' in §3; these should be cleaned up before resubmission.
  2. [Table 1, Table 2] The data counts in Tables 1 and 2 lack clear units; the text refers to 'approximately 1 billion tokens' for pre-training, but the table entries (e.g., 2,711,712) appear to be sample counts. Please clarify whether these are cases, samples, or tokens, and state the exact train/test split sizes.
  3. [§5.4, Figure 6] The claim that QWen3.5-27B shows a 'weaker positive correlation' than Mr3D-VL between ROUGE-L and BERTScore is made without reporting correlation coefficients; please provide the quantitative values and, ideally, confidence intervals.
  4. [Appendix, Figures .11–.12] Figure cross-references appear inconsistent: Figure .11 is referenced for both the multi-resolution token injection experiment and the DINO-vs-CLIP pretraining experiment, while Figure .12 is also referenced for DINO-vs-CLIP. Please renumber the figures and correct the references.
  5. [§5.3] The relative improvement claims are inconsistent with the tables: for example, BERTScore 0.856 vs. 0.733 is about 17% relative improvement, not 'over 20%', and open-ended BLEU-4 0.525 vs. 0.112 is far more than 30%. Please report absolute differences or define the relative-change formula explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

SOTA claim is partially self-referential: the test labels come from the same LLM-generated SFT pipeline used to train Mr3D-VL, while baselines are not fine-tuned on that distribution and are evaluated via 2D slicing.

  1. fitted input called prediction [§4.2.1 (Post-Training Data), §5.1 (Settings), §5.3 (Experimental Results), Table 2]
    "The test data primarily originated from the fine-tuning dataset constructed during the post-training phase, with detailed statistics provided in Table 2. ... The Q&A data is further categorized into open-ended questions and multiple-choice questions, with the latter directly reusing the Q&A dataset from the pre-training phase. ... For open-ended questions, in addition to the aforementioned generation-based evaluation metrics, we employed QWen3.5-27B as an external expert judge."

    Mr3D-VL's post-training is supervised fine-tuning on the dataset summarized in Table 2 (about 100k open-ended QA, 100k multiple-choice, and 89k report-generation samples). The evaluation then uses test data that 'primarily originated from the fine-tuning dataset constructed during the post-training phase,' so the ground-truth reports and answers are generated by the same LLM-based parsing/rewriting/QA-simulation pipeline that produced the SFT targets; §4.2.1 further states that the multiple-choice items are 'directly reusing the Q&A dataset from the pre-training phase.' The baselines are not SFT-tuned to this synthetic distribution: Qwen3.5 and Lingshu are evaluated as 2D slice stacks, and Hulu-Med, despite native 3D support, is also evaluated through the multi-image 2D protocol.

full rationale

The paper's central claim is empirical: 'Mr3D-VL surpassed all competing models' on mpMRI report generation and VQA. The load-bearing evaluation is partially circular because the test set is drawn from the same post-training SFT dataset that defines the training objective. The paper states this explicitly in §5.1: 'The test data primarily originated from the fine-tuning dataset constructed during the post-training phase.' Mr3D-VL was fine-tuned on nearly 300k samples of that dataset (Table 2), while the baselines are zero-shot and, for 3D-capable Hulu-Med, evaluated through 2D slicing. Thus the reported SOTA margins are not evidence of general 3D mpMRI capability; they are evidence of fit to the LLM-generated label distribution. The multiple-choice subtask is even more directly self-referential because §4.2.1 says these items are 'directly reusing the Q&A dataset from the pre-training phase,' meaning the evaluation labels come from the same question-generation machinery used to create training data. The open-ended accuracy scores are additionally assigned by QWen3.5-27B, an LLM judge, rather than by human experts. No load-bearing self-citation chain or uniqueness theorem is invoked: the architecture components cite external work (DINOv2, MedNext, Qwen3, DeepStack), and the design choices are not justified by the authors' own prior claims. The circularity is therefore localized to the evaluation protocol rather than to the architecture derivation, which keeps the score at 6 rather than higher.

