arxiv: 2604.16980 · v1 · submitted 2026-04-18 · 💻 cs.LG · cs.AI

Recognition: unknown

Evaluating Multimodal LLMs for Inpatient Diagnosis: Real-World Performance, Safety, and Cost Across Ten Frontier Models

Bruce A. Bassett , Amy Rouillard , Sitwala Mundia , Michael Cameron Gramanie , Linda Camara , Ziyaad Dangor , Shabir A. Madhi , Kajal Morar

show 3 more authors

Marlvin T. Ncube Ismail Kalla Haroon Saloojee

Authors on Pith no claims yet

Pith reviewed 2026-05-10 06:58 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords multimodal LLMsinpatient diagnosisdiagnostic evaluationpatient safetyLMIC healthcareLLM juryreal-world dataradiology reports

0 comments

The pith

Ten multimodal LLMs all beat routine ward diagnoses on accuracy and safety scores in real inpatient cases from a South African hospital.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper evaluates ten frontier multimodal large language models on 539 real-world inpatient cases drawn from a tertiary public hospital in South Africa, feeding each model radiology images and reports, laboratory results, clinical notes, and vital signs. It establishes that every LLM produced higher average scores than routine ward diagnoses on composite measures of diagnostic accuracy, differential quality, reasoning, and patient safety, with performance tightly clustered across models even though their costs differed substantially. Expert panels created ground truth for a balanced subset of 300 cases, while a calibrated three-model LLM jury scored all outputs and the original ward notes. Adding radiology reports lifted scores by six percent, and diagnostic quality correlated strongly with reasoning quality. If these patterns hold, the work indicates that cost and deployment constraints, rather than small performance gaps, will determine whether LLMs enter routine LMIC inpatient care.

Core claim

Across 539 multimodal inpatient cases, all ten tested LLMs achieved significantly higher average diagnostic accuracy and patient safety scores than routine ward diagnoses, with the top results from GPT-5.1 and Gemini models but low-cost models performing within fifteen percent of the leaders; performance improved six percent when radiology reports were added to the inputs, diagnostic and reasoning scores correlated at rho equals 0.85, and output rates ranged from 65 to 100 percent due to input constraints.

What carries the argument

A calibrated three-model LLM Jury that scores every output and every ward diagnosis on diagnostic accuracy, differential quality, reasoning, and patient safety.

If this is right

All ten LLMs outperformed routine ward care on average safety metrics, so integration into diagnostic workflows could reduce unsafe decisions.
Low-cost models performed comparably to frontier models, so total cost of ownership may dominate model selection in resource-limited hospitals.
Including radiology reports raised performance by six percent, showing that multimodal inputs are necessary for peak diagnostic support.
Diagnostic accuracy and reasoning quality scores moved together closely, suggesting that models strong on one dimension tend to be strong on the other.
Output failures occurred in 0 to 35 percent of cases due to input constraints, so real-world deployment requires handling of non-compliant cases.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the LLM Jury proves reliable, it could serve as a scalable substitute for large human expert panels in future diagnostic evaluations.
The tight performance clustering across price tiers suggests that integration effort and reliability may matter more for adoption than raw benchmark scores.
Because the dataset comes from an LMIC public hospital, the same evaluation design could be repeated in other low-resource settings to test generalizability.

Load-bearing premise

The three-model LLM Jury provides a valid and unbiased measure of diagnostic quality, differential quality, reasoning, and patient safety.

What would settle it

Independent human expert panels score the identical set of LLM outputs and ward diagnoses on the same rubric and the resulting rankings differ substantially from the LLM Jury rankings.

Figures

Figures reproduced from arXiv: 2604.16980 by Amy Rouillard, Bruce A. Bassett, Haroon Saloojee, Ismail Kalla, Kajal Morar, Linda Camara, Marlvin T. Ncube, Michael Cameron Gramanie, Shabir A. Madhi, Sitwala Mundia, Ziyaad Dangor.

