Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Image-plus-text reasoning beats text-only for chest X-ray reports

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 23:39 UTC pith:FBBFDFQJ

load-bearing objection A genuinely new interleaved multimodal CXR dataset and a clean ablation, but the headline 6% RadGraph gain rests on a single 1097-sample split with no error bars, and the data isn't released. the 4 major comments →

arxiv 2602.12843 v2 pith:FBBFDFQJ submitted 2026-02-13 cs.CV

MMRad-22K: A Structured Multimodal Evidence Dataset for Chest X-ray Report Generation

Yichen Zhao , Zelin Peng , Fenghe Tang , Piao Yang , Yu Huang , Wei Shen This is my paper
classification cs.CV
keywords chest X-ray report generationmultimodal chain-of-thoughtinterleaved vision-language reasoningmedical datasetradiology groundingAnole-RadCoTRadGraphMIMIC-CXR
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces MMRad-IVL-22K (also called MMRad-22K), a dataset of 21,994 chest X-ray cases that transforms report generation into an interleaved process: each textual observation about an anatomical region is paired with a cropped image of that region and a bounding box. The central claim is that report generation guided by this multimodal chain-of-thought, where a model re-examines local image evidence at each reasoning step, is more clinically accurate than generation guided by text-only reasoning, with about a 6% improvement on the RadGraph clinical metric. If correct, this means medical AI systems should be built to revisit visual evidence during reasoning rather than inspect the image once and then reason purely in language. The paper also fine-tunes a unified vision-language model, Anole-RadCoT, on this dataset and reports that it outperforms several open-source general and medical LVLMs on report quality and reasoning-consistency metrics.

Core claim

On its own terms, the paper argues that bounding-box coordinates are a 'pseudo-visual' solution because they remain text; what a model actually needs is native access to the localized image region during each reasoning step. The dataset provides this by cropping region images and interleaving them with textual rationales and coordinates, and controlled experiments show that advanced closed-source LVLMs improve when given these image crops. A model fine-tuned on the dataset, Anole-RadCoT, learns to emit its own interleaved text-and-image traces, achieving high generation, grounding, and semantic consistency, and its final reports surpass those of existing baselines. The paper's conclusion is

What carries the argument

The central object is the interleaved reasoning trace: a multi-step chain in which each step contains a textual thought, a bounding box, and a localized image crop from the original X-ray, with steps aggregated into four anatomical domains (lung, mediastinal, bone, other) spanning 35 fine-grained sub-regions. This structure lets a unified LVLM alternate between generating text and generating or consuming image tokens, grounding each diagnostic statement in a specific visual region rather than in abstract coordinates.

Load-bearing premise

The entire benefit rests on the premise that the LLM-generated and machine-audited reasoning traces accurately reflect how radiologists inspect regions and what the true regional findings are; systematic errors in the traces would make the multimodal advantage an artifact of the generation pipeline.

What would settle it

Train or prompt a model on a version of MMRad-IVL-22K in which each localized image crop is replaced by a random crop from the same X-ray (or from a different patient) while keeping the same text; if the multimodal advantage over text-only CoT persists, the visual content is not doing the work. Alternatively, have a panel of independent radiologists mark region-level findings on a sample of the audited cases and measure agreement with the dataset's traces.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If multimodal CoT is genuinely superior, future medical report generators should be designed as unified multimodal models that can emit and attend to image crops during reasoning, rather than text-only reasoners.
  • The dataset offers a benchmark for evaluating grounding and reasoning consistency, not just final report text, enabling comparisons of whether a system's rationales match the visual evidence.
  • The reported 6% RadGraph gain suggests that clinical entity extraction benefits from visual grounding, potentially reducing hallucinations in automated radiology reports.
  • A 7B-parameter model trained on this data reaches competitive report quality, implying that region-level interleaved supervision is an efficient route to improving medical LVLMs without scaling model size.
  • The construction pipeline—summarize, audit, aggregate—provides a reusable recipe for building interleaved datasets for other imaging modalities or clinical workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The advantage of multimodal CoT may partly stem from forcing the model to attend to smaller, higher-resolution crops rather than the full image; a testable extension would isolate the effect of interleaving by comparing against a full-image multi-crop attention baseline.
  • Because the traces are synthetically generated and audited by another LLM plus one human expert, the 'reasoning' may reflect the annotator model's style rather than actual radiologist cognition; collecting eye-tracking or verbal-protocol data from radiologists would validate the trace structure.
  • The consistency metrics introduced here—generation, grounding, and semantic consistency—could become a general protocol for evaluating any interleaved vision-language medical system, not just report generation.
  • The claimed non-substitutability of visual evidence could be stress-tested by degrading the localized crops (e.g., blurring or mis-cropping) and measuring whether the performance gain disappears proportionally.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces MMRad-IVL-22K, a dataset of 21,994 chest X-ray diagnostic traces derived from MIMIC-CXR, with each trace containing region-level textual observations, bounding boxes, localized image crops, and a final report. The construction pipeline summarizes GEMeX-ThinkVG groundings via DeepSeek-v3, applies multi-stage verification (self-reflection, cross-audit by Qwen2.5-72B, human adjudication), then aggregates the traces into anatomical groups. The authors claim this is the first large-scale dataset for natively interleaved vision-language reasoning in CXR interpretation. They report two experiments: (1) a controlled comparison showing that closed-source LVLMs generate better reports when given multimodal (text + image) CoT evidence than text-only CoT evidence, with a claimed ~6% relative RadGraph improvement; and (2) a benchmark where Anole-RadCoT, a model fine-tuned on MMRad-IVL-22K, outperforms several open-source LVLMs on NLG, clinical, and consistency metrics. The paper concludes that multimodal CoT is a 'non-substitutable component' of reliable medical AI.

