Pith. sign in

REVIEW 4 major objections 5 minor 15 references

Position: Open and Closed Large Language Models in Healthcare

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper argues that open and closed large language models have settled into complementary roles in healthcare: closed models such as GPT-4 dominate high-stakes diagnostic tasks like radiology and medical imaging, while open models such…

desk verdict The paper's descriptive dataset is useful, but its central claim about closed LLMs dominating high-performance healthcare rests on unnormalized topic counts and looks like a base-rate effect. read the letter →

arxiv 2501.09906 v1 pith:HZDRBU4D submitted 2025-01-17 cs.CY

Jiawei Xu , Ying Ding , Yi Bu This is my paper · ORCID
classification cs.CY
keywords largelanguagemodelsopen-sourceAIclosed-sourcehealthcaremedicalimagingmentalhealthtopicmodelingarXiv
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 seeks to establish a division of labor in medical AI: closed-weight language models are winning the high-performance, high-stakes tasks such as radiology and multimodal diagnostics, while open-weight models are powering specialized, cost-effective applications such as mental health support and patient communication. The evidence is a quantitative analysis of 6,198 arXiv papers that mention specific LLMs in their titles and abstracts, with a focus on 404 medical papers. Topic modeling shows that papers mentioning closed models concentrate on imaging and diagnostics, whereas papers mentioning open models concentrate on conversational and mental-health applications. If the pattern holds, it means a research community has sorted model choices by task demands: top reasoning where errors are costly, adaptability where personalization matters.

What carries the argument

The argument is carried by a quantitative comparison of 6,198 arXiv papers that mention specific open- or closed-weight LLMs in titles and abstracts, drawn from a foundation-model ecosystem graph that classifies each model's access type. From the 404 medical-related papers, BERTopic, a neural topic model using class-based TF-IDF, clusters research themes, and the resulting topic distributions for papers mentioning closed versus open models are compared. The load-bearing claim is that these mention distributions reflect how each type of model is actually being used in healthcare research.

What would settle it

A benchmark study that compared open- and closed-weight models on a standard radiology diagnostic suite and found open models matching or exceeding closed models on accuracy, or a survey of deployed clinical systems showing open models widely used in imaging, would undermine the paper's central complementarity claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that the scientific community's use of LLMs in healthcare has split along the open/closed axis into complementary roles. Closed-weight models, exemplified by GPT-4, lead in high-complexity, high-stakes tasks such as radiology and medical imaging, where strong reasoning and precision are paramount. Open-weight models, exemplified by the LLaMA series, are the tools of choice for specialized, lower-cost applications such as mental health support, conversational agents, and patient communication, because their released weights allow fine-tuning on targeted datasets. This complementarity, the paper argues, is visible in the topics of medical arXiv papers: those naming closed models concentrate on diagnostic imaging, while those naming open models concentrate on personalized, interactive care.

Load-bearing premise

The analysis treats a model being named in an arXiv paper's title or abstract as evidence that the model is actually used and valued in that healthcare area; if researchers mention closed models mainly because they are well-known, or report clinical work outside preprint servers, the claimed division of labor is not established.

Editorial extensions

If this is right

  • If the division of labor is real, closed models will remain the default tools for radiology and other high-accuracy diagnostic tasks as long as they lead in reasoning ability.
  • Open models will continue to proliferate in niche healthcare areas because released weights make fine-tuning cheap, which was the driver of their exponential growth since LLaMA's release.
  • The complementary pattern points toward hybrid pipelines that pair a closed model's diagnostic reasoning with open models customized for patient-facing communication.
  • Research attention is now concentrated differently: closed models draw more overall scientific interest, especially in fields that use rather than build models, while open models dominate in machine-learning subfields that optimize training and adaptation.

Reading between the lines

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

  • A testable extension: measure actual model use in clinical deployments or code and API logs rather than paper mentions; the mention-based pattern may overstate closed models' real-world share in diagnostics.
  • If the boundary is set by task stakes rather than capability ceiling, improvements in open models will shrink but not erase the closed-model niche, since closed deployment also offers accountability and data-handling assurances that matter in clinical settings.
  • The same complementarity may apply beyond healthcare, in regulated domains such as legal advice or finance, where closed models handle high-stakes analysis and open models power accessible customer-facing tools.
  • The paper's 'complementary roles' framing suggests a resourcing implication: institutions may reasonably support both open and closed routes rather than treating open source as a universal substitute.
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 / 5 minor

Summary. This position paper compares how open-source and closed-source large language models are discussed in the healthcare literature. Using the Stanford CRFM Ecosystem Graph for model counts and arXiv metadata for papers mentioning specific model names, the authors report that closed LLMs (e.g., GPT-4) receive more scientific attention overall and, in the medical subset, appear predominantly in papers on radiology and medical imaging, while open LLMs (e.g., LLaMA) appear more in papers on mental health and patient communication. From these patterns, they conclude that closed models lead high-performance diagnostic applications while open models enable specialized fine-tuned applications, and they suggest hybrid approaches for the future.

