Pith. sign in

REVIEW 4 major objections 5 minor 5 cited by

A Novel Evaluation Benchmark for Medical LLMs: Illuminating Safety and Effectiveness in Clinical Domains

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 30-criterion clinical benchmark scores six LLMs at 57.2% overall, with safety lagging effectiveness.

desk verdict A genuine Chinese medical LLM benchmark whose headline rankings sit on an unvalidated LLM judge that is also a contestant. read the letter →

arxiv 2507.23486 v3 pith:AMHKRN5U submitted 2025-07-31 cs.CL

Shirui Wang , Zhihui Tang , Huaxia Yang , Qiuhong Gong , Tiantian Gu , Hongyang Ma , Yongxin Wang , Wubin Sun
show 30 more authors
Zeliang Lian Kehang Mao Yinan Jiang Zhicheng Huang Lingyun Ma Wenjie Shen Yajie Ji Yunhui Tan Chunbo Wang Yunlu Gao Qianling Ye Rui Lin Mingyu Chen Lijuan Niu Zhihao Wang Peng Yu Mengran Lang Yue Liu Huimin Zhang Haitao Shen Long Chen Qiguang Zhao Si-Xuan Liu Lina Zhou Hua Gao Dongqiang Ye Lingmin Meng Youtao Yu Naixin Liang Jianxiong Wu
This is my paper
classification cs.CL
keywords largelanguagemodelsclinicalsafetyevaluationeffectivenessbenchmarkmedicalLLMLLM-as-judgerisk-weightedscoringdecisionsupport
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 claims that the clinical readiness of large language models can be measured on two tracks, safety and effectiveness, through a benchmark called CSEDB built from 2,069 open-ended questions spanning 26 departments and 30 expert-weighted criteria. When six LLMs were tested, the average total score was 57.2%, safety averaged 54.7%, effectiveness averaged 62.3%, and all models dropped 13.3% in high-risk scenarios. The authors take these numbers as evidence that current models are not yet safe enough for high-risk clinical tasks, and that a domain-specific medical model can substantially outperform general-purpose ones. If the benchmark is valid, it gives hospitals and regulators a standardized, risk-weighted yardstick for comparing models and for deciding where model use is unsafe.

What carries the argument

The load-bearing mechanism is a rubric-based open-ended QA pipeline: each of the 30 criteria has either a binary safe-versus-unsafe rule or a graded multi-rule scoring formula, and each question's score is multiplied by a clinical-severity weight before aggregation. An automated LLM judge applies the rubrics to each question, model response, and reference answer, then normalizes the result to a 0–1 scale. Because all model rankings and the headline percentages flow through this judge, the reliability checks—Worst-at-k stability and Macro-F1 agreement with clinician scoring—are what carry the argument.

What would settle it

Ask independent clinicians to hand-score a stratified sample of 200 CSEDB responses from non-oncology departments and compare those scores with the automated judge's; if agreement drops materially below the oncology Macro-F1, the benchmark's cross-specialty rankings are not stable.

Watch

Extended reading notes

Core claim

The paper establishes CSEDB as a dual-track evaluation: 17 safety metrics covering critical-illness recognition and medication safety, and 13 effectiveness metrics covering guideline adherence and diagnostic or therapeutic optimization, with each question assigned a 1–5 clinical-risk weight. Thirty-two specialist physicians authored and reviewed the open-ended questions, and an automated judge scored responses against rubric-based reference answers. On this benchmark, six LLMs averaged 57.2% total, with safety (54.7%) lower than effectiveness (62.3%); the significant 13.3% drop on high-risk items was consistent across models. The paper also reports that the domain-specific medical model scored highest on both tracks, while general-purpose models clustered together at moderate levels.

Load-bearing premise

The benchmark's scores are only as trustworthy as the automated judge's ability to replicate clinical expert judgments on all 30 criteria, yet the paper validates that judge on 411 oncology items only, and there its Macro-F1 (0.601) falls slightly below the human-consensus baseline (0.625).