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

No new physical entities are introduced; the model and its 4D position encoding are architectural constructs. The main unpaid-for inputs are the synthetic-data accuracy assumption and the evaluation protocol, both of which are load-bearing for the performance claim.

free parameters (4)
  • Resampled voxel spacing = 1 x 1 x 5 mm^3
    Chosen to match thick-slice clinical MRI; sets z-axis resolution and therefore the spatial information available to the model (§4).
  • Input crop size = [32, 192, 192]
    Fixed input volume dimensions after resampling; determines token count and memory footprint (§4).
  • Multi-resolution feature scales = 1/4, 1/8, 1/16
    Three pyramid levels from the visual encoder are fused in the projection; hand-selected in the appendix section on multi-resolution vision token injection.
  • Pretraining stage data volumes = 2,711,712 / 936,404 / 406,130 samples
    Data budgets for stages S0, S1, and S2 are set by the authors and affect the trained model (§4.1, Table 1).
assumptions (4)
  • domain assumption Synthetic LLM-generated reports and QA pairs are clinically accurate enough to serve as training and evaluation ground truth.
    The entire pre-training and post-training pipeline relies on LLM rewriting of reports and QA simulation; the paper lists automated quality control only as future work (§2.3, §4.1.2, §4.2.1, Future Work).
  • domain assumption A shared modality-agnostic visual encoder pre-trained with DINOv2 captures both intra-modal and inter-modal features relevant for MRI.
    The design assumes that one encoder for all MR modalities learns useful cross-modal representations without paired supervision (§3).
  • ad hoc to paper 4D rotary position encoding with a shared time-or-depth axis correctly unifies text and 3D volume positions in the LLM attention space.
    The paper claims this unifies positional information across modalities but provides no formal analysis; the appendix describes a different 4D scheme over x, y, z, and channel (§3, Appendix).
  • domain assumption Slicing 3D volumes into 2D stacks is a valid way to evaluate 3D-native and 2D-native baselines on equal footing.
    The evaluation protocol for all baselines assumes 2D slicing does not unfairly disadvantage models designed for volumetric input (§5.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mr3D-VL: A generalist vision language foundation model for Multiparametric 3D Magnetic Resonance Imaging." pith.science (2026). https://pith.science/paper/CQD6PHAT

@misc{pith2026260812689,
  author       = {Pith},
  title        = {Pith review of: Mr3D-VL: A generalist vision language foundation model for Multiparametric 3D Magnetic Resonance Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CQD6PHAT}},
  note         = {Machine review of arXiv:2608.12689}
}
read the original abstract

Multi-parametric magnetic resonance imaging (mpMRI) is a cornerstone for brain tumor diagnosis and treatment, yet current AI models face critical limitations: their lack of natural language interaction and interpretability impedes spatial information integration and cross-modal reasoning required clinically. Key challenges arise from significant physical meaning differences across modalities, spatial misalignment due to scan intervals, and the need for complex multi-feature interpretation in tasks like glioma grading. While visual-language models (VLMs) show promise in cross-modal understanding, existing methods focus mainly on 2D image modeling, neglecting direct perception of 3D volumetric space. Although 3D VLMs have been proposed for report generation and feature alignment in 3D CT imaging, mpMRI applications demand collaborative inference across multiple imaging modalities-a requirement unmet by current solutions. To address this, we introduce Mr3D-VL, a dedicated visual-language foundation model for multi-parametric 3D MRI. With 4 billion parameters, it employs an unsupervised pre-trained shared 3D encoder and 4D rotational positional embedding for dual modality-spatial integration. Its cross-modal projection layer uses a multi-resolution feature implantation strategy to enhance feature perception across resolutions. Experimental results show significant improvements over existing 4B/7B/30B domain-specific and general-purpose models in text generation tasks, achieving a BERTScore of 0.856 for report generation, with question-answering accuracy at 0.713 and multiple-choice accuracy at 0.912.

