Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Generalization of Medical Large Language Models through Cross-Domain Weak Supervision

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a three-component training recipe—curriculum ordering, dual-stage memory, and low-rank adapters—lets medical LLMs beat four baselines on question answering, preference classification, and response generation.

desk verdict The paper's central mechanism is self-defeating as written, the title promises a method that never appears, and the experimental claims have no supporting artifacts. read the letter →

arxiv 2502.00832 v1 pith:M6AVXTTV submitted 2025-02-02 cs.CL

Robert Long , Eric Gonzalez , Harrison Fuller This is my paper · ORCID
classification cs.CL
keywords largelanguagemodelsmedicalNLPcurriculumlearningparameter-efficientfine-tuningLoRAdual-stagememoryweak-to-stronggeneralization
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 proposes Incremental Curriculum-Based Fine-Tuning (ICFT), a training recipe for adapting general-purpose large language models to medicine. The intended transition is from 'weak' general knowledge to 'strong' medical expertise: training data are ordered from general medical texts through clinical guidelines and case studies to real patient-doctor dialogues, while the model retains recent dialogue context in a short-term memory, keeps frequently accessed knowledge in a long-term memory, and updates only a low-rank adapter instead of the full network. The authors claim ICFT beats standard fine-tuning, plain LoRA, a dictionary-memory method, and a dual-process memory baseline on all three tasks they test, with a 2.58-point ROUGE-1 gain and a 6.61-point higher win rate over the strongest baseline, plus better generalization to an unseen rare-disease dataset. If these results hold, a parameter-efficient, incremental schedule could make medical LLM adaptation cheaper and less prone to catastrophic forgetting, which matters for clinical deployment.

What carries the argument

The central objects are three: (1) a domain adapter defined by a down-projection matrix $W_{down}$ and an up-projection matrix $W_{up}$ with a nonlinearity between them, added to the base weights as $\Delta W_{adapter} = W_{up} W_{down}$; (2) dual-stage memory coordination, in which a short-term memory stores recent dialogue rounds with a fixed capacity $K$ and a long-term memory stores frequently accessed knowledge items above a frequency threshold $\theta$, with attention-based retrieval combining both into a weighted representation; and (3) LoRA fine-tuning, where the weight update is a product of two low-rank matrices $A$ and $B$. The paper also defines a consistency loss that penalizes the distance between base and adapted outputs, but the memory modules and curriculum are the components whose ablation produces the reported performance drops.

What would settle it

Re-run ICFT with the consistency loss weight varied from zero to a value that actually enforces the constraint; if enforcing the constraint collapses the adapter toward the identity and the reported medical-QA gains disappear, the claimed knowledge-injection mechanism is not what drives the results. A second check: read the training loop in the code (if released) to see whether Eq. (3) is ever back-propagated.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that ordering training data by increasing medical complexity and combining that with a dual-stage memory and LoRA-based parameter-efficient fine-tuning yields state-of-the-art performance across medical QA, preference classification, and response generation. The authors report that ICFT reaches 41.59 ROUGE-1 on QA, 72.89% preference classification accuracy, and a 91.53% response-generation win rate, while training only 0.55% of the parameters required for full fine-tuning. They attribute the gains to the progressive curriculum, the short-term/long-term memory coordination, and the adapter-based tuning, and they report that ablating memory coordination or the curriculum lowers QA ROUGE-1 by roughly 1.7 and 3.5 points respectively, and that the full model shows lower factual, contextual, and fluency error rates and higher response diversity.

Load-bearing premise

The framework assumes that the consistency loss in Eq. (3), which forces the adapted model's output to match the base model's output, can be minimized while the adapter still injects new medical knowledge; the paper never states the loss weight or shows how both objectives are jointly satisfiable.