Significance. The question the paper addresses—whether open and closed LLMs occupy complementary roles in healthcare—is timely and of genuine interest to the NeurIPS community. The descriptive dataset assembled from 6,198 arXiv papers and 201 models is a useful resource, and the paper is commendably transparent about its retrieval strategy and data sources. However, the central qualitative claim currently rests on unnormalized topic counts and unvalidated topic labels, so the significance of the specific complementary-roles conclusion is not yet established. The paper reads more as a set of hypotheses grounded in a preliminary descriptive analysis than as a demonstrated result, and the analyses would need to be strengthened before the conclusions can be relied upon.

major comments (4)
  1. [Section 3, Figures 3 and 4] The key claim that closed LLMs dominate high-performance tasks such as radiology and medical imaging while open LLMs dominate mental health and patient communication is not supported by the analysis as presented. Figure 4 reports raw topic concentrations within each model class without normalizing for the substantially different base rates: Figures 2 and 3 show that closed models have many more total medical mentions. Under the null model that researchers mention whichever model is more visible or more capable uniformly across all topics, closed models would appear more often in every high-volume specialty purely because of their higher base rate. To support the complementary-roles claim, the authors should report topic proportions within each class relative to a baseline (e.g., the model class share of all medical-LLM mentions), or perform a statistical contrast such as a chi-square or log-odds ratio per topic that accounts for the overall popularity difference.
  2. [Section 3, Figure 1] The assertion that the number of open LLMs grows exponentially while the number of closed LLMs increases linearly is made on the basis of visual inspection of Figure 1, with no model fitting, no growth-rate estimate, and no uncertainty quantification. Since this claim is used to motivate the 'democratization' narrative, it is load-bearing and should be substantiated. For example, the authors could fit count models (e.g., Poisson or negative binomial regressions with a log link) to the annual model counts and report rate ratios and confidence intervals, or explicitly state that the distinction is qualitative and based on visual inspection rather than a fitted trend.
  3. [Sections 2.2 and 3] The paper uses the presence of a model name in titles and abstracts as a proxy for 'using, evaluating, or mentioning' the model, and then interprets topic co-occurrence as evidence of actual use and fine-tuning. In particular, the claim that open LLMs 'enable researchers to fine-tune models for specific domains, such as mental health and patient communication' is not verified anywhere: a mental-health paper that mentions LLaMA may be evaluating the base model, comparing it with GPT-4, or mentioning it only in a literature-review sentence. The analysis should either include a validation step that checks for fine-tuning-related terms (e.g., 'fine-tun', 'LoRA', 'instruction-tuned') in the open-model medical papers, or the conclusions should be weakened to state that open models receive attention in these topic areas without claiming that the papers actually fine-tune them.
  4. [Section 3, BERTopic analysis] The BERTopic labels in Figure 4 are presented without any validation. No topic coherence scores, manual inspection details, representative terms per topic, or inter-annotator agreement are reported. Because the topic labels are the direct evidence for the paper's central healthcare claim, the authors should at least show the top terms defining each topic cluster and report a validation measure (e.g., topic coherence or a manual labeling exercise), so that a reader can judge whether labels such as 'radiology and medical imaging' accurately reflect the underlying documents.
minor comments (5)
  1. [Abstract] The abstract states that the paper 'analyzes the evolving roles' and reaches firm conclusions, but the evidence is only descriptive. Consider adding a phrase such as 'suggest' or 'provide evidence for' to align the abstract with the strength of the analysis.
  2. [Table 1] The notation '147 (103)' and '54 (44)' is not explained. Clarify that the first number is the total identified models and the second is the number matched in arXiv titles/abstracts.
  3. [Figure 3] The left panel is labeled 'cumulative ratio' but the denominator is not defined in the text. Specify whether it is the ratio of medical LLM papers to all LLM papers or to all medical papers, and state the base population.
  4. [References] Some references are incomplete: the GPT-4 technical report [9] lacks a version or date beyond '2023', and reference [5] (Vicuna blog) would benefit from a stable URL or a DOI if available.
  5. [Section 2.2] The paper uses only arXiv data, which is a significant limitation for healthcare research, where much applied work is published in clinical journals not indexed by arXiv. This limitation is not acknowledged; at minimum, it should be stated in Section 2.2 or Section 4.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a descriptive topic-analysis position piece, and its conclusions are interpretive rather than definitionally forced.

full rationale

The paper does not contain a derivation chain in the sense of equations or fitted parameters that are later renamed as predictions. Its evidence consists of arXiv mention counts (Section 2.2) and a BERTopic topic model over medical LLM papers (Section 3, Figure 4). The central claim that closed LLMs dominate high-performance tasks such as radiology while open LLMs support fine-tuned applications such as mental health is an interpretation of topic distributions, not a quantity that those distributions define by construction. No self-citation is load-bearing: the external sources (CRFM ecosystem graph, BERTopic, model papers) are used as data and tools, not as authority for the paper's main claim. The main weaknesses are inferential: mention counts are used as a proxy for use and capability, and the topic patterns are not normalized for the larger overall attention received by closed models. These are concerns about validity and confounds, not circularity. There is no equation or construction by which the conclusion equals the input, so the appropriate finding is no significant circularity.

