Toward Vibe Medicine: A Self-Evolving Multi-Agent Framework for Clinical Decision Support

Chenbin Liu; Hanqi Jiang; Hongying Feng; Jingyuan Chen; Jinyang Huang; Junhao Chen; Lian Zhang; Liao Zhang; Lifeng Chen; Qianxue Zhang

Toward Vibe Medicine: A Self-Evolving Multi-Agent Framework for Clinical Decision Support

T0 review · 4 major / 0 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read A three-agent medical AI that learns from its own cases across memory, model, and code can beat static systems on complex clinical decisions.

desk verdict Coherent multi-agent clinical stack with a clear evolution story, but the superiority claims rest on synthetic cases, LLM judges, and a 1.5B proxy with mixed results. read the letter →

arxiv 2606.15504 v2 pith:BHU4YM4U submitted 2026-04-01 cs.AI

Qianxue Zhang , Yiming Ren , Shihuan Qin , Xiao Zhang , Liao Zhang , Jinyang Huang , Zhengliang Liu , Chenbin Liu

show 13 more authors

Hongying Feng Jingyuan Chen Yuzhen Ding Weihang You Hanqi Jiang Yi Pan Yifan Zhou Junhao Chen Lifeng Chen Wei Liu Tianming Liu Zengren Zhao Lian Zhang

This is my paper · ORCID

classification cs.AI

keywords VibeMedicineLargeLanguageModelMulti-AgentSystemsSelf-EvolvingAgentsAISafetyClinicalDecisionSupportPrecision

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

Most medical AI systems are frozen after training: they cannot turn real patient outcomes, clinician rejections, and past failures into lasting improvement. This paper introduces VIBEMed, a three-agent pipeline that mimics clinical workflow—one agent diagnoses, one plans treatment under safety constraints, and a third turns completed encounters into reusable experience. That experience drives continuous updates at three levels: hierarchical clinical memory for immediate adaptation, staged model fine-tuning for lasting capability gains, and automated code changes for new functions—all behind an architecture-level safety sandbox that isolates execution environments and patient sessions. On thirty synthetic complex cases the full cascade scores higher and more consistently than a single model or either agent alone, and after evolution a small backbone improves on several MedBench tasks that demand integrated treatment planning and long-term personalization. The authors present this as a concrete route from static medical AI toward systems that accumulate clinical experience the way clinicians do.

What carries the argument

The three-level self-evolution mechanism (memory-level hierarchical storage and retrieval, model-level staged LoRA-SFT then DPO from CEMA reflection datasets, code-level sandbox-validated generation) together with the CDA–TEA–CEMA role pipeline and session-isolated safety sandbox.

What would settle it

A prospective multi-center trial on real patient encounters that measures whether the full evolved VIBEMed pipeline reduces clinically unsafe recommendations and improves clinician-adopted plan quality relative to a frozen single-model baseline of the same backbone, with independent human expert review of safety and accuracy.

Watch

Extended reading notes

Core claim

VIBEMed shows that a multi-agent clinical decision system with an explicit self-evolution loop—Clinical Diagnostic Agent, Therapeutic Execution Agent, and Clinical Evolution Manager Agent updating memory, model parameters via SFT/DPO, and code—outperforms static single-model pipelines on complex cases that require integrated reasoning and longitudinal planning, while an architecture-level safety sandbox keeps updates and patient data isolated.

Load-bearing premise

That synthetic cases with planted errors scored by another LLM, plus evolution experiments on a 1.5B proxy model judged the same way, are faithful enough proxies for real multi-institution clinical reliability and for gains that will transfer to larger production models.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful part of this paper is the packaging: CDA for structured diagnosis with uncertainty, TEA with guideline memory plus a three-tier safety filter, and CEMA that turns clinician adoption/rejection and outcomes into reflection data for memory, LoRA-SFT, and DPO. The architecture-level sandbox (dev/val/prod isolation, session-scoped memory, audit logs) is more carefully specified than most medical multi-agent write-ups. That design is real engineering work and worth reading if you care about hospital-deployable agent loops.

What is new is mostly the integration and the “vibe medicine” framing, not the individual pieces. Multi-agent medical pipelines, RAG over guidelines, hierarchical memory, LoRA/DPO, and sandboxed code gen are all established. The paper’s contribution is a concrete end-to-end clinical workflow that tries to close the loop from encounter to persistent learning without constant human prompt engineering.

The soft spots are exactly where the stress-test points, and the authors mostly admit them. Table 1 (30 Claude-generated cases with planted upstream errors, scored by another LLM) is a clean engineering signal that the cascade can catch error propagation better than a single model, but it is not external validity. Table 2 on a 1.5B proxy shows real gains on MedTreat and MedPHM and real regressions on MedAnalysis and MedDiag; that undercuts any blanket “evolution improves clinical reasoning” claim. No released code/data, no real multi-site records, no clinician-blinded outcomes, and incomplete ablation of the three evolution layers and the safety mechanisms. Free parameters (adoption thresholds, dataset sizes, refresh cadences) are stated but not stress-tested.