Editorial extensions

If this is right

  • If ICFT works as claimed, medical LLM adaptation could require only about 0.55% of the trainable parameters of full fine-tuning, making specialty deployment feasible on modest hardware.
  • The reported generalization to an unseen rare-disease dataset suggests a curriculum order could substitute for simply adding more flat medical text.
  • Lower factual, contextual, and fluency error rates combined with higher distinct-$n$ diversity would make patient-facing dialogue systems less repetitive and safer.
  • The two-tier memory design could carry over to other high-stakes domains where recent context and frequently accessed facts both matter, such as legal or financial advice.

Reading between the lines

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

  • The consistency loss in Eq. (3) actually punishes the adapter for changing the model's output on the training samples, so the reported gains may be carried by the curriculum and memory modules rather than by the adapter's knowledge injection; a clean test is to run ICFT with the consistency term removed and compare the QA scores.
  • The paper's title advertises cross-domain weak supervision, but no weak-supervision signal or cross-domain auxiliary task is defined in the method; one natural reading is that the 'weak' stage is general-domain pretraining and the 'strong' stage is medical fine-tuning, which is a much weaker notion than the title suggests.
  • Because the memory retrieval equations (6)-(8) are presented without a training loss, the memory modules may function at inference time only; a testable extension is to see whether the reported retrieval accuracies are sensitive to the STM capacity $K$ and the LTM frequency threshold $\theta$, as a genuine memory mechanism should degrade predictably when $K$ is shrunk.
  • If the curriculum ordering is the real driver, then a version of ICFT with the memory modules removed but the same data ordering should retain most of the QA gains; that would isolate the 'weak to strong' contribution from the memory contribution.
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 paper proposes Incremental Curriculum-Based Fine-Tuning (ICFT), a framework that combines curriculum learning, a dual-stage short-term/long-term memory mechanism, and LoRA-based parameter-efficient fine-tuning to adapt large language models to medical NLP. It reports experiments on medical question answering, preference classification, and response generation, claiming consistent improvements over SFT, LoRA, Dict-Mem, and DPeM, supplemented by ablations, human evaluation, efficiency, generalization, and error/diversity analyses. The central claim is that ICFT achieves state-of-the-art performance across these tasks.

Significance. The problem addressed is important: catastrophic forgetting and data efficiency in domain adaptation of LLMs to medicine are real concerns, and the high-level idea of staged curriculum training with PEFT is a plausible direction. However, the paper as written does not establish the claimed result. The formal method contains a direct contradiction between the consistency loss and the knowledge-injection objective, a matrix-dimension error in the adapter update, an underspecified memory module, and no experimental details that would allow reproduction of Table 1. No code or data are provided, and no statistical measures accompany the tables. If the method and experiments were corrected and rerun, the framework could be a useful incremental-learning contribution, but in its current form the empirical claims are not supported.

