Pith. sign in

REVIEW 5 major objections 5 minor 29 references

EQPO: Equitable Group Relative Policy Optimization for Clinical Reasoning

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A fairness-aware RL scaling rule, applied to the reward advantage in GRPO, is claimed to let clinical vision-language models shrink demographic diagnostic gaps while preserving or improving overall accuracy.

desk verdict Abstract and body describe two different papers—EQPO vs FairGRPO—and the table that tests the no-label variant contradicts its fairness claim; the core result as stated is unsupported. read the letter →

arxiv 2510.19893 v2 pith:DBVVU72G submitted 2025-10-22 cs.LG

Shiqi Dai , Wei Dai , Jiaee Cheong , Paul Pu Liang This is my paper · ORCID
classification cs.LG
keywords fairreinforcementlearningclinicalvision-languagemodelsdemographicbiasgroup-relativepolicyoptimizationrewardscalingunsupervisedclusteringmedicalimagingpredictiveparity
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

This paper tries to establish that reinforcement-learning post-training of clinical vision-language models can be made fairness-aware without sacrificing accuracy. It proposes a method, called EQPO in the title and FairGRPO in the body, that rescales each sample's RL advantage by the representation and difficulty of its demographic group so that underrepresented and harder subgroups receive stronger learning signals. When demographic labels are absent, it clusters samples by their reward patterns to recover implicit groups. Across seven diagnostic benchmarks spanning X-ray, CT, dermoscopy, mammography, and ultrasound, the authors report that this reduces the spread of F1 across groups by up to 43.9% and the largest cross-group F1 gap by up to 42.7% on one model, and cuts predictive-parity gaps by 27.2% while raising F1 by 12.5% on another. A sympathetic reader would care because it suggests fairness can be built into the RL objective itself rather than patched on after training.

What carries the argument

The load-bearing mechanism is a two-tier temperature factor: for a domain g and a group gamma (explicit or implicit), T(g,t) = sqrt(N(g,t) * mean_reward) and T(gamma,g,t) = sqrt(N(gamma,g,t) * mean_reward). The normalized reward is divided by the product of these temperatures, s_scaled = normalized_reward / max(T_domain * T_group, epsilon), so smaller and harder groups receive amplified advantages; the scaled advantages are then renormalized batch-wide. The other piece is reward-vector K-means clustering: for unlabeled prompts, the vector of rewards across rollouts is the feature, so clusters group cases with similar difficulty patterns rather than similar pixels, and the elbow method select

What would settle it

Record the true demographic labels, run the no-label variant, and compute an adjusted Rand index or cluster purity between the discovered clusters and true age and gender groups; an index near zero would show that the unsupervised clusters do not recover demographics. A complementary check is to shuffle cluster assignments during training: if the F1-gap reductions persist, the gains come from difficulty reweighting rather than equitable grouping.

Watch

Extended reading notes

Core claim

The central discovery is that a small, cheap modification to GRPO—replacing its uniform within-group normalization with a hierarchical inverse-temperature scaling based on subgroup count and mean reward—shifts the optimization landscape so that majority groups no longer dominate gradients. The discovery also includes an unsupervised component: when demographic labels are missing, K-means on the vectors of per-rollout rewards discovers latent subpopulations, and the paper claims these implicit groups are stable and align with masked age and gender attributes. The empirical result is that fairness metrics improve throughout training and overall diagnostic performance is not hurt; in the labele

Load-bearing premise

The load-bearing premise is that, when demographic labels are missing, the reward-based k-means clusters correspond to the actual demographic subpopulations; the full-text no-label row already shows worse predictive parity and equalized odds than vanilla GRPO on MedGemma, so if the clusters merely track case difficulty, the no-label fairness claim collapses.