Figures

Figures reproduced from arXiv: 2608.12689 by the authors.

Figure 1
Figure 1. The Mr3D-VL framework integrates a vision encoder, a corss modality projection and a language model [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of Pre-training & Post-training [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Data distribution of original mpMRI data [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Accuracy evaluation of-the-art (SOTA) performance across all metrics, surpassing existing models by 30% in each benchmark: BLEU-4 (0.525), ROUGE-L (0.738), BERTScore (0.794), METEOR (0.747), and CIDEr (0.837). For open-ended questions, in addition to the aforementioned…
Figure 5
Figure 5. Figure 5: Observing model performance (BERTScore & METEOR) variations in report generation across input [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Examining performance variations of models on radiology report generation using BERTScore and [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 34 canonical work pages

  1. [1]

    Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on , pages=

    Real-time segmentation of on-line handwritten arabic script , author=. Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on , pages=. 2014 , organization=

    2014

  2. [2]

    Soft Computing and Pattern Recognition (SoCPaR), 2014 6th International Conference of , pages=

    Fast classification of handwritten on-line Arabic characters , author=. Soft Computing and Pattern Recognition (SoCPaR), 2014 6th International Conference of , pages=. 2014 , organization=

    2014

  3. [3]

    arXiv preprint arXiv:1804.09028 , year=

    Estimate and Replace: A Novel Approach to Integrating Deep Neural Networks with Existing Applications , author=. arXiv preprint arXiv:1804.09028 , year=

    arXiv

  4. [4]

    Samuel, Jim and Ali, G. G. Md. Nawaz and Rahman, Md. Mokhlesur and Esawi, Ek and Samuel, Yana , TITLE =. Information , VOLUME =. 2020 , NUMBER =

    work page 2020

  5. [5]

    2021 , publisher=

    Garvey, Myles D and Samuel, Jim and Pelaez, Alexander , journal=. 2021 , publisher=

    work page 2021

  6. [6]

    arXiv , year=

    Bertscore: Evaluating text generation with bert , author=. arXiv , year=

    work page

  7. [7]

    Ali, G. G. Md. Nawaz and Rahman, Md. Mokhlesur and Hossain, Md. Amjad and Rahman, Md. Shahinoor and Paul, Kamal Chandra and Thill, Jean-Claude and Samuel, Jim , TITLE =. Healthcare , VOLUME =. 2021 , NUMBER =

    work page 2021

  8. [8]

    2024 , issue_date =

    Su, Jianlin and Ahmed, Murtadha and Lu, Yu and Pan, Shengfeng and Bo, Wen and Liu, Yunfeng , title =. 2024 , issue_date =. doi:10.1016/j.neucom.2023.127063 , journal =

    arXiv 2024

Show all 71 references
  1. [9]

    2024 , eprint=

    DeepStack: Deeply Stacking Visual Tokens is Surprisingly Simple and Effective for LMMs , author=. 2024 , eprint=

    2024

  2. [10]

    2025 , eprint=

    Qwen3-VL Technical Report , author=. 2025 , eprint=

    2025

  3. [11]

    2025 , eprint=

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency , author=. 2025 , eprint=

    2025

  4. [12]

    2025 , eprint=

    Seed1.5-VL Technical Report , author=. 2025 , eprint=

    2025

  5. [13]

    2026 , eprint=

    Kimi K2.5: Visual Agentic Intelligence , author=. 2026 , eprint=

    2026

  6. [14]

    2024 , eprint=

    DINOv2: Learning Robust Visual Features without Supervision , author=. 2024 , eprint=

    2024

  7. [15]

    2024 , archivePrefix=

    Analyzing to discover origins of CNNs and ViT architectures in medical images , author=. 2024 , archivePrefix=

    2024

  8. [16]

    2021 , eprint=

    Efficient Training of Visual Transformers with Small Datasets , author=. 2021 , eprint=

    2021

  9. [17]

    2023 , eprint=

    Toward a Deeper Understanding: RetNet Viewed through Convolution , author=. 2023 , eprint=

