Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

RARL: Improving Medical VLM Reasoning and Generalization with Reinforcement Learning and LoRA under Data and Hardware Constraints

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper argues that rewarding a small medical vision-language model for reasoned intermediate thoughts, not just correct answers, lets it beat supervised fine-tuning by about 7.78% on reasoning tasks and by roughly 27% on unseen…

desk verdict Useful resource-constrained RL recipe for medical VLMs, but the paper's central reasoning reward is never defined, making the claimed causal gains unverifiable. read the letter →

arxiv 2506.06600 v2 pith:A2CWY3QM submitted 2025-06-07 cs.CV

Tan-Hanh Pham , Chris Ngo This is my paper
classification cs.CV
keywords medicalvision-languagemodelsreinforcementlearningreasoning-awarerewardLoRAGRPOVQAlow-resourcetraininggeneralization
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 is trying to establish that a lightweight medical vision-language model can be made to reason, generalize, and stay clinically transparent without large datasets or large compute. It proposes RARL, a reasoning-aware reinforcement learning pipeline that fine-tunes Qwen2-VL-2B-Instruct with LoRA and GRPO on a single A100 GPU, scoring both the final answer and the intermediate reasoning. The reported results are about 7.78% higher than supervised fine-tuning on reasoning-focused tasks, about 27% higher on unseen datasets, and about 4% higher than traditional RL fine-tuning. If those numbers hold, resource-constrained hospitals could deploy their own explainable medical image assistants rather than depending on massive pretrained models and GPU clusters.

What carries the argument

The machinery that carries the argument is the RARL reward bundle: a format reward, a length reward, an accuracy reward, and the Reasoning Aware Reward, which together push the model to emit structured, clinically sensible thinking before a final answer. Training uses GRPO, whose group-normalized advantage replaces a value network and clips policy updates, applied only to the LoRA adapters on the language module while the vision encoder stays frozen. A second piece is prompting diversity: during training the model sees explanation-required, short-form, and open-ended prompts, and during inference a reasoning prompt steers the model to think first.

What would settle it

Ablate the reasoning-aware reward by replacing it with the plain length reward, retrain on the same 716 samples, and compare reasoning scores from the LLM judges and human evaluation; if reasoning quality does not drop, the claimed mechanism behind the 7.78% gain is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that RL can teach a small medical VLM to reason step-by-step when the reward function explicitly inspects the reasoning, not just the answer. RARL combines a format reward for <think>...</think> and <answer>...</answer> structure, a length reward, an accuracy reward using exact matching or BERTScore, and a Reasoning Aware Reward that scores the clinical coherence and informativeness of the intermediate thoughts. Training on 716 curated medical samples from Silvar-Med with GRPO and LoRA on a single 40GB A100, the model beats SFT by 7.78% on reasoning-focused test questions and generalizes to unseen benchmarks (VQA-RAD, SLAKE, VQA-Med 2019) with roughly 27% better performance than SFT and about 4% better than traditional RL. The paper also reports that prompting diversity during training and reasoning prompting during inference are necessary to get these gains.

Load-bearing premise

The load-bearing premise is that the reasoning-aware reward really measures how clinically coherent and informative the model's intermediate explanation is, and the paper never spells out how that reward is computed.

Editorial extensions

If this is right

  • A small VLM can beat supervised fine-tuning on reasoning tasks while training on a single GPU, making the approach usable by smaller clinical institutions.
  • The roughly 27% improvement on unseen datasets suggests reasoning-guided RL is a viable alternative to collecting more annotated medical data for generalization.
  • The measured drop in performance when reasoning prompting is removed at inference means the learned reasoning behavior is prompt-dependent, so deployment should keep the reasoning prompt.
  • The low-resource recipe of LoRA, a frozen vision encoder, and only 716 training samples indicates parameter-efficient RL is a practical route for domain adaptation beyond medicine.