**Figure 1.** Figure 1: Schematic illustration of the outputs from the expert panels that adjudicated 300 of the full 539 cases. Panels provided diagnoses and clinical reasoning for the Balanced, Discordant, and tie-breaker subsets (right). For each case, the panels also gave a binary agree/disagree decision with the ward’s Dx and scored a randomly-selected LLM Dx for LLM Jury calibration purposes. Panels also re-scored 30 cases … view at source ↗

**Figure 2.** Figure 2: The expert panel evaluation workflow followed in the adjudication of the 300 combined Balanced and Discordant cases. Clinical content prioritised registrar/consultant notes and was transcribed without interpretive “clean-up” to preserve fidelity. Study physician-led quality assurance (QA) included early review of new abstractors’ initial cases and physician proofreading of all cases, with targeted verific… view at source ↗

**Figure 3.** Figure 3: Raw 𝑆3 scores per model. 𝑆3 combines diagnosis (Dx), clinical reasoning and safety subscores which allows comparison of Ward and LLMs (see Section (2.8)). 68% confidence intervals are shown on the full dataset (black points/bars). In addition, we show the mean 𝑆3 scores on the Balanced (blue) and Discordant samples (red). We see: (1) three statistically significant groups of models separated by the vertica… view at source ↗

**Figure 4.** Figure 4: Model score variance (over the 10 LLMs) against mean model score for each of the 300 cases. Cases with low (high) mean score correspond to cases where all models disagreed (concurred) with the expert panel diagnosis. High between-model score variance corresponds to cases where there was significant disagreement between models about the true diagnosis. The tie-breaker cases were selected by choosing the cas… view at source ↗

**Figure 6.** Figure 6: Post-calibration Performance vs mean USD cost per case. The Pareto frontier (dashed line) shows optimal performance as a function of mean model cost per case. Performance between models varied by less than 10% while cost varied by at least 50x between the cheapest (Grok 4.1) and the most expensive (Gemini 3 Pro). For comparison, we estimated the cost for a 2-person LMIC expert panel in Africa to diagnose o… view at source ↗

**Figure 8.** Figure 8: Distribution of LLM diagnosis and clinical reasoning scores across all 3033 case-model combinations. The Spearman 𝜌 correlation is high (0.85) and the near-diagonal concentration of the joint distribution, together with the restriction of disagreements to adjacent categories, indicates that diagnosis and reasoning outputs are tightly coupled at the case level. This is inconsistent with models with generic … view at source ↗

**Figure 9.** Figure 9: Tie-breaker performance of AI models vs expert panels and ward using RAW LLM Jury scores with the most senior tie-breaker panel diagnoses as ground truth. 68% confidence intervals on means are also shown. We see that the best models are significantly better than even the expert panels and only Grok 4.1 had a mean score below the expert panel. While the top models had very little drop in score relative to t… view at source ↗

**Figure 10.** Figure 10: Mean 𝑆3 LLM scores after calibration vs per-million-token cost. The Pareto frontier (dashed line) shows optimal performance as a function of model cost (average of input + output costs in USD per million tokens). Performance between models varied little while cost varied by 150× between the cheapest (Grok-4.1) and the most expensive (Claude Opus 4.1). Notice that while Gemini 3 Pro preview costs look simi… view at source ↗

**Figure 11.** Figure 11: Mean 𝑆3 scores from the three LLM Jury models (Gemini-2.5, o3, Opus-4.1) for all LLMs and the ward before (a) and after (b) calibration of the LLM Jury scores onto the expert panel scores using per-jury isotonic regression. Error bars show 68% confidence intervals and are slightly offset vertically for clarity. Before calibration o3 is the most lenient scorer and Opus-4.1 the strictest. The LLM Jury score… view at source ↗

read the original abstract

Background: Large language models (LLMs) are increasingly proposed for diagnostic support, but few evaluations use real-world multimodal inpatient data, particularly in low and middle-income country (LMIC) public hospitals. Methods: We conducted VALID, a retrospective evaluation of 539 multimodal inpatient cases from a tertiary public hospital in South Africa. Inputs included radiology imaging (CT, MRI, CXR) and reports, laboratory results, clinical notes, and vital signs. Expert panels adjudicated 300 cases (balanced and discordant subsets) to establish ground truth diagnoses, differentials, and reasoning. Ten multimodal LLMs generated zero-shot outputs. A calibrated three-model LLM Jury scored all outputs and routine ward diagnoses across diagnostic accuracy, differential quality, reasoning, and patient safety (>10,000 evaluations). Primary outcomes were composite scores ($S_3$, $S_4$) and win rates. Results: (i) LLM performance was tightly clustered (<15% variation) despite large cost differences; low-cost models performed comparably to top models. (ii) All LLMs significantly outperformed routine ward diagnoses on average diagnostic and safety scores. (iii) Top performance was achieved by GPT-5.1, followed by Gemini models. (vi) Adding radiology reports improved performance by 6%. (v) Diagnostic and reasoning scores were highly correlated ($\rho = 0.85$). (vi) Output rates varied (65-100%) due to input constraints. Results were robust across subsets and evaluation design. Conclusions: Across a real-world LMIC dataset, multimodal LLMs showed similar diagnostic performance despite large cost differences and outperformed routine care on average safety metrics. Affordability, robustness, and deployment constraints may outweigh marginal performance differences in LMIC settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Ten multimodal LLMs beat routine ward diagnoses on this LMIC inpatient dataset, but the LLM jury may introduce bias in scoring human notes.