Editorial extensions

If this is right

  • General-purpose LLMs should not be trusted for high-risk clinical decisions, since every tested model lost about 13.3% of its score when case severity rose.
  • Safety and effectiveness should be reported as separate numbers rather than one blended score, because safety averaged 54.7% versus 62.3% for effectiveness and the gap was consistent.
  • A domain-specific medical model can be much safer than general-purpose models, but the advantage is not uniform across departments or patient populations, so model selection must be contextual.
  • Structured system prompts can raise both safety and effectiveness scores significantly, giving a low-cost way to improve deployed models before retraining.
  • Weight-stratified reporting makes weak spots visible at the criterion level, such as combination-therapy decisions, converting a single benchmark score into a targeted improvement list.

Reading between the lines

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

  • If CSEDB becomes a standard tool, its current numbers should be treated as provisional outside oncology because judge agreement with physicians was tested on only 411 oncology criteria.
  • The universal 13.3% high-risk drop implies a practical deployment rule: route high-severity queries away from automated answers until a model demonstrates stability on high-weight items.
  • The 30 shared criteria could be re-used to build specialty-specific sub-benchmarks, turning the framework into a diagnostic map rather than one aggregate grade.
  • Because the benchmark is single-turn and text-only, real-world multi-turn conversations could reveal an even wider safety-effectiveness gap than the paper reports.
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. The paper introduces CSEDB, a two-dimensional benchmark for medical LLMs that scores safety (17 indicators) and effectiveness (13 indicators) across 2,069 open-ended clinical questions spanning 26 departments, with risk weights 1–5 assigned by expert consensus. Six LLMs (Deepseek-R1-0528, OpenAI-o3, Gemini-2.5-Pro, Qwen3-235B-A22B, Claude-3.7-Sonnet, and MedGPT) were evaluated using Deepseek-R1 as an automated LLM-as-a-judge. The headline findings are an average total score of 57.2%, safety 54.7%, effectiveness 62.3%, a 13.3% performance drop in high-risk scenarios (p < 0.0001), and consistent superiority of the domain-specific model MedGPT. The authors also report repeatability (Worst-at-k), inter-doctor consistency on 411 oncology criteria, and a prompt-engineering improvement study.

Significance. If the automated scoring is valid, CSEDB would be a useful contribution: it is a large, clinically motivated, risk-weighted, open-ended benchmark with released data and code, and the two-dimensional safety/effectiveness decomposition addresses a real gap in medical LLM evaluation. The bootstrap-based inference and the attempt to calibrate the judge against human oncologists are also strengths. However, the central quantitative claims—the aggregate scores, the model rankings, and the high-risk performance drop—all pass through a single LLM judge whose validity is demonstrated only on a narrow slice of the data. The significance of the empirical results is therefore conditional on evidence that the authors have not yet supplied.