major comments (5)
  1. [Section 3.1, Eqs. (1)-(3), and Section 3.4, Eq. (11)] The consistency loss L_consistency in Eq. (3) penalizes the mean squared difference between the base-model output Phi(x_i; W) and the adapted output Phi(x_i; W + Delta_W_adapter). Minimizing this term drives Delta_W_adapter toward zero, which directly opposes the stated goal of injecting medical knowledge through the adapter. Eq. (11) sums L_consistency, L_task, and L_fine-tune with no weighting coefficients, and the paper never reports a loss weight, stop-gradient, early-stopping rule, or any other mechanism that would prevent the adapter from collapsing to the base model. Consequently the improvements in Table 1 cannot be attributed to the training objective as written.
  2. [Section 3.1, Eq. (2)] Eq. (2) is not dimensionally consistent. With W_down in R^{d x r} and W_up in R^{r x d}, the product Delta_W_adapter = W_up W_down has shape r x r, whereas W in Phi(x; W) is a weight matrix of the base model and should be updated by a matrix of the same shape. The paper also does not specify how the adapter output f_adapter(x) of Eq. (1) is added to the transformer or how Delta_W_adapter is added to W. As written, the adapted parameter update is not implementable.
  3. [Section 3.2, Eqs. (6)-(8)] The memory module is not formally connected to the language model. The retrieval equations use q, D_STM, D_LTM, a_STM, a_LTM, W_STM, and W_LTM without specifying their dimensions, how the memory content is represented, or how the combined representation z is fed into Phi or the adapter. The 'retrieval accuracy' reported in Table 6 is never defined. Without this integration, the memory mechanism is not a well-defined component and its contribution cannot be verified.
  4. [Section 4.1 and Table 1] The experimental section omits the base LLM, dataset splits, curriculum ordering and difficulty schedule, hyperparameters (learning rate, batch size, number of epochs, LoRA rank r, regularization coefficient lambda, STM capacity K, LTM threshold theta, consistency-loss weight), and the evaluation protocol. All tables report point estimates with no variance, confidence intervals, significance tests, or number of runs. For example, Table 1 reports ICFT's QA ROUGE-1 as 41.59 but gives no standard deviation or repetition count, and no code or data are released. The central empirical claim of state-of-the-art performance is therefore not supported by the information in the manuscript.
  5. [Title and Abstract vs. Section 3] The title and abstract promise 'cross-domain weak supervision,' but no weak-supervision labels, cross-domain transfer mechanism, or weakly supervised training signal appears anywhere in Section 3. The phrase 'Weak to Strong' in the introduction refers to moving from weak general knowledge to strong domain expertise, which is a different concept from weak supervision in the machine-learning sense. The claimed contribution is thus not represented by the described method.
minor comments (5)
  1. [Section 3.3, Eq. (9)] The citation of LoRA points to reference [9], which is a healthcare embedding paper rather than the LoRA method; the LoRA formulation should be cited to its original source.
  2. [Section 4.4] The human evaluation uses only two experts and 100 responses and reports no inter-annotator agreement or detailed scoring rubric; this should be described or the claim of superior response quality should be tempered.
  3. [Table 2] The ablation row 'w/o LoRA Fine-Tuning' is not explained; if ICFT is built on LoRA, the configuration and remaining components for that row should be specified.
  4. [Section 4.6] The 'unseen medical dataset' is not described in terms of size, sources, or relation to the training data, so the generalization claim cannot be assessed.
  5. [Throughout] There are minor language and typesetting issues in the text, such as 'particularl y' and 'fine-tuning,' and several acronyms are introduced without definitions; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the paper's limitations are missing experimental detail and internal consistency of the loss, not circularity.

full rationale

The ICFT paper does not derive any result from assumptions that already contain the result. Its central claims are empirical comparisons against baselines; no parameter is fitted to a target and then reported as a prediction of that same target, and no load-bearing premise is justified solely by a self-citation. The only candidate issue is the consistency loss in Eq. (3), which penalizes the difference between base and adapted outputs and could in principle cancel the knowledge-injection update unless the loss weight or an optimization strategy is specified. That is a potential internal inconsistency or reproducibility gap, not an instance of the paper's output being equivalent to its input by construction. Section 3.4 simply sums the losses in Eq. (11) with no weights, and Section 4.1 omits training hyperparameters; these omissions make the reported gains in Tables 1-8 difficult to attribute to the described method, but they do not make the claims circular. There is also no self-citation chain: the cited related work is not authored by Long, Gonzalez, or Fuller, and no uniqueness or existence theorem from prior work is used to force the framework's design. Accordingly, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

Everything the framework claims to add (adapter, dual memory, curriculum) is described at the level of equations that do not close, with no code, data, or hyperparameters. The free parameters are all unreported tuning choices. The invented memory modules have no independent evidence of being implementable.