Significance. If the central empirical claim withstands scrutiny, the dataset and the formulation of interleaved radiologist-style reasoning would be a useful resource for training and evaluating report-generation LVLMs. The controlled text-only vs multimodal comparison is a reasonable design for isolating the contribution of visual evidence. The authors also deserve credit for the multi-stage verification pipeline and for attempting to ground each reasoning step with a bounding box and localized image, which is clinically motivated. However, the primary quantitative evidence rests on a small single split with no statistical inference, and the synthetic origin of the traces raises generalizability concerns. As reported, the conclusion that multimodal evidence is 'non-substitutable' is not established by the data.

major comments (4)
  1. [Sec. 3.1, Table 2] The central claim that multimodal CoT 'significantly outperforms' text-only CoT is not supported by the reported statistics. All results are on a single 5% stratified subset (n=1,097) with no confidence intervals, significance tests, or multiple seeds. The absolute RGe gains are small (GPT-5: 0.316→0.333; Qwen3-VL-Plus: 0.306→0.320), and RadGraph F1 estimates on ~1,100 reports are noisy; these gaps may be within sampling variation. The stratified split is by reasoning step, not by patient, so MIMIC-CXR studies from the same patient may appear in both conditions, potentially inflating consistency. Please provide paired tests, bootstrap CIs, multiple evaluation seeds, and a patient-level split.
  2. [Sec. 2.1, Fig. 1] The dataset's reasoning traces are generated by DeepSeek-v3 summarizing GEMeX-ThinkVG, then audited by Qwen2.5-72B and only 500 cases by one human expert. This means the 'multimodal evidence' is synthetic, and the downstream advantage of multimodal CoT (Table 2) may partly reflect that the same generation pipeline defines both the evidence and the evaluation prompt. Please report inter-rater agreement for the 500-case audit, the number of human-adjudicated revisions, and test whether multimodal CoT remains superior when the reasoning traces are produced by an independent annotation protocol or when reports are generated from raw images without CoT conditioning.
  3. [Sec. 3.2, Table 3] The benchmarking is not controlled. Anole-RadCoT is fine-tuned on 95% of MMRad-IVL-22K and evaluated on the same distribution's 5% test split, while the other LVLMs are zero-shot; MiniGPT-Med is even marked as having test-set overlap. The consistency metrics (Gen., Sem., Grd.) are computed only for unified models, so the comparison does not support the conclusion that Anole-RadCoT has 'superior reasoning consistency and report quality' beyond that expected from in-distribution fine-tuning. Please add a fair comparison protocol (e.g., fine-tune baselines on the same training set, or evaluate zero-shot for all models) and report the consistency-metric definition and computation details.
  4. [Sec. 3.2, Consistency Metrics] 'Grounding Consistency' is defined as 'Mean Intersection over Union (mIoU) between the images extracted via predicted bounding boxes and the ground truth localized images.' IoU is defined for boxes, not images; if you mean IoU of the bounding boxes, say so, and if you compare image crops, specify the visual similarity measure and normalization. Similarly, 'Semantic Consistency' is defined as the sum of two BiomedCLIP similarities, and reported values above 1 (e.g., 1.226) need a range and per-component breakdown to be interpretable.
minor comments (6)
  1. [Title/Abstract] Inconsistent naming: the abstract and title use 'MMRad-22K' while the body uses 'MMRad-IVL-22K' throughout. Please unify.
  2. [Sec. 3.1] '6% increase in the RadGraph metric' should state relative improvement and distinguish which RadGraph variant (RGe, RGer, RGber) is being referenced.
  3. [Fig. 1] Figure 1 is difficult to read; labels and boxes are tiny, and the relationship between 'Reasoning Image 1/2' and the final dataset format is unclear.
  4. [Sec. 3.2] Typo: 'understaning-only' should be 'understanding-only'.
  5. [References] Reference [10] supports the PACS turnaround claim, but the cited study concerns pneumothorax AI performance; please provide a direct citation for median report turnaround time.
  6. [Reproducibility] The project page does not appear to include downloadable data; dataset and evaluation code should be released for reproducibility. Also report the random seed used for the 5% split.