major comments (4)
  1. [Methods 3.2, Methods 3.9] The calibration of the Deepseek-R1 scoring engine is described only qualitatively. Section 3.2 states that agreement with human scores was assessed using the Kappa coefficient and prompt rules were iteratively adjusted "until the consistency reached a predefined threshold," but neither the threshold nor the achieved Kappa values are reported. Section 3.9 gives formulas for agreement on 4,303 doctor-reviewed rules but presents no numerical results. Because every model score, ranking, and the 13.3% high-risk drop are produced by this judge, the absence of quantitative evidence of judge accuracy is a load-bearing gap that must be filled.
  2. [Section 2.5.2, Figure 5] The only expert-consistency validation reported is on 411 oncology criteria, where Deepseek-R1 achieves macro-F1 0.601 against a group-consensus baseline of 0.625. Yet the judge is applied to all 2,069 cases, 30 criteria, and 26 departments, including obstetrics, pediatrics, psychiatry, and pharmacy. The manuscript provides no evidence that judge accuracy generalizes across departments, risk levels, or criterion types. The authors should report judge agreement stratified by department, risk weight, and binary vs. graded criteria, or otherwise justify why the oncology result transfers to the entire benchmark.
  3. [Section 2.2 and Methods 3.2] Deepseek-R1-0528 is both the automated judge and one of the six evaluated models. This creates a plausible self-scoring path: the judge may be more lenient toward its own output style, inflating its scores and shifting all pairwise comparisons and aggregate rankings. The manuscript does not address this confound. A sensitivity analysis with an independent judge (e.g., a second LLM or a subset of human-scored cases) or an explicit analysis of judge bias toward Deepseek-R1 responses is needed before the comparative claims can be accepted.
  4. [Section 3.6.2] The description of error bars is internally unclear: the text says error bars are derived from bootstrap resamples but then states they are "computed as the standard deviation of the bootstrap resamples divided by the square root of the sample size," which is a standard error formula, not a bootstrap confidence interval. Since the paper reports 95% confidence intervals and bootstrap p-values, the precise construction of the intervals should be stated consistently, and the figures should indicate whether the intervals are percentile-based or normal-approximation intervals.
minor comments (5)
  1. [Section 2.5.2] The notation "Deepseek-R1 (M2)" is confusing because M1–M5 are described as human oncologists; it should be made explicit that M2 refers to the LLM judge or the physician labels should be relabeled to avoid ambiguity.
  2. [Abstract and Section 2.2] The abstract states that "domain-specific medical LLMs" showed consistent advantages, but only one domain-specific model (MedGPT) was evaluated; the plural is misleading.
  3. [Introduction, reference [17]] The text says the work builds on "OpenAI's healthBench," but reference [17] is Arora et al., HealthBench, which is not an OpenAI product; the attribution should be corrected.
  4. [Methods 3.2] The calibration subset is not described in terms of size or composition; please specify how many cases were used, how they were sampled, and whether they overlapped with the 411 oncology validation set.
  5. [Section 5, Data Availability] The paper refers to Supplementary Tables S1–S10 and an appendix table but does not include them in the manuscript; the repository link should be checked to ensure all cited tables are actually present and accessible.

Circularity Check

1 steps flagged · score 6.0 of 10

Deepseek-R1 is used as both the automated scoring engine and one of the six benchmarked models, so the reported LLM scores and rankings are partly produced by the judge scoring itself; the rest of the benchmark construction is externally grounded in physician rubrics.

  1. self definitional [Methods 3.2 (Model Auto-Scoring Evaluation) and Results 2.2 (test model list)]
    "we adopted the “LLM-as-Judge” paradigm [26,27], utilizing a commercial large language model, Deepseek-R1, to construct the automated scoring engine. ... we employed Deepseek-R1-0528, OpenAI-o3 (20250416), Gemini-2.5-Pro (20250506), Qwen3-235B-A22B, Claude-3.7-Sonnet (20250219) and MedGPT (MG-0623, Medlinker) as the test models."

    Deepseek-R1 is simultaneously the automated scoring engine and one of the six models being scored. Consequently, every reported Deepseek-R1 score, and every ranking or aggregate that depends on it, is produced by Deepseek-R1 judging Deepseek-R1's own outputs. The measurement is not independent: the reported benchmark score of the judge model reduces, by construction, to the judge's self-assessment under its own rubric, so any comparison between Deepseek-R1 and the other five models is confounded by self-scoring. The paper's oncology validation (MF1 0.601 vs group consensus 0.625) shows rough alignment with human judges but does not measure or remove self-preference; no analysis of judge-model identity effects is reported.

full rationale

The benchmark artifact itself is not circular: the 2,069 questions, 30 criteria, rubrics, and risk weights were constructed and reviewed by 32 specialist physicians and an expert committee, so the central measurement target is externally grounded. However, the entire scoring layer is Deepseek-R1, and Deepseek-R1 is also one of the six models whose scores are reported. The resulting rankings, including the 57.2% average, the safety/effectiveness split, and the 13.3% high-risk drop, are therefore not independent measurements for the judge model and are confounded for all models by the judge's preferences. The oncology-only validation (MF1 0.601 vs group consensus 0.625) and the unreported Kappa threshold in Methods 3.2 are validity gaps rather than circularity per se, and no self-citation chains or imported uniqueness theorems appear. On balance, one structural self-scoring circularity affects the central comparative claim, while the benchmark construction retains substantial independent content, supporting a partial-circularity score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The benchmark is an evaluation method, not a theoretical derivation, so the ledger is dominated by human-chosen weights and domain assumptions rather than invented physical entities. The most important non-independent element is the LLM-as-judge itself, which is both a measurement instrument and one of the measured objects.