Math and citations are fine for a systems paper; nothing load-bearing is formally wrong. This is for people building clinical agent stacks or institutional LLM adaptation, not for someone looking for a settled clinical efficacy result. I would send it to peer review: the design is coherent enough and the gaps are fixable with better evaluation. Engage if you work on medical multi-agents; treat the empirical superiority claims as provisional until they run on real data and larger models.

Referee Report

4 major / 0 minor

Summary. The paper proposes VIBEMed, a multi-agent clinical decision-support framework built around three specialized agents (CDA for diagnosis, TEA for treatment planning with guideline memory and three-tier safety constraints, and CEMA for distilling encounter outcomes into reflection data) plus a three-level self-evolution stack (memory, model via LoRA-SFT/DPO, and code) and an architecture-level safety sandbox (execution isolation and session-level memory isolation). The central claim is that this combination yields superior, safer, experience-driven clinical decisions relative to static single-model pipelines, especially for integrated and longitudinal tasks. Empirical support consists of (i) a 30-case synthetic multi-agent ablation (Table 1) in which the CDA+TEA cascade scores 39.20/40 versus 35.13 for a direct LLM, and (ii) evolution experiments on a 1.5B proxy model evaluated on MedBench tasks (Table 2), with gains on MedTreat, MedPHM, and MedDiffer but declines on MedAnalysis and MedDiag, plus a qualitative oncology interface demonstration.

Significance. If the architecture and evolution claims hold under stronger evaluation, the work would be a useful systems contribution: it articulates a concrete multi-agent clinical workflow, a hierarchical memory design, staged SFT/DPO evolution with an LLM-as-judge promotion gate, and system-level isolation mechanisms that go beyond prompt-only medical agents. The multi-agent ablation is directionally informative about error propagation, and the explicit limitations section is appropriately candid. The manuscript does not, however, yet deliver machine-checked proofs, released code/data, or prospective multi-institution clinical outcomes; significance therefore remains conditional on validation that the synthetic and 1.5B-proxy results transfer to production-scale models and real clinical settings.

major comments (4)

§3.1 and Table 1: The multi-agent superiority claim rests on 30 Claude-generated synthetic cases with planted upstream errors, scored 1–5 by another LLM (Qwen3.6 Plus) on eight rubrics. This design cannot establish real-world clinical reliability or safety; it mainly shows that a cascade can recover planted errors under LLM-judged rubrics. At minimum the paper should (a) report inter-judge agreement with clinicians on a subset, (b) include a non-LLM baseline or human-scored sample, and (c) temper abstract/conclusion language that implies real-world superiority from this table alone.
§3.2 and Table 2: Evolution is demonstrated only on a 1.5B DeepSeek-Distill-Qwen2.5 proxy with LoRA-SFT (~4k treatment + 10k general samples); gains on MedTreat (+3.9) and MedPHM (+12.4) are offset by regressions on MedAnalysis (−3.0) and MedDiag (−2.0). The claim that scaling laws make these gains predictive for larger production models is asserted but not tested. Either run the same protocol on at least one larger backbone, or reframe the result strictly as a small-model feasibility study and remove transfer language from the abstract and §3.2.
§2.3–2.4 and §4: The three-level evolution stack and architecture-level safety sandbox are load-bearing for the “safe continuous evolution” claim, yet neither is ablated or stress-tested (no rollback incidents, no isolation breach tests, no contribution of memory vs. model vs. code). The limitations section acknowledges this; the manuscript should either add hierarchical ablations / safety probes or substantially narrow claims that the sandbox “guarantees” reliable clinical operation.
Evaluation circularity (§3.1–3.2): Cases are LLM-generated, pipelines are LLM-based, promotion uses LLM-as-a-judge, and MedBench scoring uses three further LLMs. Reflection data for DPO is drawn from the same system’s interactions. Without an external clinician-labeled holdout or prospective outcome measure, the superiority claim for “experience-driven clinical decision support” remains weakly grounded. A concrete fix is a small clinician-blinded preference study on a fixed case set before and after evolution.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical multi-agent systems paper with no derivation chain that reduces predictions to inputs by construction; evaluation self-reference is a validity concern, not circularity under the defined patterns.

full rationale

VIBEMed is an architecture-and-evaluation paper, not a first-principles derivation. It proposes three agents (CDA/TEA/CEMA), a three-level evolution loop (memory/model/code), and a safety sandbox, then reports empirical scores on 30 synthetic cases (Table 1) and MedBench subsets after LoRA-SFT of a 1.5B proxy (Table 2). No equation, uniqueness theorem, or fitted parameter is later re-presented as an independent prediction; the performance numbers are measured outputs of the described pipelines, not forced by definition. Self-citations (e.g., vibe-coding/research inspirations) are motivational, not load-bearing uniqueness claims. Synthetic LLM-generated cases scored by another LLM and reflection data drawn from the system’s own adoption/outcome rules create self-referential evaluation risk, but that is a correctness/external-validity issue, not a reduction of a claimed derivation to its inputs. Per the rules, honest non-finding of circularity is appropriate; score 1 only for the mild self-referential flavor of the CEMA reflection dataset construction, which does not force the reported MedBench deltas.