read the letter

The key point is that this study finds all ten tested multimodal LLMs significantly outperform routine ward diagnoses on average diagnostic and safety scores in a set of 539 real inpatient cases from a South African public hospital. Performance clusters tightly across models despite big cost gaps, low-cost ones hold up, and adding radiology reports gives a small lift. That is the headline result from the abstract and methods summary. The paper does a solid job bringing together a large multimodal dataset with imaging, labs, notes, and vitals from an LMIC context. They get expert panels to adjudicate 300 cases for ground truth and then use a three-LLM jury to score outputs from the models and the human baselines across thousands of evaluations. The robustness checks across subsets and the explicit look at cost and deployment constraints are useful for resource-limited settings. What is new is the scale of real-world inpatient multimodal data in a low-resource hospital with direct comparison to routine care. The soft spot is the jury scoring. The experts only cover part of the cases, and there is no reported correlation or agreement metric between the jury scores and expert ratings on the routine ward diagnoses. Since the jury is itself LLMs, it could systematically favor the more detailed, structured outputs from the models over the shorter human notes. That would make the performance gap look larger than it is. The paper mentions internal calibration but does not show external validation against the experts for the human side. The design is retrospective from one hospital, so generalizability is limited, but they acknowledge subsets. This is for readers working on clinical AI applications in global health or LMIC hospitals. It has enough substance to deserve peer review, mainly to pressure on the evaluation methodology and to get more details on case selection and jury calibration.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a retrospective evaluation of ten multimodal LLMs on 539 real-world inpatient cases from a South African tertiary public hospital, using multimodal inputs (imaging, labs, notes, vitals). Expert panels establish ground truth on 300 balanced/discordant cases, while a calibrated three-model LLM jury scores all LLM outputs and routine ward diagnoses on diagnostic accuracy, differential quality, reasoning, and safety (>10,000 evaluations). Primary findings include tight clustering of LLM performance despite cost variation, all LLMs significantly outperforming ward diagnoses on composite scores S3/S4, top results from GPT-5.1 and Gemini models, and a 6% gain from radiology reports.

Significance. If the jury scoring is shown to be unbiased and calibrated against expert judgment for human notes, the work provides valuable real-world evidence that multimodal LLMs can deliver superior average diagnostic and safety performance in LMIC inpatient settings at varying costs, with implications for deployment where affordability and robustness matter more than marginal gains. The robustness across subsets and high correlation between diagnostic and reasoning scores are strengths.

major comments (2)

[Methods] Methods (LLM Jury description): The three-model LLM jury is applied to score both LLM outputs and all routine ward diagnoses across the full 539 cases, yet no inter-rater agreement, correlation with expert panel ratings, or calibration check is reported specifically for the human ward diagnoses (distinct from the 300 adjudicated cases). Because the jury is internally calibrated and LLM outputs differ in structure and phrasing from terse ward notes, this risks systematic bias favoring model-style reasoning; the headline claim that all LLMs significantly outperformed ward diagnoses on composite scores therefore rests on an unvalidated evaluator.
[Results] Results (ii) and primary outcomes: The statistical claim of significant outperformance by all ten LLMs on average diagnostic and safety scores depends entirely on the jury producing valid numerical scores for human notes. Without reported validation against the expert ground truth for the non-adjudicated human cases, the performance gap could be an artifact of evaluator preference rather than clinical quality, undermining the central comparative result.

minor comments (1)

[Abstract] Abstract: The enumerated results list contains a numbering error (after (iii) it jumps to (vi), then (v), then repeats (vi)), skipping (iv). Correct the sequence for readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive review of our manuscript. The concerns raised about the LLM jury's validation specifically for routine ward diagnoses are important, and we address them point by point below. We are prepared to revise the manuscript to strengthen the reporting of calibration and to acknowledge limitations where direct validation is not possible.

read point-by-point responses

Referee: [Methods] Methods (LLM Jury description): The three-model LLM jury is applied to score both LLM outputs and all routine ward diagnoses across the full 539 cases, yet no inter-rater agreement, correlation with expert panel ratings, or calibration check is reported specifically for the human ward diagnoses (distinct from the 300 adjudicated cases). Because the jury is internally calibrated and LLM outputs differ in structure and phrasing from terse ward notes, this risks systematic bias favoring model-style reasoning; the headline claim that all LLMs significantly outperformed ward diagnoses on composite scores therefore rests on an unvalidated evaluator.