    2023

  10. [18]

    2025 , eprint=

    Comparative Analysis of Vision Transformers and Convolutional Neural Networks for Medical Image Classification , author=. 2025 , eprint=

    2025

  11. [19]

    2023 , archivePrefix=

    Convolutional Neural Network in Medical Image Analysis: A Review , author=. 2023 , archivePrefix=

    2023

  12. [20]

    2024 , issn =

    Reviewing 3D convolutional neural network approaches for medical image segmentation , journal =. 2024 , issn =

    2024

  13. [21]

    2024 , eprint=

    PaliGemma: A versatile 3B VLM for transfer , author=. 2024 , eprint=

    2024

  14. [22]

    2023 , eprint=

    LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day , author=. 2023 , eprint=

    2023

  15. [23]

    2024 , eprint=

    DeepSeek-VL: Towards Real-World Vision-Language Understanding , author=. 2024 , eprint=

    2024

  16. [24]

    2024 , eprint=

    MiniCPM-V: A GPT-4V Level MLLM on Your Phone , author=. 2024 , eprint=

    2024

  17. [25]

    2025 , eprint=

    Gemma 3 Technical Report , author=. 2025 , eprint=

    2025

  18. [26]

    2025 , eprint=

    MedGemma Technical Report , author=. 2025 , eprint=

    2025

  19. [27]

    2025 , eprint=

    Hulu-Med: A Transparent Generalist Model towards Holistic Medical Vision-Language Understanding , author=. 2025 , eprint=

    2025

  20. [28]

    2023 , eprint=

    RoFormer: Enhanced Transformer with Rotary Position Embedding , author=. 2023 , eprint=

    2023

  21. [29]

    2024 , eprint=

    A Formal Perspective on Byte-Pair Encoding , author=. 2024 , eprint=

    2024

  22. [30]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  23. [31]

    2024 , eprint=

    Qwen2 Technical Report , author=. 2024 , eprint=

    2024

  24. [32]

    2015 , eprint=

    Magnetic Resonance Imaging: Principles and Techniques: Lessons for Clinicians , author=. 2015 , eprint=

    2015

  25. [33]

    2023 , eprint=

    ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders , author=. 2023 , eprint=

    2023

  26. [34]

    2024 , eprint=

    MedNeXt: Transformer-driven Scaling of ConvNets for Medical Image Segmentation , author=. 2024 , eprint=

    2024

  27. [35]

    2023 , eprint=

    RQUGE: Reference-Free Metric for Evaluating Question Generation by Answering the Question , author=. 2023 , eprint=

    2023

  28. [36]

    2002 , publisher =

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing , title =. 2002 , publisher =. doi:10.3115/1073083.1073135 , booktitle =

    2002

  29. [37]

    2021 , eprint=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. 2021 , eprint=

    2021

  30. [38]

    2015 , eprint=

    U-Net: Convolutional Networks for Biomedical Image Segmentation , author=. 2015 , eprint=

    2015

  31. [39]

    2023 , archivePrefix=

    Multiparametric MRI , author=. 2023 , archivePrefix=

    2023

  32. [40]

    2025 , eprint=

    Contrastive Language-Image Pre-Training Model based Semantic Communication Performance Optimization , author=. 2025 , eprint=

    2025

  33. [41]

    2025 , eprint=

    Instruction Tuning for Large Language Models: A Survey , author=. 2025 , eprint=

    2025

  34. [42]

    2025 , eprint=

    Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning , author=. 2025 , eprint=

    2025

  35. [43]

    2025 , eprint=

    Demystifying Long Chain-of-Thought Reasoning in LLMs , author=. 2025 , eprint=

    2025

  36. [44]

    2026 , eprint=

    Multimodal LLM With Hierarchical Mixture-of-Experts for VQA on 3D Brain MRI , author=. 2026 , eprint=

    2026

  37. [45]

    2024 , eprint=

    3D-CT-GPT: Generating 3D Radiology Reports through Integration of Large Vision-Language Models , author=. 2024 , eprint=