free parameters (6)
  • LoRA/adapter rank r = not reported
    Determines capacity of adapter and LoRA; no value or tuning procedure is given (Section 3.1, 3.3).
  • Regularization coefficient lambda = not reported
    Appears in Eq. (10); no value or sensitivity analysis is provided (Section 3.3).
  • STM capacity K = not reported
    Maximum number of stored dialogue rounds in Eq. (4); no value is given (Section 3.2).
  • LTM frequency threshold theta = not reported
    Required for adding knowledge to long-term memory in Eq. (5); no value is given (Section 3.2).
  • Consistency loss weight (implicitly 1) = not reported
    Eq. (11) sums losses with equal weight; no weighting or justification is given, and the consistency loss conflicts with task adaptation (Section 3.4).
  • Curriculum ordering and difficulty schedule = not reported
    The paper describes a progressive curriculum but never defines difficulty ordering or stage transitions (Section 3.4).
assumptions (4)
  • ad hoc to paper Minimizing L_consistency (Eq. 3) is compatible with injecting medical knowledge into the adapted model.
    The loss forces adapted outputs toward base outputs, working against the adapter's purpose; this is assumed without discussion in Section 3.1.
  • ad hoc to paper The dual-stage memory equations (4)-(8) describe a differentiable component that can be trained jointly with the LLM.
    No forward/backward pass details or integration with the loss are given in Section 3.2.
  • domain assumption HealthCareMagic and iCliniq are suitable and available datasets for the reported QA evaluation.
    The datasets are named but not described, split, or linked in Section 4.1.
  • standard math Standard LoRA parameterization preserves general knowledge while adapting to medical tasks.
    This is a generally accepted property of LoRA, though the paper provides no specific evidence for its setting.
invented entities (2)
  • Short-Term Memory (STM) module
    purpose: Stores recent dialogue rounds for context-aware inference (Eq. 4).
    Defined only by a set update rule and attention retrieval; no implementation, initialization, or ablation-specific setting is given in Section 3.2.
  • Long-Term Memory (LTM) module
    purpose: Stores frequently accessed medical knowledge based on a frequency threshold (Eq. 5).
    No storage representation, update mechanism, or evidence that it can be trained is provided in Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalization of Medical Large Language Models through Cross-Domain Weak Supervision." pith.science (2026). https://pith.science/paper/M6AVXTTV

@misc{pith2026250200832,
  author       = {Pith},
  title        = {Pith review of: Generalization of Medical Large Language Models through Cross-Domain Weak Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M6AVXTTV}},
  note         = {Machine review of arXiv:2502.00832}
}
read the original abstract

The advancement of large language models (LLMs) has opened new frontiers in natural language processing, particularly in specialized domains like healthcare. In this paper, we propose the Incremental Curriculum-Based Fine-Tuning (ICFT) framework to enhance the generative capabilities of medical large language models (MLLMs). ICFT combines curriculum-based learning, dual-stage memory coordination, and parameter-efficient fine-tuning to enable a progressive transition from general linguistic knowledge to strong domain-specific expertise. Experimental results across diverse medical NLP tasks, including question answering, preference classification, and response generation, demonstrate that ICFT consistently outperforms state-of-the-art baselines, achieving improvements in both accuracy and efficiency. Further analysis reveals the framework's ability to generalize to unseen data, reduce errors, and deliver diverse, contextually relevant medical responses. These findings establish ICFT as a robust and scalable solution for adapting LLMs to the medical domain, offering practical benefits for real-world healthcare applications.

Discussion (0). Continue with ORCID 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. SWE-Bench-CL: Continual Learning for Coding Agents

    cs.LG 2025-06 conditional novelty 5.0 of 10

    SWE-Bench-CL reorganizes SWE-Bench Verified into 8 time-ordered sequences of 273 total tasks to measure continual learning in coding agents, adding CL-specific metrics and a semantic memory agent design.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    CoRR abs/2310.07282 (2023)

    Sharaf, S., Anoop, V.S.: An analysis on large language models in healthcare: A case study of biobert. CoRR abs/2310.07282 (2023). https://doi.org/10.48550/ARXIV.2310.07282, https://doi.org/10.48550/arXiv.2310.07282

    arXiv doi:10.48550/arxiv.2310.07282 2023

  2. [2]

    arXiv preprint arXiv:2501.0137 7 (2025)

    Zhou, Y., Song, L., Shen, J.: Training medical large visio n-language models with abnormal-aware feedback. arXiv preprint arXiv:2501.0137 7 (2025)

    work page arXiv 2025

  3. [3]