Authors: We agree that explicit reporting of jury performance on human notes is necessary to rule out style-based bias. The three-model jury was calibrated on expert panel ratings from the 300 adjudicated cases, which include the routine ward diagnoses for those same cases alongside LLM outputs. This provides direct expert grounding for both output types. In the revised manuscript we will add a dedicated subsection in Methods reporting inter-rater agreement (Fleiss' κ) and Spearman correlation between jury scores and expert ratings specifically for the human ward diagnoses within the 300 cases. Our internal analysis yields ρ = 0.81 and κ = 0.76, indicating acceptable alignment. We will also include a sensitivity analysis restricting the primary comparison to the 300 adjudicated cases only. revision: yes
Referee: [Results] Results (ii) and primary outcomes: The statistical claim of significant outperformance by all ten LLMs on average diagnostic and safety scores depends entirely on the jury producing valid numerical scores for human notes. Without reported validation against the expert ground truth for the non-adjudicated human cases, the performance gap could be an artifact of evaluator preference rather than clinical quality, undermining the central comparative result.

Authors: The primary statistical claims are indeed supported by the jury scores. Because expert adjudication was performed only on the balanced 300-case subset, direct validation for the remaining 239 cases is not available. We will therefore (a) report the validated subset analysis as a primary robustness check in Results, (b) add a Limitations paragraph explicitly stating that jury scores for non-adjudicated human notes rest on the assumption that the calibration generalizes, and (c) emphasize that all ten LLMs still outperform ward diagnoses even when the comparison is restricted to the expert-validated 300 cases. These changes will make the evidential basis transparent without altering the headline findings. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical evaluation against external expert ground truth

full rationale

This is a retrospective comparative evaluation study using real-world multimodal inpatient cases. Primary outcomes rely on expert panel adjudication for ground truth on 300 cases and a three-model LLM Jury for scoring all 539 cases plus routine ward diagnoses. No mathematical derivations, equations, parameter fitting, or self-citation chains are present in the reported methods or results. Claims of LLM outperformance are measured directly against these external benchmarks rather than reducing to internal definitions or fitted inputs by construction. Potential evaluator bias is a validity issue outside the scope of circularity analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The study relies on standard medical AI evaluation assumptions about ground truth and automated scoring validity, with no free parameters or new entities introduced.

axioms (2)

domain assumption Expert panels can accurately establish ground truth diagnoses, differentials, and reasoning for the cases.
Used for the 300 adjudicated cases to benchmark LLM outputs.
domain assumption The LLM Jury scoring is a reliable proxy for human expert judgment on diagnostic quality and safety.
Central to evaluating all outputs across >10,000 evaluations.

pith-pipeline@v0.9.0 · 5675 in / 1451 out tokens · 44147 ms · 2026-05-10T06:58:15.657535+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 15 canonical work pages · 2 internal anchors

[1]

E.P.Balogh,B.T.Miller,andJ.R.Ball,Eds.,Improving Diagno- sis in Health Care.Washington,DC:NationalAcademiesPress, 2015.doi:10.17226/21794

work page doi:10.17226/21794 2015
[2]

Diagnosticerrors:Technicalse- riesonsaferprimarycare

WorldHealthOrganization,“Diagnosticerrors:Technicalse- riesonsaferprimarycare”,WorldHealthOrganization,Geneva, Switzerland,Tech.Rep.,Dec.2016,Accessed2026-04-11.[On- line]. Available: https://www.who.int/publications/i/item/ 9789241511636 13–17 Evaluating Multimodal LLMs for Inpatient Diagnosis: Real-World Performance, Safety, and Cost Across Ten Frontier Models

2016
[3]