free parameters (3)
  • Risk weights (1-5) assigned to each of the 30 criteria = Integer weights 1-5, set by expert consensus
    Chosen by hand through three rounds of expert consultation; the weighted total score formula (Section 3.4) depends directly on these weights, and changing them changes every model ranking.
  • Effectiveness gate sub-weights (70%-20%-10%) = 70%-20%-10% for high-value diagnostic decisions, intermediate management, and patient experience
    Introduced in Methods Section 3.1 as a weighting scheme for effectiveness metrics; no empirical or external validation of the chosen split is provided.
  • Graded scoring rule weights within individual cases = Varies per case, e.g., [5,4,3] in the CKD medication example
    Predefined by experts for each graded-scoring item; the ScoreGraded formula in Section 3.3.2 sums these weights, so they directly determine case-level scores.
assumptions (4)
  • domain assumption The 30 criteria and reference answers produced by 32 physicians are a complete and correct gold standard for clinical safety and effectiveness.
    The whole benchmark depends on this; no inter-rater reliability among the 32 physicians is reported, so the stability of the gold standard is unknown.
  • domain assumption Deepseek-R1's automated scores approximate expert clinical judgment across all 30 criteria.
    Validity is checked only on an oncology subset (MF1 0.601 versus group consensus 0.625); the Kappa calibration threshold and its measured value are not reported (Methods 3.2, 3.9).
  • domain assumption Single-turn open-ended QA approximates real clinical decision contexts.
    The paper itself acknowledges in the Discussion that real clinical communication is multi-turn, so single-turn scores may underestimate or miss real-world biases.
  • domain assumption All six models received equivalent evaluation conditions.
    Temperature was set to 1.0 for all and default prompts were used, but the prompt-engineering experiment shows scores are sensitive to prompt design; the standard prompt itself is not fully shown in the main text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Evaluation Benchmark for Medical LLMs: Illuminating Safety and Effectiveness in Clinical Domains." pith.science (2026). https://pith.science/paper/AMHKRN5U

@misc{pith2026250723486,
  author       = {Pith},
  title        = {Pith review of: A Novel Evaluation Benchmark for Medical LLMs: Illuminating Safety and Effectiveness in Clinical Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AMHKRN5U}},
  note         = {Machine review of arXiv:2507.23486}
}
read the original abstract

Large language models (LLMs) hold promise in clinical decision support but face major challenges in safety evaluation and effectiveness validation. We developed the Clinical Safety-Effectiveness Dual-Track Benchmark (CSEDB), a multidimensional framework built on clinical expert consensus, encompassing 30 criteria covering critical areas like critical illness recognition, guideline adherence, and medication safety, with weighted consequence measures. Thirty-two specialist physicians developed and reviewed 2,069 open-ended Q&A items aligned with these criteria, spanning 26 clinical departments to simulate real-world scenarios. Benchmark testing of six LLMs revealed moderate overall performance (average total score 57.2%, safety 54.7%, effectiveness 62.3%), with a significant 13.3% performance drop in high-risk scenarios (p < 0.0001). Domain-specific medical LLMs showed consistent performance advantages over general-purpose models, with relatively higher top scores in safety (0.912) and effectiveness (0.861). The findings of this study not only provide a standardized metric for evaluating the clinical application of medical LLMs, facilitating comparative analyses, risk exposure identification, and improvement directions across different scenarios, but also hold the potential to promote safer and more effective deployment of large language models in healthcare environments.

Figures

Figures reproduced from arXiv: 2507.23486 by the authors.