Editorial extensions

If this is right

  • If the reported numbers hold, clinical vision-language models can be RL post-trained for equity with no accuracy penalty: the labeled variant beats vanilla GRPO in F1 on both tested base models.
  • Fairness becomes a train-time property: the training trajectory shows the cross-group F1 gap shrinking as RL proceeds, unlike baselines whose gap grows with training.
  • The no-label variant offers a route for fairness improvements when demographic recording is incomplete, because reward-only clustering supplies group structure.
  • Because the added computation is negligible, the method can be attached to any GRPO-style clinical reasoning pipeline.
  • The released models are claimed to set a new state of the art on the seven-benchmark suite while having markedly smaller demographic gaps.

Reading between the lines

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

  • Editorial note: the title and abstract name the method EQPO, while the full text calls it FairGRPO and the released model FairMedGemma; the abstract-level numbers for QoQ-Med3-8B do not appear in the full-text Table 2, so readers should separate abstract claims from the presented tables.
  • The full text's no-label row (FairGRPO ND) shows worse predictive parity and equalized odds than vanilla GRPO on MedGemma; if the reward-based clusters track difficulty rather than demographics, the no-label variant is better described as difficulty-aware weighting than as a demographic fairness result.
  • A natural next experiment is to hold out demographic labels and measure cluster purity against age and gender; high purity would validate the equity interpretation, while low purity would require restating the fairness claim.
  • The inverse-square-root temperature form suggests a tunable family of fairness-utility tradeoffs; sweeping the exponent could give practitioners a dial for deployment contexts with binding equity constraints.
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 / 5 minor

Summary. The manuscript as identified by its arXiv title and abstract proposes EQPO (Equitable Group Relative Policy Optimization), a hierarchical reinforcement learning method that extends GRPO for clinical VLLMs. The abstract claims that EQPO reduces F1 standard deviation by 43.9% and the maximum cross-group F1 gap by 42.7% on QoQ-Med3-8B, narrows predictive parity by 27.2% on MedGemma-4B, and improves F1 by 12.5% even without demographic labels, with implicit clusters that are stable and aligned with masked demographic attributes. The full text, however, presents a different method, FairGRPO, evaluated on Qwen-2.5-VL-7B and MedGemma-4B, and reports no results for QoQ-Med3-8B and none of the two F1-based headline numbers. Within the body, the no-label variant (FairGRPO ND) is claimed to improve fairness, but Table 2 shows that on MedGemma it yields worse predictive parity, equal opportunity difference, false-positive-rate difference, sigma-F1, and Delta-F1 than vanilla GRPO. The paper releases code/models, but the central fairness-without-labels claim is not supported by the reported experiments.

Significance. Fairness-aware RL post-training for clinical vision-language models is an important and timely problem, and the paper's empirical setup—seven datasets across five modalities, comparison with GRPO, RLOO, REINFORCE++, Group DRO, and resampling—is a useful and ambitious framework. The proposed idea of adaptively scaling advantages by group representation and difficulty is plausible and worth studying. However, the submitted manuscript cannot support its significance claim: the title/abstract and the body describe different methods and models, the no-label variant's fairness metrics contradict the abstract, and the cluster-alignment assertion is not demonstrated. If the method and the experiments were properly aligned and the contradictions resolved, this could be a valuable contribution, but the current text does not permit such an assessment.