Reading between the lines

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

  • Editorial inference: the paper never isolates the Reasoning Aware Reward in an ablation, so the cleanest testable implication is that replacing it with a pure length reward should remove most of the 7.78% reasoning gain if the mechanism is real.
  • Editorial inference: because the reward bundle is task-agnostic, the pipeline could transfer to other low-data safety-critical visual domains, such as pathology or dermatology, provided each domain supplies a few hundred reasoning-annotated examples.
  • Editorial inference: the observed gap between reasoning scores and final-answer scores suggests an LLM judge grades explanation quality and answer correctness almost independently, so a follow-up could reward only reasoning that leads to a correct final answer.
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

4 major / 6 minor

Summary. The paper proposes RARL, a framework that fine-tunes Qwen2-VL-2B-Instruct using GRPO reinforcement learning, LoRA, and custom reward functions for medical VQA and clinical reasoning. Training is performed on a single A100 GPU with a small curated dataset derived from Silvar-Med, and evaluation uses an LLM-as-judge framework plus human evaluation. The paper claims RARL outperforms supervised fine-tuning by about 7.78% on reasoning-focused tasks and improves generalization on unseen datasets by up to 27% over SFT and about 4% over traditional RL.

Significance. If the reported gains were fully supported, RARL would be a valuable contribution for resource-constrained medical imaging, demonstrating a practical path to more explainable small VLMs. The public release of code and data, the emphasis on single-GPU training, and the combination of RL with LoRA are commendable and address a real need. However, the central reasoning reward is not specified, the evaluation protocol lacks critical details, and several headline numbers are not consistent with the reported tables, so the claimed improvements are currently unverifiable.

major comments (4)
  1. [Section 2.2.2, Eq. (1)] The Reasoning Reward, which is the defining component of RARL, is described only in qualitative terms; no equation, metric, algorithm, or judge prompt is provided. Since GRPO in Eq. (1) optimizes a scalar reward r_i, the causal contribution of this reward to the reported gains cannot be assessed or reproduced. Specify the exact function (e.g., a rule-based heuristic, BERTScore against reference reasoning, or an LLM-as-judge call with the full prompt and thresholds), including any hyperparameters such as coefficient weights.
  2. [Tables 3 and 6, Section 3.2] The 'Human Evaluation' columns report reasoning and final accuracy without any protocol: number of evaluators, medical qualifications, instructions, blinding, or inter-rater reliability. The 7.78% improvement claimed in the abstract and conclusion is not interpretable without this information. Additionally, no confidence intervals or significance tests are reported for any table, so the differences between methods may be within sampling variability.
  3. [Abstract, Section 3.3, Section 4] The headline quantitative claims are inconsistent with the data in the tables. The abstract and conclusion state that RARL outperforms SFT by approximately 7.78% on reasoning-focused tasks, but the only 7.78% mentioned in Section 3.3 is the performance decrease when switching from reasoning to non-reasoning prompting, and the table values do not reproduce this number. Similarly, the claimed 27% generalization improvement and 4% gain over RL are not derivable from Table 3; for example, VQA-RAD final accuracy changes from 26.16% to 43.90% (about +68% relative) and VQA-Med from 13.80% to 49.80% (about +260% relative). Clarify the exact baselines and calculations for each claimed percentage.
  4. [Section 3.1, Table 2] The 'RL only (w/o LoRA)' baseline is taken from Reference [32] with 15,000 training samples, whereas the proposed 'RL w/ LoRA' is trained on at most 5,000 samples. This data-scale mismatch makes the comparison in Table 2 unfair and does not support the claim that RL with LoRA performs 'competitively' with RL without LoRA. Re-run the RL-only baseline under the same training data size, prompts, and evaluation protocol, or clearly report it as a literature comparison rather than an experimental baseline.
minor comments (6)
  1. [Table 4] The table contains the typo 'Grounth truth' and should read 'Ground truth'.
  2. [Appendix A] The Open-Ended Answer Prompt contains 'askes', which should be 'asks'.
  3. [Section 3.3] The sentence 'encourages a the model get some certain knowledge' is ungrammatical and unclear; please revise.
  4. [Section 3.2] The phrase 'the other datasets do not content explanation' should be 'do not contain explanations'.
  5. [Section 3.3] The statement 'the model increase by around 4.5%' does not specify the baseline or which metric is being compared.
  6. [Appendix B] The LLM-as-judge prompt requests binary reasoning and prediction scores, but the paper reports percentages; explain how continuous scores are obtained or averaged.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the reported gains are measured against external ground-truth labels and held-out/benchmark data, not read off from fitted parameters; the undefined Reasoning Reward is a reproducibility gap, not a demonstrated circularity.