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

The paper does not introduce new parameters or entities. Its conclusions rest on external datasets, a citation-matching heuristic, and an interpretive link from publication topics to capability, which are the main hidden assumptions.

assumptions (4)
  • domain assumption arXiv titles/abstracts mentioning specific LLM names serve as a valid proxy for scientific community attention and model usage.
    Used throughout Section 2.2 and Section 3 to measure popularity and infer application areas; this assumption excludes non-arXiv and non-title/abstract mentions.
  • domain assumption The CRFM Ecosystem Graph for Foundation Models provides a complete and accurate list of significant LLMs with correct open/closed labels as of April 2024.
    The entire model census in Section 2.1 derives from this external dataset.
  • domain assumption BERTopic topic clusters correspond to meaningful research directions that can be labeled as radiology, mental health, etc.
    Topic labels in Figure 4 are presented without validation details such as coherence scores or manual agreement.
  • domain assumption Publication counts in medical arXiv papers can be interpreted as evidence of performance dominance for closed models.
    The paper infers superior reasoning and high-stakes application leadership from topic prevalence rather than benchmark results (Section 3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Position: Open and Closed Large Language Models in Healthcare." pith.science (2026). https://pith.science/paper/HZDRBU4D

@misc{pith2026250109906,
  author       = {Pith},
  title        = {Pith review of: Position: Open and Closed Large Language Models in Healthcare},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZDRBU4D}},
  note         = {Machine review of arXiv:2501.09906}
}
read the original abstract

This position paper analyzes the evolving roles of open-source and closed-source large language models (LLMs) in healthcare, emphasizing their distinct contributions and the scientific community's response to their development. Due to their advanced reasoning capabilities, closed LLMs, such as GPT-4, have dominated high-performance applications, particularly in medical imaging and multimodal diagnostics. Conversely, open LLMs, like Meta's LLaMA, have gained popularity for their adaptability and cost-effectiveness, enabling researchers to fine-tune models for specific domains, such as mental health and patient communication.

Figures

Figures reproduced from arXiv: 2501.09906 by the authors.

Figure 1
Figure 1. Growth rates of open and closed LLMs. grown exponentially, with those fine-tuned on existing models (orange bars) contributing significantly to this trend. The growth rate of closed models initially outpaced open models until 2023, after which open models surged, primarily driven by fine-tuning efforts. Closed LLMs gained prominence in June 2020 with the release of GPT-3 [4] by OpenAI. With its massive size (175 B p… view at source ↗
Figure 2
Figure 2. Percentage of arXiv papers mentioning open LLMs or closed LLMs from 2019 onwards, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Cumulative ratio (left) and counts (right) of medical papers in LLMs. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Topic modeling results for medical LLM papers. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 7 canonical work pages

  1. [1]

    Llama 2: Open foundation and fine-tuned chat models

    Meta AI. Llama 2: Open foundation and fine-tuned chat models. ArXiv, abs/2307.09288, 2023

    arXiv 2023

  2. [2]

    Jiang, Jia Deng, Stella Biderman, and Sean Welleck

    Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen Marcus McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics. ArXiv, abs/2310.10631, 2023

    arXiv

  3. [3]

    Creel, and Percy Liang

    Rishi Bommasani, Dilara Soylu, Thomas Liao, Kathleen A. Creel, and Percy Liang. Ecosystem graphs: The social footprint of foundation models. ArXiv, abs/2303.15772, 2023

    arXiv 2023

  4. [4]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz Litwin,...

    arXiv 2005

  5. [5]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, march 2023

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, march 2023. URL https://lmsys. org/blog/2023-03- 30-vicuna, 3(5), 2023

    work page 2023

  6. [6]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics , 2019

    work page 2019

  7. [7]

    Palm: Scaling language modeling with pathways

    Google. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res. , 24:240:1– 240:113, 2022

    work page 2022

  8. [8]

    Bertopic: Neural topic modeling with a class-based tf-idf procedure

    Maarten Grootendorst. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794, 2022

    arXiv 2022

Show all 15 references
  1. [9]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. ArXiv, 2023

    2023

  2. [10]

    Language models are unsupervised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. ArXiv, 2019

    2019

  3. [11]

    Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J

    Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1–140:67, 2019

    2019

  4. [12]

    Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model

    Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Anand Korthikanti, Elton Zhang, Rewon Child, Reza Yazdani Aminabadi, Julie Bernauer, Xia Song, Mohammad Shoeybi, Yuxiong He, ...

    2022 arXiv

  5. [13]

    The gradient of generative ai release: Methods and considerations

    Irene Solaiman. The gradient of generative ai release: Methods and considerations. In Proceed- ings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , FAccT ’23, pages 111–122. Association for Computing Machinery, 2023

    2023

  6. [14]

    Alpaca: A strong, replicable instruction- following model

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpaca: A strong, replicable instruction- following model. Stanford Center for Research on F oundation Models. https://crfm. stanford. edu/2023/03/13/...

    2023

  7. [15]

    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, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation langua...

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.