major comments (5)
  1. [Abstract vs. §§1–5] The title and abstract describe EQPO, EquiQoQ-Med3-8B, EquiMedGemma-4B, and the headline numbers '43.9% F1 standard deviation reduction' and '42.7% maximum cross-group F1 gap reduction.' The full text is a different paper on FairGRPO/FairMedGemma evaluated on Qwen-2.5-VL-7B and MedGemma-4B; the terms EQPO, EquiQoQ-Med3-8B, and the two F1 headline numbers never appear in the body, and QoQ-Med3-8B is not in Table 1. The abstract also claims implicit groups 'align with masked demographic attributes,' but no such analysis is present. Because the central advertised claims are not backed by any experimental evidence in the manuscript, this is a load-bearing inconsistency, not a presentation issue.
  2. [§4.2, Table 2] Table 2 directly contradicts the no-label fairness claim. For MedGemma, FairGRPO ND (clustering, no demographic labels) has PP 25.15 vs. GRPO 22.42, EOD 11.56 vs. 6.476, FPRDiff 5.692 vs. 4.820, sigma-F1 .0547 vs. .0418, and Delta-F1 .1067 vs. .0795—all worse than vanilla GRPO. The text's assertion that 'FairGRPO ND improves fairness and performance even when no demographic information is passed' is therefore false for every standard fairness metric in the table. The 27.2% PP reduction cited in the abstract comes from labeled FairGRPO vs. GRPO+DRO (16.04 to 11.67), not from the no-label variant. Also, §4.2's claim that 'EOD improves by 23.8% on MedGemma' is contradicted by Table 2, where FairGRPO's EOD (6.663) is worse than GRPO's (6.476).
  3. [§3(ii), abstract] The method's no-label component rests on the assumption that reward-based k-means clusters recover latent demographic subpopulations. The abstract states that 'the discovered implicit groups remain stable and align with masked demographic attributes,' but the full text contains no experiment comparing the inferred clusters with held-out demographic labels, no stability analysis, and no cluster-quality metric. Without that evidence, Eq. (1)'s cluster-based temperature T(gamma,g,t) reweights by reward similarity/difficulty, not necessarily by demographic equity. This is the central mechanism behind the 'fairness without labels' claim, and it is currently unsubstantiated.
  4. [§4.1, Appendix C] All reported results come from a single training run per condition, with no seeds, error bars, or confidence intervals. Several test sets are small—e.g., CheXpert 225 records, PAD-UFES-20 459 records, COVID-BLUES 96 records—and per-demographic-group cells are much smaller. Many of the claimed improvements in §4.2 and Fig. 3 (e.g., '73.08% improvement on 75+ populations in PAD-UFES-20') are therefore point estimates of unknown reliability. Given that the headline numbers are precise percentages, this lack of statistical grounding is a major concern for any quantitative conclusion.
  5. [§4.2, Table 2 (Qwen row)] The text says FairGRPO 'outperforms classical bias mitigation methods in both fairness and diagnosis performance,' but on Qwen-2.5-VL-7B, FairGRPO's PP is 16.80, worse than GRPO's 11.39 and GRPO+DRO's 14.51. Even on the labeled variant, the comparison is not uniformly favorable; the paper's own table shows the claim as stated is too strong.
minor comments (5)
  1. [Throughout] Numerous typos and inconsistent terminology: 'dermoscropy,' 'faieness,' 'mdoel,' and inconsistent use of EQPO/EquiQoQ-Med3-8B in the abstract versus FairGRPO/FairMedGemma in the body.
  2. [Figure 2 caption] The caption labels two subfigures as '(e)'—one for per-step runtime and one for the performance-fairness tradeoff—and the in-text references to these panels are confusing.
  3. [Eq. (2)] The epsilon in max(T(g,t)*T(gamma,g,t), epsilon) is not defined in the method text; it should be stated as a numerical stability constant and its value given.
  4. [References] The VinDr-Mammo dataset is cited as 'Nguyen et al., 2021' but the reference points to VinDr-CXR, not VinDr-Mammo.
  5. [§7] The reproducibility statement says trained model weights 'will be made available upon publication,' while the abstract and introduction say the models are 'released.' The distinction should be stated consistently.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the core reweighting is an empirical algorithm design and no derivation reduces to its inputs, though the abstract/full-text mismatch is a separate integrity concern.