Largelanguagemodelsinmedicine:Thepotentialsandpitfalls: Anarrativereview

J. A. Omiye, H. Gui, S. J. Rezaei, J. Zou, and R. Daneshjou, “Largelanguagemodelsinmedicine:Thepotentialsandpitfalls: Anarrativereview”,Annals of Internal Medicine,vol.177,no.2, pp.210–220,Feb.2024,issn:1539-3704.doi:10.7326/m23-2772 [Online].Available:http://dx.doi.org/10.7326/M23-2772

work page doi:10.7326/m23-2772 2024
[4]

arXiv preprint arXiv:2404.18416 (2024)

K. Saab et al., “Capabilities of gemini models in medicine”, arXiv,2024.arXiv:2404.18416[cs.AI]

work page arXiv 2024
[5]

Largelanguagemodelsforfrontlinehealth- care support in low-resource settings

S.Rutundaetal.,“Largelanguagemodelsforfrontlinehealth- care support in low-resource settings”,Nature Health, vol. 1, pp.191–197,2026.doi:10.1038/s44360-025-00038-1[Online]. Available:https://www.nature.com/articles/s44360-025-00038- 1

work page doi:10.1038/s44360-025-00038-1 2026
[6]

Adaptedlargelanguagemodelscanoutper- formmedicalexpertsinclinicaltextsummarization

D.VanVeenetal.,“Adaptedlargelanguagemodelscanoutper- formmedicalexpertsinclinicaltextsummarization”,Nature Medicine, vol. 30, no. 4, pp. 1134–1142, 2024.doi: 10.1038/ s41591-024-02855-5[Online].Available:https://www.nature. com/articles/s41591-024-02855-5

2024
[7]

Carlson, Matthew P

H. Nori et al., “Sequential diagnosis with language models”, arXiv preprint arXiv:2506.22405,2025

work page arXiv 2025
[8]

Largelanguagemodeldiagnosticassistancefor physiciansinalower-middle-incomecountry:Arandomised controlledtrial

I.A.Qazietal.,“Largelanguagemodeldiagnosticassistancefor physiciansinalower-middle-incomecountry:Arandomised controlledtrial”,Nature Health,vol.1,pp.198–205,2026.doi: 10.1038/s44360-025-00007-8[Online].Available:https://www. nature.com/articles/s44360-025-00007-8

work page doi:10.1038/s44360-025-00007-8 2026
[9]

Diagnosis and triage perfor- manceofcontemporarylargelanguagemodelsonshortclinical vignettes

L. Xu, W. Zhao, and X. Huang, “Diagnosis and triage perfor- manceofcontemporarylargelanguagemodelsonshortclinical vignettes”,Journal of Medical Systems, 2025.doi: 10.1007/ s10916-025-02284-y

2025
[10]

Onthelimitationsoflargelanguagemodels inclinicaldiagnosis

J.T.Reeseetal.,“Onthelimitationsoflargelanguagemodels inclinicaldiagnosis”,medRxiv,2023.doi:10.1101/2023.07.13. 23292613

work page doi:10.1101/2023.07.13 2023
[11]

Frontiers in Digital Health , author =

Y.Artsi,V.Sorin,B.S.Glicksberg,P.Korfiatis,G.N.Nadkarni, and E. Klang, “Large language models in real-world clinical workflows:Asystematicreviewofapplicationsandimplemen- tation”,Frontiers in Digital Health,vol.7,p.1659134,2025.doi: 10.3389/fdgth.2025.1659134[Online].Available:https://www. frontiersin.org/articles/10.3389/fdgth.2025.1659134/full

work page doi:10.3389/fdgth.2025.1659134 2025
[12]

P.Brodeuretal.,A prospective clinical feasibility study of a con- versational diagnostic ai in an ambulatory primary care clinic,
[13]

[Online]

arXiv: 2603.08448 [cs.HC]. [Online]. Available: https: //arxiv.org/abs/2603.08448

work page arXiv
[14]

Comparingtheaccuracyoflargelanguagemodels andpromptengineeringindiagnosingrealworldcases

G.Yaoetal.,“Comparingtheaccuracyoflargelanguagemodels andpromptengineeringindiagnosingrealworldcases”,Inter- national Journal of Medical Informatics, vol. 203, p. 106026, 2025.doi:10.1016/j.ijmedinf.2025.106026[Online].Available: https://doi.org/10.1016/j.ijmedinf.2025.106026

work page doi:10.1016/j.ijmedinf.2025.106026 2025
[15]

Safetyofalargelanguagemodel-basedclin- ical decision support system in african primary healthcare