Figure 1
Figure 1. Overall research design workflow. 2.2 Core Performance Comparison: Overall Model Scores on Safety and Effectiveness To investigate the performance of various LLMs on CSEDB framework, we employed Deepseek￾R1-0528, OpenAI-o3 (20250416), Gemini-2.5-Pro (20250506), Qwen3-235B-A22B, Claude-3.7- Sonnet (20250219) and MedGPT (MG-0623, Medlinker) as the test models. All evaluations 5 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Comparative Performance of Models across safety and effectiveness gates. A. LLMs performance comparison across three evaluation metrics. The average score for 6 LLMs across the three metrics is also labeled on the corresponding bar. Error bars represent the 95% weighted bootstrap confidence intervals. P-values are derived from weighted bootstrap tests for all pairwise comparisons, adjusted using the Holm correction.… view at source ↗
Figure 3
Figure 3. Comparison of LLM performance based on weighted categories. A. LLMs performance comparison by weight categories. Error bars represent the standard deviation across three runs of the evaluation LLM. B. LLMs performance comparison between normal and high￾risk scenarios. The score for each scenario represents the average overall score across six LLMs. P-values are derived from bootstrap tests for all pairwise compariso… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of LLM performance across different departments and pop￾ulations. Safety and effectiveness score are calculated by different departments (A) and pop￾ulations (B) for each LLM individually. The abbreviations for 26 clinical departments are as follows: Cardiol…
Figure 5
Figure 5. Figure 5: Evaluating the trustworthiness of model grading. Worst-at-k performance for various LLM models, up to k=10. A.The Worst-at-k metric quantifies model stability by estimating the expected worst-case performance when selecting k responses, where lower scores indicate high…
Figure 6
Figure 6. Figure 6: Comparison of LLM performance by safety and effectiveness score. Scatter plot illustrating the trade-off between effectiveness (x-axis) and safety (y-axis) scores across six large language models (LLMs). Each point represents a model, with the values in parentheses ind…
Figure 7
Figure 7. Figure 7: Comparison of safety (left) and effectiveness (right) score before and after prompt engineering optimization. P-values are derived from weighted, bootstrap tests for all pairwise comparisons, adjusted using the Holm correction. ** p ≤ 0.01 ; * p ≤ 0.05 16 [PITH_FULL_I…

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures

    cs.AI 2026-04 conditional novelty 7.0 of 10

    Frontier models systematically withhold clinically necessary guidance from laypeople that they provide to physicians on identical facts, and LLM judges fail to detect that omission harm.

  2. MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    MedVIGIL introduces a clinician-supervised benchmark showing medical VLMs frequently give fluent answers on broken visual evidence, with top models 14 points below human radiologists on the composite score.

  3. MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    MedVIGIL provides a 300-case evaluation suite with 2556 probes that measures silent failures in medical VLMs under broken evidence, showing the best model at 69.2 on the composite score versus a human radiologist at 83.3.

  4. IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures

    cs.AI 2026-04 unverdicted novelty 6.0 of 10

    AI models exhibit identity-contingent withholding, providing better clinical guidance on benzodiazepine tapering to physicians than laypeople in identical scenarios, with a measured decoupling gap of +0.38 and 13.1 pe...

  5. Deployment-Centered Evaluation: Predicting Query-Level Rejection Risk in a Clinical LLM System

    cs.AI 2026-06 unverdicted novelty 5.0 of 10

    A pre-response classifier predicts user rejection risk for clinical LLM outputs with AUROC 0.719 over 4.5 months of deployment data by incorporating deployment-specific context.

Reference graph

Works this paper leans on

32 extracted references · 8 canonical work pages · cited by 3 Pith papers

  1. [1]

    A., Gui, H., Rezaei, S

    Omiye, J. A., Gui, H., Rezaei, S. J., Zou, J. & Daneshjou, R. Large Language Models in Medicine: The Potentials and Pitfalls : A Narrative Review. Ann Intern Med177, 210-220, doi:10.7326/m23-2772 (2024)

    doi:10.7326/m23-2772 2024

  2. [2]

    et al.Towards accurate differential diagnosis with large language models.Nature 642, 451-457, doi:10.1038/s41586-025-08869-4 (2025)

    McDuff, D. et al.Towards accurate differential diagnosis with large language models.Nature 642, 451-457, doi:10.1038/s41586-025-08869-4 (2025)

    doi:10.1038/s41586-025-08869-4 2025

  3. [3]