    CoRR abs/2409.16860 (2024)

    Khalid, A., Khalid, A., Khalid, U.: The role of language mo d- els in modern healthcare: A comprehensive review. CoRR abs/2409.16860 (2024). https://doi.org/10.48550/ARXIV.2409.16860, https://doi.org/10.48550/arXiv.2409.16860

    arXiv doi:10.48550/arxiv.2409.16860 2024

  4. [4]

    In: The Thirteenth Intern ational Conference on Learning Representations (2024), https://openreview.net/pdf?id=N1vYivuSKq

    Zhou, Y., Shen, J., Cheng, Y.: Weak to strong generalizati on for large language models with multi-capabilities. In: The Thirteenth Intern ational Conference on Learning Representations (2024), https://openreview.net/pdf?id=N1vYivuSKq

    work page 2024

  5. [5]

    arXiv preprin t arXiv:2410.19732 (2024)

    Zhou, Y., Rao, Z., Wan, J., Shen, J.: Rethinking visual dep endency in long-context reasoning for large vision-language models. arXiv preprin t arXiv:2410.19732 (2024)

    arXiv 2024

  6. [6]

    In: Findings of the Association for Comput ational Linguistics, 12 R

    Zhou, Y., Li, X., Wang, Q., Shen, J.: Visual in-context lea rning for large vision- language models. In: Findings of the Association for Comput ational Linguistics, 12 R. Long et al. ACL 2024, Bangkok, Thailand and virtual meeting, August 11- 16, 2024. pp. 15890– 15902. Association for Computational Linguistics (2024)

    work page 2024

  7. [7]

    CoRR abs/2401.06775 (2024)

    Nazi, Z.A., Peng, W.: Large language models in health- care and medical domain: A review. CoRR abs/2401.06775 (2024). https://doi.org/10.48550/ARXIV.2401.06775, https://doi.org/10.48550/arXiv.2401.06775

    arXiv doi:10.48550/arxiv.2401.06775 2024

  8. [8]

    arXiv preprint arXiv:2404.03611 (2 024)

    Wang, Q., Wang, C., Lai, Z., Zhou, Y.: Insectmamba: Insect pest classification with state space model. arXiv preprint arXiv:2404.03611 (2 024)

    arXiv

Show all 33 references
  1. [9]

    Taylor, N., Schofield, D., Kormilitzin, A., Joyce, D.W., N evado-Holgado, A.J.: Developing healthcare language model embedding spac es. Artif. Intell. Medicine 158, 103009 (2024). https://doi.org/10.1016/J.ARTMED.2024.103009, https://doi.org/10.1016/j.artmed.2024.103009

    2024

  2. [10]

    Artificial Intel- ligence in Health 1(2), 16–28 (2024)

    Mumtaz, U., Ahmed, A., Mumtaz, S.: Llms-healthcare: Cur rent applications and challenges of large language models in various medical spec ialties. Artificial Intel- ligence in Health 1(2), 16–28 (2024)

    2024

  3. [11]

    arXiv preprint arXiv :2311.08734 (2023)

    Zhou, Y., Geng, X., Shen, T., Tao, C., Long, G., Lou, J.G., Shen, J.: Thread of thought unraveling chaotic contexts. arXiv preprint arXiv :2311.08734 (2023)

    2023 arXiv

  4. [12]

    In: ICASSP 2021-2021 IEEE Inter national Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Zhou, Y., Tao, W., Zhang, W.: Triple sequence generative adversarial nets for unsu- pervised image captioning. In: ICASSP 2021-2021 IEEE Inter national Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 75 98–7602. IEEE (2021)

    2021

  5. [13]

    CoRR abs/2306.07377 (2023)

    Nicholas, G., Bhatia, A.: Lost in translation: Large lan - guage models in non-english content analysis. CoRR abs/2306.07377 (2023). https://doi.org/10.48550/ARXIV.2306.07377, https://doi.org/10.48550/arXiv.2306.07377