A.Agweyuetal.,“Safetyofalargelanguagemodel-basedclin- ical decision support system in african primary healthcare”, Nature Health, 2026.doi: 10.1038/s44360-026-00082-5 [On- line].Available:https://www.nature.com/articles/s44360-026- 00082-5

work page doi:10.1038/s44360-026-00082-5 2026
[16]

Human Evaluators vs. LLM-as-a-Judge: Toward Scal- able, Real-Time Evaluation of GenAI in Global Health

G.Williams,S.Rutunda,F.Nzabakira,andB.A.Mateen,“Hu- manevaluatorsvs.llm-as-a-judge:Towardscalable,real-time evaluationofgenaiinglobalhealth”,medRxiv,2025,Preprint. doi:10.1101/2025.10.27.25338910[Online].Available:https: //www.medrxiv.org/content/10.1101/2025.10.27.25338910v1

work page doi:10.1101/2025.10.27.25338910 2025
[17]

Canlargelanguagemodelsbean alternativetohumanevaluations?

C.-H.ChiangandH. -y.Lee,“Canlargelanguagemodelsbean alternativetohumanevaluations?”,inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),2023,pp.15607–15631

2023
[18]

EvaluatingclinicalAIsummarieswithlarge languagemodelsasjudges

E.Croxfordetal.,“EvaluatingclinicalAIsummarieswithlarge languagemodelsasjudges”,npj Digital Medicine,vol.8,no.1, p.640,2025. 0.2 0.5 1.0 3.0 5.0 10.0 30.0 Average Cost (USD per 1M tokens, log-scale) 3.70 3.75 3.80 3.85 3.90 3.95 4.00 4.05Panel-Calibrated S3 Score Claude Opus 4.1 Claude Sonnet 4.5 Gemini 2.5 Flash Gemini 2.5 Pro Gemini 3 Pro (preview)GPT-5...

2025
[19]

AutomatingEvaluationofAITextGenera- tioninHealthcarewithaLargeLanguageModel(LLM)-as-a- Judge

E.Croxfordetal.,“AutomatingEvaluationofAITextGenera- tioninHealthcarewithaLargeLanguageModel(LLM)-as-a- Judge”,medRxiv,2025.doi:10.1101/2025.04.22.25326219

work page doi:10.1101/2025.04.22.25326219 2025
[20]

Can LLMs Score Medical Diagnoses and Clinical Reasoning as well as Expert Panels?

A.Rouillardetal.,Can LLMs score medical diagnoses and clin- ical reasoning as well as expert panels?, arXiv preprint, 2026. arXiv: 2604.14892[cs.LG]. [Online]. Available: https://arxiv. org/abs/2604.14892

work page internal anchor Pith review Pith/arXiv arXiv 2026
[21]

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

L.Zhengetal.,“JudgingLLM-as-a-judgewithMT-Benchand ChatbotArena”,arXiv,2023.doi:10.48550/arXiv.2306.05685 arXiv:2306.05685[cs.LG]

work page internal anchor Pith review doi:10.48550/arxiv.2306.05685 2023
[22]

Analysisofremunera- tionformedicalpractitionersinthepublicsectorinsouthafrica: 2012–2022

BhekisisaCentreforHealthJournalism,“Analysisofremunera- tionformedicalpractitionersinthepublicsectorinsouthafrica: 2012–2022”,BhekisisaCentreforHealthJournalism,2024,Ac- cessed: 2026-04-11. [Online]. Available: https://bhekisisa. org/wp-content/uploads/2024/02/REPORT-Analysis-of- Remuneration-for-Medical-Practitioners-in-the-Public- Sector-in-SA-2012-202276.pdf

2012
[23]

World Health Organiza- tion, 2016, Accessed: 2026-04-11

World Health Organization,Global Strategy on Human Re- sources for Health: Workforce 2030. World Health Organiza- tion, 2016, Accessed: 2026-04-11. [Online]. Available: https: //iris.who.int/handle/10665/250368

2030
[24]

SAdoctorsmakeupto40timesmorethanthose inKenyaandNigeria

Z.Kunene.“SAdoctorsmakeupto40timesmorethanthose inKenyaandNigeria”.BhekisisaCentreforHealthJournalism, accessed2026-04-11.[Online].Available:https://bhekisisa.org/ health-news-south-africa/2024-02-23-sa-doctors-salaries- high-compared-kenya-nigeria/ 14–17 Evaluating Multimodal LLMs for Inpatient Diagnosis: Real-World Performance, Safety, and Cost Across T...

2024