    Bedi, S. et al. Testing and Evaluation of Health Care Applications of Large Language Models: A Systematic Review. Jama 333, 319-328, doi:10.1001/jama.2024.21700 (2025)

    arXiv 2025

  4. [4]

    et al.Foundation models for generalist medical artificial intelligence

    Moor, M. et al.Foundation models for generalist medical artificial intelligence. Nature 616, 259-265, doi:10.1038/s41586-023-05881-4 (2023)

    doi:10.1038/s41586-023-05881-4 2023

  5. [5]

    Tordjman, M. et al. Comparative benchmarking of the DeepSeek large language model on medical tasks and clinical reasoning. Nat Med, doi:10.1038/s41591-025-03726-3 (2025)

    doi:10.1038/s41591-025-03726-3 2025

  6. [6]

    Dada, A. et al. MeDiSumQA: Patient-Oriented Question-Answer Generation from Dis- charge Letters. arXiv e-prints, doi:10.48550/arXiv.2502.03298 (2025)

    work page Pith review arXiv doi:10.48550/arxiv.2502.03298 2025

  7. [7]

    et al.Adapted large language models can outperform medical experts in clin- ical text summarization

    Van Veen, D. et al.Adapted large language models can outperform medical experts in clin- ical text summarization. Nat Med 30, 1134-1142, doi:10.1038/s41591-024-02855-5 (2024)

    doi:10.1038/s41591-024-02855-5 2024

  8. [8]

    Ive, J. et al. Clean & Clear: Feasibility of Safe LLM Clinical Guidance. arXiv:2503.20953 (2025). <https://ui.adsabs.harvard.edu/abs/2025arXiv250320953I>. 26

    work page Pith review arXiv 2025

Show all 32 references
  1. [9]

    de Hond, A. et al. From text to treatment: the crucial role of validation for generative large language models in health care. Lancet Digit Health6, e441-e443, doi:10.1016/s2589- 7500(24)00111-0 (2024)

    2024 doi

  2. [10]

    Hager, P. et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat Med 30, 2613-2622, doi:10.1038/s41591-024-03097-1 (2024)

    2024 doi

  3. [11]

    & Cho, B

    Lee, J., Park, S., Shin, J. & Cho, B. Analyzing evaluation methods for large language models in the medical field: a scoping review. BMC Med Inform Decis Mak 24, 366, doi:10.1186/s12911-024-02709-7 (2024)

    2024 doi

  4. [12]

    Liu, M. et al. MedBench: A Comprehensive, Standardized, and Reliable Benchmarking System for Evaluating Chinese Medical Large Language Models. arXiv:2407.10990 (2024). <https://ui.adsabs.harvard.edu/abs/2024arXiv240710990L>

    2024 arXiv

  5. [13]

    Ying, Z. et al. SafeBench: A Safety Evaluation Framework for Multimodal Large Language Models. arXiv:2410.18927 (2024). <https://ui.adsabs.harvard.edu/abs/ 2024arXiv241018927Y>

    2024 arXiv

  6. [14]

    et al.Agent-SafetyBench: Evaluating the Safety of LLM Agents

    Zhang, Z. et al.Agent-SafetyBench: Evaluating the Safety of LLM Agents. arXiv:2412.14470 (2024). <https://ui.adsabs.harvard.edu/abs/2024arXiv241214470Z>

    2024 arXiv

  7. [15]

    Deniz, F. et al. aiXamine: Simplified LLM Safety and Security. arXiv:2504.14985 (2025). <https://ui.adsabs.harvard.edu/abs/2025arXiv250414985D>

    2025 arXiv

  8. [16]

    Gaber, F. et al. Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis. NPJ Digit Med 8, 263, doi:10.1038/s41746-025-01684-1 (2025)

    2025 doi

  9. [17]

    Arora, R. K. et al. HealthBench: Evaluating Large Language Models Towards Im- proved Human Health. arXiv:2505.08775 (2025). <https://ui.adsabs.harvard.edu/ abs/2025arXiv250508775A>

    2025 arXiv

  10. [18]

    Liu, L. et al. Towards Automatic Evaluation for LLMs ' Clinical Capabilities: Metric, Data, and Algorithm. arXiv:2403.16446 (2024). <https://ui.adsabs.harvard.edu/abs/ 2024arXiv240316446L>

    2024 arXiv

  11. [19]