full rationale

The paper's central method, FairGRPO, is an empirical RL reweighting scheme: it normalizes GRPO advantages, computes temperature factors from group counts and mean rewards, and rescales advantages. The temperature factors are functions of the same rewards used in the advantages, but this is intrinsic to the algorithm's design rather than a fitted parameter being later renamed as a prediction. The unsupervised clustering step defines implicit groups from reward vectors and then scales by cluster size and reward; the paper's reported fairness metrics, however, are evaluated on explicit demographic labels, so the claim that clustering improves demographic fairness is an empirical assertion, not a definitional equivalence. No uniqueness theorem, ansatz, or load-bearing conclusion is imported from the authors' prior work; the self-citations (e.g., Dai et al. 2025b,c) appear in background and are not used to justify the algorithm's correctness. The abstract and the full text describe different methods and models (EQPO/QoQ-Med3-8B vs. FairGRPO/Qwen-2.5-VL/MedGemma), and Table 2 shows FairGRPO ND has worse predictive parity and EOD than vanilla GRPO; these are serious correctness and reporting inconsistencies, but they are not instances of circular derivation and therefore do not raise the circularity score. Honest non-finding: the derivation chain is self-contained in the sense that no step is equivalent to its inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central method depends on hand-chosen scaling choices (exponents, epsilon) and on the assumptions that reward clustering recovers demographic structure and that test sets are adequate; the paper provides no sensitivity analysis, and Table 2 contradicts the no-label fairness claim.

free parameters (4)
  • scaling exponents in T = sqrt(N·r̄) = 0.5, 0.5
    The multiplicative form and square-root exponents in Eq. 1 are chosen without ablation or theoretical justification; different exponents would change the strength of reweighting.
  • epsilon in max(..., ε) = not reported
    Eq. 2 uses ε without stating its value; affects scaling for small groups.
  • age bins (18-25, 26-50, 51-75, 76+) = 25-year bins
    Arbitrary binning; all fairness metrics and conclusions depend on this choice.
  • number of clusters in k-means = elbow method, per iteration
    Elbow method is subjective; no cluster counts or stability reported.
assumptions (5)
  • standard math GRPO advantage normalization (standard)
    Section 3 Background: A_GRPO = (r - μ_G)/σ_G, standard.
  • domain assumption Binary accuracy reward adequately measures clinical diagnostic quality
    Reward design in Sec. 3: 1 if final answer correct; ignores clinical severity, uncertainty, and reasoning quality.
  • ad hoc to paper Reward-based k-means clusters correspond to demographic groups
    Sec. 3(ii): k-means on reward vectors is assumed to discover latent subpopulations; the claimed alignment with masked demographics is not demonstrated in the body.
  • domain assumption Test sets large enough for stable per-group F1 estimates
    CheXpert test 225 records; Hemorrhage 515; per age-gender cells can be tiny; no confidence intervals.
  • domain assumption Answer extraction from free-text traces is lossless
    Evaluation extracts diagnoses from reasoning traces (App. B.1) but parsing method is not described; errors would bias F1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EQPO: Equitable Group Relative Policy Optimization for Clinical Reasoning." pith.science (2026). https://pith.science/paper/DBVVU72G

@misc{pith2026251019893,
  author       = {Pith},
  title        = {Pith review of: EQPO: Equitable Group Relative Policy Optimization for Clinical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBVVU72G}},
  note         = {Machine review of arXiv:2510.19893}
}
read the original abstract

Medical AI systems demonstrated impressive diagnostic performance, yet they routinely show uneven accuracy across demographic groups, disadvantaging underrepresented populations. Although multimodal reasoning foundation models have pushed clinical diagnosis forward, reinforcement learning-based post-training tends to absorb and magnify the biases present in majority-dominated training corpora. We propose Equitable Group Relative Policy Optimization (EQPO), a hierarchical reinforcement learning method that encourages balanced learning across heterogeneous clinical populations by adaptively reweighting samples according to subgroup representation, task difficulty, and data source. As demographic annotations are frequently missing in real-world clinical data, EQPO additionally applies unsupervised clustering to recover latent subpopulations when they are unavailable. On 7 diagnostic benchmarks covering 5 modalities (X-ray, CT, dermoscopy, mammography, ultrasound), EQPO reduces F1 standard deviation by 43.9% and the maximum cross-group F1 gap by 42.7% on QoQ-Med3-8B over vanilla GRPO, and narrows predictive parity gaps by 27.2% on MedGemma-4B over bias-mitigated RL baselines while raising F1 by 12.5% even without any demographic labels. Examining the training trajectory shows that EQPO steadily improves fairness over the course of optimization, in contrast to baseline methods whose fairness degrades as training proceeds, and the discovered implicit groups remain stable and align with masked demographic attributes. We further release EquiMedGemma-4B and EquiQoQ-Med3-8B, equitability-aware clinical VLLMs that attain state-of-the-art accuracy with markedly smaller demographic gaps.