Circularity Check

0 steps flagged

No circularity: the multimodal-vs-text CoT comparison is an empirical result anchored to MIMIC-CXR ground-truth reports, not a consequence of dataset construction.

full rationale

The paper's central claims are (1) building MMRad-IVL-22K from GEMeX-ThinkVG traces, and (2) showing that adding regional image crops to text rationales improves report generation. Neither reduces to its inputs by construction. The controlled experiment in Sec. 3.1 compares text-only (full image + regional text) vs multimodal (full image + regional text + regional vision) on the same 5% stratified subset, and report quality is measured with standard NLG and RadGraph metrics against the MIMIC-CXR ground-truth reports, which are external to the reasoning-trace generation. The 6% RadGraph gain is an empirical observation, not a fitted parameter renamed as a prediction. For Anole-RadCoT, training on 95% of the dataset and evaluating on a held-out 5% split is standard; consistency metrics compare against the dataset's own annotation schema, which is legitimate dataset construction rather than self-definitional reasoning. The foundational dataset GEMeX-ThinkVG [17] and the Anole-Zebra-CoT [14] base model are external works with no author overlap; there is no load-bearing self-citation or imported uniqueness theorem. The main weakness—lack of error bars on the 6% claim—is a statistical-support concern, not a circularity concern. The derivation chain is self-contained with respect to the paper's stated assumptions.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The central claims rest on the quality of LLM-synthesized reasoning traces and on standard radiology-report metrics. No new physical entities are introduced. Several fine-tuning hyperparameters are fixed without sensitivity analysis, so the reported model performance is partly a function of these choices.

free parameters (5)
  • LoRA rank r = 16
    Chosen for Anole-RadCoT fine-tuning; no sensitivity analysis shown, so reported results may depend on this choice.
  • LoRA scaling factor alpha = 32
    Hyperparameter for fine-tuning; listed in Sec. 3.2 without ablations.
  • Learning rate = 1e-5
    Fine-tuning hyperparameter; not explored.
  • Training steps = 50,000
    Fine-tuning budget; no early-stopping or ablation details.
  • Batch size = 2
    Fine-tuning batch size; may affect convergence.
axioms (3)
  • domain assumption MIMIC-CXR ground-truth reports are an accurate reference for radiological findings.
    All NLG and CE metrics are computed against these ground-truth reports; if they are noisy, the measured improvements inherit that noise.
  • domain assumption GEMeX-ThinkVG reasoning traces and bounding boxes are sufficiently reliable to serve as the foundation for MMRad-IVL-22K.
    Section 2.1 directly builds the new dataset by summarizing GEMeX-ThinkVG; any systematic grounding errors propagate.
  • domain assumption The Likert-scale validation by one human expert and Gemini-3.0-pro is a valid measure of clinical fidelity.
    Section 2.1 reports high agreement between the two evaluators, but the AI evaluator is not a clinician, so this does not independently establish clinical truth.

pith-pipeline@v1.3.0-alltime-deepseek · 9438 in / 10629 out tokens · 91545 ms · 2026-08-02T23:39:50.459083+00:00 · methodology

0 comments
read the original abstract

Chest X-ray (CXR) reporting follows a region-based clinical workflow in which radiologists inspect anatomical regions and integrate localized findings into a final report. However, existing resources for CXR report generation provide these supervision signals in fragmented forms. We introduce MMRad-22K, a dataset that organizes regional textual observations, anatomical grounding coordinates, localized image evidence, and report targets into structured multimodal evidence units for CXR report generation. To motivate this formulation, we first compare different evidence formats for report generation and find that structured multimodal evidence is generally more useful than text-only or bounding box-based evidence. We then adapt a unified LVLM backbone using MMRad-22K and show that adaptation with multimodal evidence outperforms both textual-evidence adaptation and end-to-end adaptation on language and clinically oriented metrics. Under the same evaluation protocol, the adapted model also reaches a performance level comparable to several open-source LVLM references. Together, these results support MMRad-22K as a practical structured multimodal resource for training and evaluating CXR report generation aligned with clinical reading workflows.