    Singhal, K. et al. Towards Expert-Level Medical Question Answering with Large Language Models. arXiv:2305.09617 (2023). <https://ui.adsabs.harvard.edu/abs/ 2023arXiv230509617S>

    2023 arXiv

  12. [20]

    et al.An evaluation framework for clinical use of large language models in patient interaction tasks

    Johri, S. et al.An evaluation framework for clinical use of large language models in patient interaction tasks. Nat Med 31, 77-86, doi:10.1038/s41591-024-03328-5 (2025)

    2025 doi

  13. [21]

    Tu, T. et al. Towards Conversational Diagnostic AI. arXiv:2401.05654 (2024). <https: //ui.adsabs.harvard.edu/abs/2024arXiv240105654T>

    2024 arXiv

  14. [22]

    et al.AgentClinic: a multimodal agent benchmark to evaluate AI in simulated 27 clinical environments

    Schmidgall, S. et al.AgentClinic: a multimodal agent benchmark to evaluate AI in simulated 27 clinical environments. arXiv:2405.07960 (2024). <https://ui.adsabs.harvard.edu/abs/ 2024arXiv240507960S>

    2024 arXiv

  15. [23]

    & Wang, Y

    Liao, Y., Meng, Y., Liu, H., Wang, Y. & Wang, Y. An Automatic Evaluation Framework for Multi-turn Medical Consultations Capabilities of Large Language Models. arXiv:2309.02077 (2023). <https://ui.adsabs.harvard.edu/abs/2023arXiv230902077L>

    2023 arXiv

  16. [24]

    Shi, X. et al. LLM-Mini-CEX: Automatic Evaluation of Large Language Model for Diag- nostic Conversation. arXiv:2308.07635 (2023). <https://ui.adsabs.harvard.edu/abs/ 2023arXiv230807635S>

    2023 arXiv

  17. [25]

    Fast, D. et al. Autonomous medical evaluation for guideline adherence of large language models. NPJ Digit Med7, 358, doi:10.1038/s41746-024-01356-6 (2024)

    2024 doi

  18. [26]

    Zheng, L. et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. arXiv:2306.05685 (2023). <https://ui.adsabs.harvard.edu/abs/ 2023arXiv230605685Z>

    2023 arXiv

  19. [27]

    et al.Automating Evaluation of AI Text Generation in Healthcare with a Large Language Model (LLM)-as-a-Judge

    Croxford, E. et al.Automating Evaluation of AI Text Generation in Healthcare with a Large Language Model (LLM)-as-a-Judge. medRxiv, doi:10.1101/2025.04.22.25326219 (2025)

    2025 doi

  20. [28]

    Wu, Y. et al. Effectiveness of various general large language models in clinical consensus and case analysis in dental implantology: a comparative study. BMC Med Inform Decis Mak 25, 147, doi:10.1186/s12911-025-02972-2 (2025)

    2025 doi

  21. [29]

    Verlingue, L. et al. Artificial intelligence in oncology: ensuring safe and effective in- tegration of language models in clinical practice. Lancet Reg Health Eur 46, 101064, doi:10.1016/j.lanepe.2024.101064 (2024)

    2024

  22. [30]

    Bedi, S. et al. MedHELM: Holistic Evaluation of Large Language Models for Medical Tasks. arXiv:2505.23802 (2025). <https://ui.adsabs.harvard.edu/abs/ 2025arXiv250523802B>

    2025 arXiv

  23. [31]

    et al.Aligning Large Language Models with Humans: A Comprehensive Survey of ChatGPT's Aptitude in Pharmacology

    Zhang, Y. et al.Aligning Large Language Models with Humans: A Comprehensive Survey of ChatGPT's Aptitude in Pharmacology. Drugs 85, 231-254, doi:10.1007/s40265-024-02124-2 (2025)

    2025 doi

  24. [32]

    T., Bardak, A

    Dinc, M. T., Bardak, A. E., Bahar, F. & Noronha, C. Comparative analysis of large language models in clinical diagnosis: performance evaluation across common and complex medical cases. JAMIA Open 8, ooaf055, doi:10.1093/jamiaopen/ooaf055 (2025). 28 Supplementary Figure S1. Com...

    2025 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.