Figures

Figures reproduced from arXiv: 2510.19893 by the authors.

Figure 1
Figure 1. FairGRPO Training Pipeline. Our method addresses fairness disparities by adaptively scaling rewards based on demographic representation and task difficulty. Starting with medical data containing both labeled demographic information and unlabeled samples, the policy model generates multiple responses for each prompt, producing both reasoning insights and clinical diagnoses. These responses are evaluated and assigned … view at source ↗
Figure 2
Figure 2. Training dynamics comparison between GRPO and FairGRPO on clinical classification [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. F1 score differences between FairGRPO and GRPO across demographic groups on MedGemma. Each bar represents the F1 score difference from the population mean for specific demographic subgroups, where a positive value means FairGRPO performs better for the given demographic group. The four age groups are binned as described in Sec. 4.1. In general, Fair￾GRPO consistently demonstrates better performance for 25 out of the… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative Examples of Model’s Reasoning Traces. We see the greatest performance boosts from underrepresented groups, including samples from older population and females. In par￾ticular, we observe the models trained with FairGRPO exhibit an accuracy improvement of 73…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 canonical work pages

  1. [3]

    10 Preprint Under Review Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923,

    arXiv

  2. [5]

    Jiaee Cheong, Aditya Bangar, Sinan Kalkan, and Hatice Gunes

    AI for Good. Jiaee Cheong, Aditya Bangar, Sinan Kalkan, and Hatice Gunes. U-fair: Uncertainty-based multi- modal multitask learning for fairer depression detection. InMachine Learning for Health (ML4H), pp. 203–218. PMLR, 2025a. Jiaee Cheong, Abtin Mogharabin, Paul Liang, Hatice Gunes, and Sinan Kalkan. Fairwell: Fair multimodal self-supervised learning f...

    arXiv

  3. [6]

    Biomedical visual instruction tuning with clinician preference alignment.arXiv preprint arXiv:2406.13173,

    Hejie Cui, Lingjun Mao, Xin Liang, Jieyu Zhang, Hui Ren, Quanzheng Li, Xiang Li, and Carl Yang. Biomedical visual instruction tuning with clinician preference alignment.arXiv preprint arXiv:2406.13173,

    arXiv

  4. [7]

    Developing icu clinical behavioral atlas using ambient intelligence and computer vision.NEJM AI, pp

    Wei Dai, Ehsan Adeli, Zelun Luo, Dev Dash, Shrinidhi Lakshmikanth, Zane Durante, Paul Tang, Amit Kaushal, Arnold Milstein, Li Fei-Fei, et al. Developing icu clinical behavioral atlas using ambient intelligence and computer vision.NEJM AI, pp. AIoa2400590, 2025a. Wei Dai, Peilin Chen, Chanakya Ekbote, and Paul Pu Liang. Qoq-med: Building multimodal clinica...

    arXiv

  5. [8]

    Graph-based patient representation for multimodal clinical data: Addressing data heterogeneity.medRxiv, pp

    Suparna Ghanvatkar and Vaibhav Rajan. Graph-based patient representation for multimodal clinical data: Addressing data heterogeneity.medRxiv, pp. 2023–12,

    2023

  6. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  7. [11]

    Ball, Katie S

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Christopher Chute, Henrik Marklund, Behzad Haghgoo, Robyn L. Ball, Katie S. Shpanskaya, Jayne Seekins, David A. Mong, Safwan S. Halabi, Jesse K. Sandberg, Ricky Jones, David B. Larson, Cur- tis P. Langlotz, Bhavik N. Patel, Matthew P. Lungren, and Andrew Y . Ng. Chexpert: A large...