Figures

Figures reproduced from arXiv: 2602.12843 by Fenghe Tang, Piao Yang, Wei Shen, Yichen Zhao, Yu Huang, Zelin Peng.

Figure 1
Figure 1. Figure 1: Pipeline of MMRad-IVL-22K. (1) Initial interleaved reasoning traces and region-text correspondences are generated and summarized based on GEMeX-ThinkVG via DeepSeek-v3. (2) Multi-stage verification is conducted through DeepSeek-v3 self￾reflection, Qwen2.5-72B cross-auditing, and domain expert adjudication to ensure clinical fidelity. (3) Refined rationales and bounding boxes are aggregated into ma￾jor anat… view at source ↗
Figure 2
Figure 2. Figure 2: Statistical overview of MMRad-IVL-22K. (a-b) Distribution of reason￾ing steps and rationale lengths. (c) Fine-grained anatomical structures. (d-e) Clinical characteristics extracted across the dataset. v3 [16] refines the initial reasoning traces via self-reflection. Since the GEMeX￾ThinkVG [17] dataset is derived from MIMIC-CXR [13], it leverages the avail￾able ground truth reports to identify and rectify… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative analysis of the interleaved reasoning process and final re [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

31 extracted references · 15 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2407.04106 (2024)

    Alkhaldi,A.,Alnajim,R.,Alabdullatef,L.,Alyahya,R.,Chen,J.,Zhu,D.,Alsinan, A., Elhoseiny, M.: Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106 (2024)

    Pith/arXiv arXiv 2024

  2. [2]

    Anthropic: Introducing claude haiku 4.5 (2025),https://www.anthropic.com/ news/claude-haiku-4-5

    2025

  3. [3]

    Advances in Neural Information Processing Systems36, 3867–3880 (2023)

    Bae, S., Kyung, D., Ryu, J., Cho, E., Lee, G., Kweon, S., Oh, J., Ji, L., Chang, E., Kim, T., et al.: Ehrxqa: A multi-modal question answering dataset for electronic health records with chest x-ray images. Advances in Neural Information Processing Systems36, 3867–3880 (2023)

    2023

  4. [4]

    5-vl technical report

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

    Pith/arXiv arXiv 2025

  5. [5]

    arXiv preprint arXiv:2505.22525 (2025)

    Chern, E., Hu, Z., Chern, S., Kou, S., Su, J., Ma, Y., Deng, Z., Liu, P.: Thinking with generated images. arXiv preprint arXiv:2505.22525 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    arXiv preprint arXiv:2505.14683 (2025)

    Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., et al.: Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    google/products-and-platforms/products/gemini/gemini-3-flash

    Google: Gemini 3 flash: frontier intelligence built for speed (2025),https://blog. google/products-and-platforms/products/gemini/gemini-3-flash

    2025

  8. [8]

    Google: Gemini 3 pro: the frontier of vision ai (2025),https://blog.google/ innovation-and-ai/technology/developers-tools/gemini-3-pro-vision/

    2025

  9. [9]

    ICLR1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

    2022

  10. [10]

    Academic radiology32(3), 1165–1174 (2025)

    Hunter, J.G., Bera, K., Shah, N., Bukhari, S.M.A., Marshall, C., Caovan, D., Rosipko, B., Gupta, A.: Real-world performance of pneumothorax-detecting artifi- cial intelligence algorithm and its impact on radiologist reporting times. Academic radiology32(3), 1165–1174 (2025)

    2025

  11. [11]

    arXiv preprint arXiv:2106.14463 (2021)

    Jain, S., Agrawal, A., Saporta, A., Truong, S.Q., Duong, D.N., Bui, T., Chambon, P., Zhang, Y., Lungren, M.P., Ng, A.Y., et al.: Radgraph: Extracting clinical enti- ties and relations from radiology reports. arXiv preprint arXiv:2106.14463 (2021)

    Pith/arXiv arXiv 2021

  12. [12]

    arXiv preprint arXiv:2504.18453 (2025)

    Jing,P.,Lee,K.,Zhang,Z.,Zhou,H.,Yuan,Z.,Gao,Z.,Zhu,L.,Papanastasiou,G., Fang, Y., Yang, G.: Reason like a radiologist: Chain-of-thought and reinforcement learning for verifiable report generation. arXiv preprint arXiv:2504.18453 (2025)

    arXiv 2025

  13. [13]

    arXiv preprint arXiv:1901.07042 (2019)

    Johnson, A.E., Pollard, T.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Peng, Y., Lu, Z., Mark, R.G., Berkowitz, S.J., Horng, S.: Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042 (2019)