full rationale

The paper's claimed derivation chain is empirical, not formal: GRPO (Eqs. 1-3) optimizes scalar rewards, and the rewards that are actually specified are explicit functions of the model output and ground truth — Format Reward (1.0/0.1/0), Length Reward (min(0.001 × token count, 1.0)), and Accuracy Reward (binary or BERTScore F1). The headline improvements are computed on a held-out test split of Silvar-Med and on external benchmarks (VQA-RAD, SLAKE, VQA-Med 2019, Path-VQA) against ground-truth answers, so no reported 'prediction' reduces to a training objective by construction. The only self-referential element is the Silvar-Med dataset (Ref. [20]), which shares authors with this paper and supplies both the training data and the in-distribution test split; however, that dataset is a published, physician-annotated resource with a separate held-out split, and the generalization claims also rest on four external benchmarks, so the citation is not load-bearing. The paper's own Limitation section concedes 'LLM-as-judge bias,' and the 'Reasoning Reward' in Section 2.2.2 is described only in prose ('evaluates the clinical coherence and informativeness of these intermediate steps') without an equation, metric, or judge prompt. This is a serious reproducibility gap that prevents a full check of the reward's causal contribution, and if that reward were secretly the same judge used at evaluation the reasoning gains would be partly by construction; but the paper provides no such reduction, and the no-speculation rule prevents treating the gap as demonstrated circularity. Separately, the abstract's 'approximately 7.78%' outperformance appears to be taken from the reasoning-prompt-to-non-reasoning-prompt gap in Section 3.3 rather than from the SFT comparison in Table 3, but that is an attribution/correctness issue, not a circularity. Overall, no circular step can be exhibited with a quotation and a specific reduction, so the circularity score is 0.

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

The central claim rests on several unstated design choices: the LoRA configuration, the length reward scale, and the unspecified reasoning reward. The largest burden is the undefined reasoning reward, as the paper's main novelty and claimed advantage depend on it. The domain assumptions about annotation quality and LLM-as-judge reliability are also load-bearing because they are not validated.

free parameters (5)
  • LoRA rank r = 8
    Set in Section 2.3; not tuned, but central to the capacity of the adapter.
  • LoRA alpha = 16
    Set in Section 2.3; standard scaling factor for LoRA.
  • Length reward coefficient and cap = 0.001 per token, cap 1.0
    Defined in Section 2.2.2; arbitrary scale that incentivizes longer outputs and may cause reward hacking.
  • GRPO hyperparameters beta, epsilon, group size G = Not reported
    Eq. (1) and (2) require these values for exact reproduction; the paper omits them.
  • Reasoning reward implementation = Undefined
    Section 2.2.2 describes a reasoning reward but gives no equation, metric, or scoring rule.
assumptions (5)
  • domain assumption Silvar-Med ground-truth reasoning annotations are clinically correct and reliable.
    The curated training and test sets come from Silvar-Med [20] with annotations by three physicians; no independent clinical verification is provided.
  • domain assumption GPT-4o mini and Gemini 1.5 Flash provide valid semantic scoring for medical reasoning and final answers.
    LLM-as-judge is used for all main metrics (Section 3.2, Appendix B); no validation against a clinician gold standard is shown.
  • domain assumption Frozen vision encoder plus LoRA on language modules is sufficient for medical VQA adaptation.
    Section 2.3 freezes the vision encoder; the paper does not compare to unfrozen or full fine-tuning.
  • domain assumption Human evaluation is conducted by qualified medical evaluators following a fixed protocol.
    Tables 3 and 6 report 'Human Evaluation' numbers but the reviewer, evaluator qualifications, and instructions are never described.
  • standard math GRPO objective (Eq. 1) and KL estimate (Eq. 3) are standard and correctly implemented.
    Uses standard GRPO from DeepSeekMath [28]; no code is provided to check implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RARL: Improving Medical VLM Reasoning and Generalization with Reinforcement Learning and LoRA under Data and Hardware Constraints." pith.science (2026). https://pith.science/paper/A2CWY3QM

@misc{pith2026250606600,
  author       = {Pith},
  title        = {Pith review of: RARL: Improving Medical VLM Reasoning and Generalization with Reinforcement Learning and LoRA under Data and Hardware Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A2CWY3QM}},
  note         = {Machine review of arXiv:2506.06600}
}
read the original abstract