    2019

  8. [13]

    Seungeun Lee, Yongwon Cho, Yuyoung Ji, Minhyek Jeon, Aram Kim, Byung-Joo Ham, and Yoon- jung Yoonie Joo

    doi: 10.1073/pnas.1919012117. Seungeun Lee, Yongwon Cho, Yuyoung Ji, Minhyek Jeon, Aram Kim, Byung-Joo Ham, and Yoon- jung Yoonie Joo. Multimodal integration of neuroimaging and genetic data for the diagnosis of mood disorders based on computer vision models.Journal of psychiatric research, 172:144–155,

    doi:10.1073/pnas.1919012117

Show all 29 references
  1. [14]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al

    doi: 10.1126/science.aax2342. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to fol- low instructions with human feedback.Advances in neural informat...

    doi

  2. [15]

    URL https://doi.org/10.1016/j.dib.2020.106221

    doi: 10.1016/j.dib.2020.106221. URL https://doi.org/10.1016/j.dib.2020.106221. Esther Puyol-Ant´on, Bram Ruijsink, Stefan K Piechnik, Stefan Neubauer, Steffen E Petersen, Reza Razavi, and Andrew P King. Fairness in cardiac mr image analysis: an investigation of bias due to dat...

    2020

  3. [17]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    arXiv

  4. [19]

    Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew B. A. McDermott, Irene Y . Chen, and Marzyeh Ghassemi. Chexclusion: Fairness gaps in deep chest x-ray classifiers. InBiocomputing 2021: Proceedings of the Pacific Symposium on Biocomputing, pp. 232–243. World Scientific, 2021a. do...

    2021 doi

  5. [21]

    doi: 10.4103/heartviews.heartviews 103

    doi

  6. [22]

    Laleh Seyyed-Kalantari, Haoran Zhang, Matthew B

    URLhttps://www.worldscientific.com/doi/ abs/10.1142/9789811232701_0022. Laleh Seyyed-Kalantari, Haoran Zhang, Matthew B. A. McDermott, Irene Y . Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served p...

    arXiv

  7. [24]

    Recorded Feb–May 2021 at Maas- tricht University Medical Center (UMC+); CC BY-NC-ND 4.0

    URL https://github.com/NinaWie/COVID-BLUES. Recorded Feb–May 2021 at Maas- tricht University Medical Center (UMC+); CC BY-NC-ND 4.0. Jenny Yang, Andrew AS Soltan, David W Eyre, and David A Clifton. Algorithmic fairness and bias mitigation for clinical machine learning with dee...

    2021

  8. [25]

    Multimodal healthcare ai: identifying and designing clinically relevant vision-language applications for radiology

    Nur Yildirim, Hannah Richardson, Maria Teodora Wetscherek, Junaid Bajwa, Joseph Jacob, Mark Ames Pinnock, Stephen Harris, Daniel Coelho De Castro, Shruthi Bannur, Stephanie Hy- land, et al. Multimodal healthcare ai: identifying and designing clinically relevant vision-language...

    2024

  9. [26]

    All models are trained with 4 NVIDIA H200 GPUs

    A APPENDIX B HYPERPARAMETERS& SETUPS In this section, we describe our setup and hyperparameters during the training of the model. All models are trained with 4 NVIDIA H200 GPUs. All experiments were conducted using the VERL (V olcano Engine Reinforcement Learning for LLMs) fra...

    2024

  10. [27]

    The dataset includes local labels for bounding boxes; however, we evaluate our models based on the 5 global labels for BI-RADS 1-5

    contains mammography collected from Hospital 108 and Hanoi Medical University Hospital in Vietnam. The dataset includes local labels for bounding boxes; however, we evaluate our models based on the 5 global labels for BI-RADS 1-5. We use a training set of 16,000 records, a tes...