    2023 doi arXiv

  6. [14]

    CoRR abs/2202.03371 (2022), https://arxiv.org/abs/2202.03371

    Müller, M., Laurent, F.: Cedille: A large autoregressiv e french language model. CoRR abs/2202.03371 (2022), https://arxiv.org/abs/2202.03371

    2022 arXiv

  7. [15]

    Ojo, J., Ogueji, K.: How good are commercial large langua ge models on african languages? In: Proceedings of the 4th Workshop on African Na tural Lan- guage Processing, AfricaNLP@ICLR 2023, Kigali, Rwanda, Ma y 1, 2023 (2023), https://openreview.net/pdf?id=MCgyGyRPEIU

    2023

  8. [16]

    CoRR abs/2408.10441 (2024)

    Chang, T.A., Arnett, C., Tu, Z., Bergen, B.K.: Gold- fish: Monolingual language models for 350 languages. CoRR abs/2408.10441 (2024). https://doi.org/10.48550/ARXIV.2408.10441, https://doi.org/10.48550/arXiv.2408.10441

    2024 doi arXiv

  9. [18]

    CoRR abs/2401.04155 (2024)

    Liu, J., Yang, M., Yu, Y., Xu, H., Li, K., Zhou, X.: Large la n- guage models in bioinformatics: applications and perspect ives. CoRR abs/2401.04155 (2024). https://doi.org/10.48550/ARXIV.2401.04155, https://doi.org/10.48550/arXiv.2401.04155

    2024 doi arXiv

  10. [19]

    In: Proceedings of the 60th Annual Meeting o f the Association for Computational Linguistics (Volume 1: Long Papers)

    Zhou, Y., Shen, T., Geng, X., Long, G., Jiang, D.: Claret: Pre-training a correlation-aware context-to-event transformer for even t-centric generation and classification. In: Proceedings of the 60th Annual Meeting o f the Association for Computational Linguistics (Volume 1: Lon...

    2022

  11. [20]

    In: Proceedings of the ACM Web Conference 2022

    Zhou, Y., Geng, X., Shen, T., Long, G., Jiang, D.: Eventbe rt: A pre-trained model for event correlation reasoning. In: Proceedings of the ACM Web Conference 2022. pp. 850–859 (2022) Title Suppressed Due to Excessive Length 13

    2022

  12. [21]

    ICT Express 5(3), 182–186 (2019)

    Spirov, A.V., Myasnikova, E.M.: Linguistic modelling o f gene regulation: Trans- lation from the language of experiments to the language of mo delling. ICT Express 5(3), 182–186 (2019). https://doi.org/10.1016/J.ICTE.2018.10.008, https://doi.org/10.1016/j.icte.2018.10.008

    2019 doi

  13. [22]

    CoRR abs/2303.00077 (2023)

    Houghton, C.J., Kazanina, N., Sukumaran, P.: Beyond the limitations of any imaginable mechanism: large language models and psycho linguistics. CoRR abs/2303.00077 (2023). https://doi.org/10.48550/ARXIV.2303.00077, https://doi.org/10.48550/arXiv.2303.00077

    2023 doi arXiv

  14. [23]

    CoRR abs/2205.07634 (2022)

    Veres, C.: A precis of language models are not models of la nguage. CoRR abs/2205.07634 (2022). https://doi.org/10.48550/ARXIV.2205.07634, https://doi.org/10.48550/arXiv.2205.07634

    2022 doi arXiv

  15. [24]

    CoRR abs/2408.15040 (2024)

    Ali, W., Pyysalo, S.: A survey of large language models fo r european languages. CoRR abs/2408.15040 (2024). https://doi.org/10.48550/ARXIV.2408.15040, https://doi.org/10.48550/arXiv.2408.15040

    2024 doi arXiv

  16. [25]

    CoRR abs/2407.16565 (2024)