    2024

  38. [46]

    2023 , eprint=

    Visual Instruction Tuning , author=. 2023 , eprint=

    2023

  39. [47]

    2025 , eprint=

    Kimi-VL Technical Report , author=. 2025 , eprint=

    2025

  40. [48]

    2026 , eprint=

    GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning , author=. 2026 , eprint=

    2026

  41. [49]

    2025 , eprint=

    Large-scale Multi-sequence Pretraining for Generalizable MRI Analysis in Versatile Clinical Applications , author=. 2025 , eprint=

    2025

  42. [50]

    2025 , eprint=

    Multi-modal Vision Pre-training for Medical Image Analysis , author=. 2025 , eprint=

    2025

  43. [51]

    2025 , eprint=

    Evaluating Causal Explanation in Medical Reports with LLM-Based and Human-Aligned Metrics , author=. 2025 , eprint=

    2025

  44. [52]

    Proceedings of the Second Workshop on Statistical Machine Translation , pages =

    Lavie, Alon and Agarwal, Abhaya , title =. Proceedings of the Second Workshop on Statistical Machine Translation , pages =. 2007 , publisher =

    2007

  45. [53]

    2015 , eprint=

    CIDEr: Consensus-based Image Description Evaluation , author=. 2015 , eprint=

    2015

  46. [54]

    2020 , eprint=

    PathVQA: 30000+ Questions for Medical Visual Question Answering , author=. 2020 , eprint=

    2020

  47. [55]

    2023 , archivePrefix=

    A visual–language foundation model for pathology image analysis using medical Twitter , author=. 2023 , archivePrefix=

    2023

  48. [56]

    2024 , eprint=

    Free Form Medical Visual Question Answering in Radiology , author=. 2024 , eprint=

    2024

  49. [57]

    2021 , eprint=

    SLAKE: A Semantically-Labeled Knowledge-Enhanced Dataset for Medical Visual Question Answering , author=. 2021 , eprint=

    2021

  50. [58]

    2025 , eprint=

    GEMeX: A Large-Scale, Groundable, and Explainable Medical VQA Benchmark for Chest X-ray Diagnosis , author=. 2025 , eprint=

    2025

  51. [59]

    2025 , eprint=

    BIOMEDICA: An Open Biomedical Image-Caption Archive, Dataset, and Vision-Language Models Derived from Scientific Literature , author=. 2025 , eprint=

    2025

  52. [60]

    2020 , eprint=

    MedICaT: A Dataset of Medical Images, Captions, and Textual References , author=. 2020 , eprint=

    2020

  53. [61]

    2024 , eprint=

    M3D: Advancing 3D Medical Image Analysis with Multi-Modal Large Language Models , author=. 2024 , eprint=

    2024

  54. [62]

    2024 , eprint=

    RadGenome-Chest CT: A Grounded Vision-Language Dataset for Chest CT Analysis , author=. 2024 , eprint=

    2024

  55. [63]

    2024 , eprint=

    BIMCV-R: A Landmark Dataset for 3D CT Text-Image Retrieval , author=. 2024 , eprint=

    2024

  56. [64]

    2025 , eprint=

    RadGPT: Constructing 3D Image-Text Tumor Datasets , author=. 2025 , eprint=

    2025

  57. [65]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  58. [66]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  59. [67]

    2023 , eprint=

    Qwen Technical Report , author=. 2023 , eprint=

    2023

  60. [68]

    2024 , eprint=

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author=. 2024 , eprint=

    2024

  61. [69]

    2021 , eprint=

    RadGraph: Extracting Clinical Entities and Relations from Radiology Reports , author=. 2021 , eprint=

    2021

  62. [70]

    2024 , eprint=

    V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding , author=. 2024 , eprint=

    2024

  63. [71]

    2026 , eprint=

    Revisiting Multimodal Positional Encoding in Vision-Language Models , author=. 2026 , eprint=

    2026

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.