    2020

  11. [28]

    Malignant

    comprises dermoscopy of skin lesions from over 2,000 patients, generated by the International Skin Imaging Collaboration (ISIC). We evaluate the models on the binary classification (“Malignant” or “Benign”) for each image, where all malignant diagnoses are histopathology–confi...

    2018

  12. [29]

    No Hemorrhage

    consists of intracranial hemorrhage CT images for 82 patients at Al Hilla Teaching Hospital, Iraq, each with brain and bone window images and approximately 30 image slices in total. We evaluate the models as binary diagnoses: “No Hemorrhage” and “Has Hemorrhage”. We use a trai...

    2020

  13. [1953]

    The HAM10000 dataset, a large collec- tion of multi-source dermatoscopic images of common pigmented skin lesions.Scientific Data, 5(1):180161, 2018a

    Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The HAM10000 dataset, a large collec- tion of multi-source dermatoscopic images of common pigmented skin lesions.Scientific Data, 5(1):180161, 2018a. doi: 10.1038/sdata.2018.161. URLhttps://www.nature.com/ articles/sdata20...

    2018 arXiv

  14. [2017]

    Medgemma technical report.arXiv preprint arXiv:2507.05201,

    13 Preprint Under Review Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, C ´ıan Hughes, Charles Lau, et al. Medgemma technical report.arXiv preprint arXiv:2507.05201,

    arXiv

  15. [2019]

    URL https://doi.org/10.1609/aaai.v33i01.3301590

    doi: 10.1609/AAAI.V33I01.3301590. URL https://doi.org/10.1609/aaai.v33i01.3301590. Shahin Jabbari, Matthew Joseph, Michael Kearns, Jamie Morgenstern, and Aaron Roth. Fairness in reinforcement learning. InInternational conference on machine learning, pp. 1617–1626. PMLR,

    doi

  16. [2020]

    Reinforce++: A simple and efficient approach for aligning large language models.arXiv preprint arXiv:2501.03262,

    11 Preprint Under Review Jian Hu. Reinforce++: A simple and efficient approach for aligning large language models.arXiv preprint arXiv:2501.03262,

    arXiv

  17. [2021]

    URL https://www.nature.com/articles/s41597-021-00815-z

    doi: 10.1038/s41597-021-00815-z. URL https://www.nature.com/articles/s41597-021-00815-z. Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generaliza...

    1911 arXiv

  18. [2022]

    Language models get a gender makeover: Mitigating gender bias with few-shot data interventions.arXiv preprint arXiv:2306.04597,

    Himanshu Thakur, Atishay Jain, Praneetha Vaddamanu, Paul Pu Liang, and Louis-Philippe Morency. Language models get a gender makeover: Mitigating gender bias with few-shot data interventions.arXiv preprint arXiv:2306.04597,

    arXiv

  19. [2023]

    Back to basics: Revisiting reinforce-style optimization for learn- ing from human feedback in llms

    Arash Ahmadian, Chris Cremer, Matthias Gall ´e, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet ¨Ust¨un, and Sara Hooker. Back to basics: Revisiting reinforce-style optimization for learn- ing from human feedback in llms. In Lun-Wei Ku, Andre Martins, and Vivek Srikuma...

    2024

  20. [2024]

    URL https://doi.org/10.18653/v1/2024.acl-long.662

    doi: 10.18653/V1/2024.ACL-LONG.662. URL https://doi.org/10.18653/v1/2024.acl-long.662. Rawan AlSaad, Alaa Abd-Alrazaq, Sabri Boughorbel, Arfan Ahmed, Max-Antoine Renault, Rafat Damseh, and Javaid Sheikh. Multimodal large language models in health care: applications, challenges...

    2024 doi

  21. [2025]

    Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862,

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862,

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.