    Buhnila, I., Sinha, A., Constant, M.: Retrieve, generat e, evaluate: A case study for medical paraphrases generation with small langua ge models. CoRR abs/2407.16565 (2024). https://doi.org/10.48550/ARXIV.2407.16565, https://doi.org/10.48550/arXiv.2407.16565

    2024 doi arXiv

  17. [26]

    CoRR abs/2410.23822 (2024)

    He, J., Li, P., Liu, G., Zhong, S.: Parameter-efficient fine -tuning medical multimodal large language models for medical visual ground ing. CoRR abs/2410.23822 (2024). https://doi.org/10.48550/ARXIV.2410.23822, https://doi.org/10.48550/arXiv.2410.23822

    2024 doi arXiv

  18. [27]

    CoRR abs/2405.18774 (2024)

    Ma, M., Yang, Y.: Llama-reg: Using llama 2 for unsu- pervised medical image registration. CoRR abs/2405.18774 (2024). https://doi.org/10.48550/ARXIV.2405.18774, https://doi.org/10.48550/arXiv.2405.18774

    2024 doi arXiv

  19. [28]

    CoRR abs/2410.20428 (2024)

    Zhang, H., An, B.: Medgo: A chinese medical large languag e model. CoRR abs/2410.20428 (2024). https://doi.org/10.48550/ARXIV.2410.20428, https://doi.org/10.48550/arXiv.2410.20428

    2024 doi arXiv

  20. [29]

    CoRR abs/2404.00578 (2024)

    Bai, F., Du, Y., Huang, T., Meng, M.Q., Zhao, B.: M3D: adva nc- ing 3d medical image analysis with multi-modal large langua ge models. CoRR abs/2404.00578 (2024). https://doi.org/10.48550/ARXIV.2404.00578, https://doi.org/10.48550/arXiv.2404.00578

    2024 doi arXiv

  21. [30]

    CoRR abs/2403.06611 (2024)

    Wu, J., Wu, X., Zheng, Y., Yang, J.: Medkp: Medical dialog ue with knowledge enhancement and clinical pathway encoding. CoRR abs/2403.06611 (2024). https://doi.org/10.48550/ARXIV.2403.06611, https://doi.org/10.48550/arXiv.2403.06611

    2024 doi arXiv

  22. [31]

    CoRR abs/2404.09027 (2024)

    Liao, Y., Jiang, S., Wang, Y., Wang, Y.: MING-MOE: en- hancing medical multi-task learning in large language mod- els with sparse mixture of low-rank adapter experts. CoRR abs/2404.09027 (2024). https://doi.org/10.48550/ARXIV.2404.09027, https://doi.org/10.48550/arXiv.2404.09027

    2024 doi arXiv

  23. [32]

    Co RR abs/2310.14151 (2023)

    Zhu, W., Wang, X., Zheng, H., Chen, M., Tang, B.: Promptcb lue: A chinese prompt tuning benchmark for the medical domain. Co RR abs/2310.14151 (2023). https://doi.org/10.48550/ARXIV.2310.14151, https://doi.org/10.48550/arXiv.2310.14151

    2023 doi arXiv

  24. [33]

    In: Demner-Fushman, D., A naniadou, S., Miwa, M., Roberts, K., Tsujii, J

    Kim, Y., Wu, J., Abdulle, Y., Wu, H.: Medexqa: Medical que stion answering bench- mark with multiple explanations. In: Demner-Fushman, D., A naniadou, S., Miwa, M., Roberts, K., Tsujii, J. (eds.) Proceedings of the 23rd Wo rkshop on Biomedi- cal Natural Language Processing, Bi...

    2024

  25. [34]

    CoRR abs/2405.10893 (2024)

    Panagoulias, D.P., Papatheodosiou, P., Palamidas, A.P ., Sanoudos, M., Tsoureli- Nikita, E., Virvou, M., Tsihrintzis, G.A.: Cognet-md, an ev aluation frame- work and dataset for large language model benchmarks in the m edical domain. CoRR abs/2405.10893 (2024). https://doi.org...

    2024 doi arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.