The growing integration of vision-language models (VLMs) in medical applications offers promising support for diagnostic reasoning. However, current medical VLMs often face limitations in generalization, transparency, and computational efficiency-barriers that hinder deployment in real-world, resource-constrained settings. To address these challenges, we propose a Reasoning-Aware Reinforcement Learning framework, \textbf{RARL}, that enhances the reasoning capabilities of medical VLMs while remaining efficient and adaptable to low-resource environments. Our approach fine-tunes a lightweight base model, Qwen2-VL-2B-Instruct, using Low-Rank Adaptation and custom reward functions that jointly consider diagnostic accuracy and reasoning quality. Training is performed on a single NVIDIA A100-PCIE-40GB GPU, demonstrating the feasibility of deploying such models in constrained environments. We evaluate the model using an LLM-as-judge framework that scores both correctness and explanation quality. Experimental results show that RARL significantly improves VLM performance in medical image analysis and clinical reasoning, outperforming supervised fine-tuning on reasoning-focused tasks by approximately 7.78%, while requiring fewer computational resources. Additionally, we demonstrate the generalization capabilities of our approach on unseen datasets, achieving around 27% improved performance compared to supervised fine-tuning and about 4% over traditional RL fine-tuning. Our experiments also illustrate that diversity prompting during training and reasoning prompting during inference are crucial for enhancing VLM performance. Our findings highlight the potential of reasoning-guided learning and reasoning prompting to steer medical VLMs toward more transparent, accurate, and resource-efficient clinical decision-making. Code and data are publicly available.

Figures

Figures reproduced from arXiv: 2506.06600 by the authors.

Figure 1
Figure 1. Overview of RARL, which incentivizes both clinically faithful intermediate reasoning and accurate final answers. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Improving Heart-Focused Medical Question Answering in LLMs via Variance-Aware Rubric Rewards with GRPO

    cs.CL 2026-04 conditional novelty 5.0 of 10

    Variance-aware continuous rubric rewards under GRPO lift Qwen3-14B heart-focused HealthBench accuracy from 0.362 to 0.502, competitive with GPT-OSS-120B.

Reference graph

Works this paper leans on

33 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [32]

    aha moment

    Hengguang Zhou, Xirui Li, Ruochen Wang, Minhao Cheng, Tianyi Zhou, and Cho-Jui Hsieh. R1-zero’s" aha moment" in visual reasoning on a 2b non-sft model. arXiv preprint arXiv:2503.05132, 2025. 11 A Prompts for Training VLMs Explanation-Required Answer Prompt: prompt = f"You are a Medical Assistant. Carefully analyze the medical image and the User’s question...

    arXiv 2025

  2. [1]

    Language models are few-shot learners

    Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020

    arXiv 2005

  3. [2]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 9

    arXiv 2023

  4. [3]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    arXiv 2023

  5. [4]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    arXiv 2024

  6. [5]

    Qwen3, April 2025

    Qwen Team. Qwen3, April 2025. URL https://qwenlm.github.io/blog/qwen3/

    work page 2025

  7. [6]

    Visual-language models for medical image analysis: A survey

    Zhihong Huang, Lei Jin, and Hongbo Zhang. Visual-language models for medical image analysis: A survey. IEEE Transactions on Medical Imaging, 42(10):2789–2804, 2023

    work page 2023

  8. [7]

    Med-flamingo: a multimodal medical few-shot learner

    Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. In Machine Learning for Health (ML4H), pages 353–367. PMLR, 2023

    2023

Show all 33 references
  1. [8]

    Towards generalist biomedical ai with med-palm m

    Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Chuck Lau, Ryutaro Tanno, et al. Towards generalist biomedical ai with med-palm m. arXiv preprint arXiv:2402.16994, 2024

    2024 arXiv

  2. [9]

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

    Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023

    2023 arXiv

  3. [10]

    Biomedgpt: A unified and generalist biomedical generative pre-trained transformer for vision, language, and multimodal tasks

    Kai Zhang, Jun Yu, Eashan Adhikarla, Rong Zhou, Zhiling Yan, Yixin Liu, Zhengliang Liu, Lifang He, Brian Davison, Xiang Li, et al. Biomedgpt: A unified and generalist biomedical generative pre-trained transformer for vision, language, and multimodal tasks. arXiv e-prints, page...

    2023

  4. [11]

    Llava-med: Training a large language-and-vision assistant for biomedicine in one day

    Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36: 28541–2...