    Pith/arXiv arXiv 1901

  14. [14]

    arXiv preprint arXiv:2507.16746 (2025)

    Li, A., Wang, C., Fu, D., Yue, K., Cai, Z., Zhu, W.B., Liu, O., Guo, P., Neiswanger, W., Huang, F., et al.: Zebra-cot: A dataset for interleaved vision language reason- ing. arXiv preprint arXiv:2507.16746 (2025)

    arXiv 2025

  15. [15]

    Advances in Neural Information Processing Systems36, 28541–28564 (2023)

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems36, 28541–28564 (2023)

    2023

  16. [16]

    arXiv preprint arXiv:2412.19437 (2024) Thinking Like a Radiologist 11

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al.: Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024) Thinking Like a Radiologist 11

    Pith/arXiv arXiv 2024

  17. [17]

    arXiv preprint arXiv:2506.17939 (2025)

    Liu, B., Zhao, X., He, A., Chen, Y., Fu, H., Wu, X.M.: Gemex-thinkvg: Towards thinking with visual grounding in medical vqa via reinforcement learning. arXiv preprint arXiv:2506.17939 (2025)

    arXiv 2025

  18. [18]

    In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision

    Liu,B.,Zou,K.,Zhan,L.M.,Lu,Z.,Dong,X.,Chen,Y.,Xie,C.,Cao,J.,Wu,X.M., Fu, H.: Gemex: A large-scale, groundable, and explainable medical vqa benchmark for chest x-ray diagnosis. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 21310–21320 (2025)

    2025

  19. [19]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Liu, K., Ma, Z., Kang, X., Li, Y., Xie, K., Jiao, Z., Miao, Q.: Enhanced con- trastive learning with multi-view longitudinal data for chest x-ray report genera- tion. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10348–10359 (2025)

    2025

  20. [20]

    Informatics in Medicine Unlocked50, 101585 (2024)

    Nicolson, A., Dowling, J., Anderson, D., Koopman, B.: Longitudinal data and a se- mantic similarity reward for chest x-ray report generation. Informatics in Medicine Unlocked50, 101585 (2024)

    2024

  21. [21]

    arXiv preprint arXiv:2510.15710 (2025)

    Ning,J.,Li,W.,Tang,C.,Lin,J.,Ma,C.,Zhang,C.,Liu,J.,Chen,Y.,Gao,S.,Liu, L., et al.: Unimedvl: Unifying medical multimodal understanding and generation through observation-knowledge-analysis. arXiv preprint arXiv:2510.15710 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    Nursing Times120(6), 38–43 (2024)

    North, Z.: How to interpret chest radiographs (x-rays): a systematic approach. Nursing Times120(6), 38–43 (2024)

    2024

  23. [23]

    OpenAI: Introducing gpt-5 (2025),https://openai.com/index/ introducing-gpt-5

    2025

  24. [24]

    arXiv preprint arXiv:2005.03642 (2020)

    Wang, C., Sennrich, R.: On exposure bias, hallucination and domain shift in neural machine translation. arXiv preprint arXiv:2005.03642 (2020)

    Pith/arXiv arXiv 2005

  25. [25]

    arXiv preprint arXiv:2504.02885 (2025)

    Wang, H., Ye, S., Lin, J., Naseem, U., Kim, J.: Lvmed-r2: Perception and reflection-driven complex reasoning for medical report generation. arXiv preprint arXiv:2504.02885 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    arXiv preprint arXiv:2512.16145 (2025)

    Wang, P., Ye, S., Naseem, U., Kim, J.: Mrg-r1: Reinforcement learning for clinically aligned medical report generation. arXiv preprint arXiv:2512.16145 (2025)

    arXiv 2025

  27. [27]

    5: Advancing open-source multimodal models in versatility, reasoning, and efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    Pith/arXiv arXiv 2025

  28. [28]

    arXiv preprint arXiv:2412.10302 (2024)

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

    Pith/arXiv arXiv 2024

  29. [29]

    arXiv preprint arXiv:2505.09388 (2025)

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

    Pith/arXiv arXiv 2025

  30. [30]

    arXiv preprint arXiv:2303.00915 (2023)

    Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., et al.: Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915 (2023)

    Pith/arXiv arXiv 2023

  31. [31]

    arXiv preprint arXiv:2404.16754 (2024)

    Zhang, X., Wu, C., Zhao, Z., Lei, J., Zhang, Y., Wang, Y., Xie, W.: Radgenome- chest ct: A grounded vision-language dataset for chest ct analysis. arXiv preprint arXiv:2404.16754 (2024)

    Pith/arXiv arXiv 2024