    2023

  5. [12]

    Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports

    Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019

    2019

  6. [13]

    Pmc-15m: A large-scale dataset for medical vision-language pretraining

    Xin Zhang, Chunyuan Li, and Haotian Liu. Pmc-15m: A large-scale dataset for medical vision-language pretraining. arXiv preprint arXiv:2305.07890, 2023

    2023 arXiv

  7. [14]

    Artificial intelligence in radiology

    Ahmed Hosny, Chintan Parmar, John Quackenbush, Lawrence H Schwartz, and Hugo JWL Aerts. Artificial intelligence in radiology. Nature Reviews Cancer, 18(8):500–510, 2018

    2018

  8. [15]

    Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning

    Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Hongwei Bran Li, Chen Chen, Cheng Ouyang, and Daniel Rueckert. Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning. arXiv preprint arXiv:2502.19634, 2025

    2025 arXiv

  9. [16]

    Med-r1: Reinforce- ment learning for generalizable medical reasoning in vision-language models

    Yuxiang Lai, Jike Zhong, Ming Li, Shitian Zhao, and Xiaofeng Yang. Med-r1: Reinforce- ment learning for generalizable medical reasoning in vision-language models. arXiv preprint arXiv:2503.13939, 2025

    2025

  10. [17]

    Explainability for artificial intelligence in healthcare: a multidisciplinary perspective

    Julia Amann, Alessandro Blasimme, Effy Vayena, Dietmar Frey, Vince I Madai, and Pre- cise4Q Consortium. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC medical informatics and decision making, 20:1–9, 2020

    2020

  11. [18]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824–24837, 2022. 10

    2022

  12. [19]

    Medcot: Medical chain of thought via hierarchical expert

    Jiaxiang Liu, Yuan Wang, Jiawei Du, Joey Tianyi Zhou, and Zuozhu Liu. Medcot: Medical chain of thought via hierarchical expert. arXiv preprint arXiv:2412.13736, 2024

    2024 arXiv

  13. [20]

    Silvar-med: A speech-driven visual language model for explainable abnormality detection in medical imaging

    Tan-Hanh Pham, Trong-Duong Bui, Minh Luu Quang, Tan Huong Pham, Chris Ngo, and Truong Son Hy. Silvar-med: A speech-driven visual language model for explainable abnormality detection in medical imaging. In Proceedings of the Computer Vision and Pattern Recognition Conference, p...

    2025

  14. [21]

    Deepseek-rl: Advancing reasoning in large language models with reinforcement learning

    Jun Yao, Ziming Wang, and Wei Chen. Deepseek-rl: Advancing reasoning in large language models with reinforcement learning. arXiv preprint arXiv:2401.08912, 2024

    2024 arXiv

  15. [22]

    Reinforcement learning for medical image analysis: Current progress and future directions

    Xiaoxiao Chen, Yufan Wang, and Li Zhang. Reinforcement learning for medical image analysis: Current progress and future directions. Medical Image Analysis, 92:103056, 2024

    2024

  16. [23]

    Generalized reward-driven policy optimization for vision-language models

    Wei Liu, Kai Zhang, and Ming Zhao. Generalized reward-driven policy optimization for vision-language models. arXiv preprint arXiv:2403.09876, 2024

    2024 arXiv

  17. [24]

    Huatuogpt-vision, towards injecting medical visual knowledge into multimodal llms at scale

    Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Ruifei Zhang, Zhenyang Cai, Ke Ji, et al. Huatuogpt-vision, towards injecting medical visual knowledge into multimodal llms at scale. arXiv preprint arXiv:2406.19280, 2024

    2024 arXiv

  18. [25]

    A dataset of clinically generated visual questions and answers about radiology images

    Jason J Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific data, 5(1): 1–10, 2018

    2018

  19. [26]

    Slake: A semantically- labeled knowledge-enhanced dataset for medical visual question answering

    Bo Liu, Li-Ming Zhan, Li Xu, Lin Qian, Jiahong Wu, and Wenbin Shi. Slake: A semantically- labeled knowledge-enhanced dataset for medical visual question answering. arXiv preprint arXiv:2102.09542, 2021

    2021 arXiv

  20. [27]

    Hasan, Vivek V

    Asma Ben Abacha, Sadid A. Hasan, Vivek V . Datla, Joey Liu, Dina Demner-Fushman, and Henning Müller. Vqa-med: Overview of the medical visual question answering task at imageclef 2019. In Working Notes of CLEF 2019 , volume 2380 of CEUR Workshop Proceedings, Lugano, Switzerland...

    2019

  21. [28]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  22. [29]

    Weinberger, and Yoav Artzi

    Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations,

  23. [30]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344–16359, 2022

    2022

  24. [31]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms

    Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processi...

    2024

  25. [2020]

    URL https://openreview.net/forum?id=SkeHuCVFDr

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.