Assumptions & free parameters 8 free parameters · 6 assumptions · 5 invented entities

The central empirical claims rest on engineering design choices and evaluation conventions rather than physical laws. Load-bearing free parameters include adoption/rejection thresholds, update cadences, LoRA/training sizes, and judge temperatures. Domain axioms include that report-level imaging text plus structured EHR fields suffice for high-stakes diagnosis, that clinician adoption and short-term outcomes are valid evolution labels, and that small-model gains predict large-model gains. Invented entities are the named agents, memory tiers, and the “vibe medicine” framing—software constructs without independent clinical evidence outside this paper.

free parameters (8)

Success adoption threshold = >80%
CEMA labels success cases when clinician adoption >80%; this cutoff defines the positive reflection set used for evolution.
Failure rejection threshold = >50%
Failure cases use clinician rejection >50% (or adverse events); defines negative/contrastive DPO pairs.
Reflection dataset size and balance = ~40k, ~50/50, weekly FIFO
CEMA maintains ~40k samples, roughly equal success/failure, FIFO weekly updates; shapes model-level evolution data.
Weekly SFT case count = 1k–3k (exp ~4k+10k)
Weekly LoRA-SFT uses 1k–3k success cases (experiment used ~4k medical + 10k general).
Monthly DPO pair count = 5k–10k
Monthly DPO uses 5k–10k paired success/failure cases from clinician decisions.
LoRA rank scaling and dropout = α=32, dropout=0.05, lr=5e-4
Training uses α=32, dropout 0.05, lr 5e-4, 3 epochs—hand-chosen adaptation hyperparameters.
Guideline refresh cadences = 6 months / 1 month
Authoritative guidelines every six months; literature/knowledge bases monthly—policy choices affecting TEA knowledge freshness.
Generation and judge temperatures = 0.30 / 0.20
DeepSeek-V3.2 generation T=0.30; Qwen3.6 Plus scoring T=0.20—affect reported pipeline scores.

assumptions (6)

domain assumption Imaging reports (not raw images) plus structured clinical text are sufficient inputs for reliable diagnostic ranking in complex cases.
Stated in CDA design (§2.2) citing limits of multimodal LLMs on raw images; load-bearing for end-to-end claims.
domain assumption Clinician adoption rates and short-term outcomes are valid labels for success/failure and for DPO preference pairs.
CEMA classification and reflection datasets (§2.2–2.3) treat adoption/outcomes as ground truth for evolution.
ad hoc to paper Performance gains from evolution on a 1.5B model predict similar gains on larger production models via scaling laws.
Explicitly used to justify the MedBench proxy experiment (§3.2, citing Kaplan et al.).
domain assumption LLM-as-a-judge scores on fixed rubrics are adequate proxies for clinical quality when screening models.
Used for Table 1 and Table 2 evaluation and for model promotion gates (§2.3, §3).
domain assumption Isolated execution environments, session-scoped memory ACLs, and immutable logs enforce clinical safety without needing fully trusted agent behavior.
Architecture-level safety sandbox claim (§2.4); not empirically stress-tested in results.
standard math Standard transformer attention, RAG, LoRA, SFT, and DPO behave as in the cited literature.
Background ML machinery assumed throughout methods.

invented entities (5)

VIBEMed / “vibe medicine” paradigm
purpose: Brand the experience-driven, self-evolving clinical AI framing analogous to vibe coding/research.
Introduced in introduction as a new paradigm; no independent clinical validation outside this system description.
Clinical Diagnostic Agent (CDA)
purpose: Role-specialized diagnosis and structured differential generation from multimodal text.
Software role defined by system prompts on a shared backbone; not a new biological or physical entity.
Therapeutic Execution Agent (TEA)
purpose: Guideline-grounded treatment planning with three-tier safety constraints and hybrid retrieval.
Defined in §2.2; evidence limited to synthetic pipeline scores and a UI demo.
Clinical Evolution Manager Agent (CEMA)
purpose: Distill encounters into success/failure/edge/routine labels and drive memory/model evolution.
Core learning component; independent clinical utility not shown beyond MedBench deltas.
Three-level self-evolution stack (memory/model/code) with architecture-level safety sandbox
purpose: Enable continuous improvement while isolating unverified updates and patient sessions.
Architectural invention of the paper; safety and per-level necessity not fully evaluated (limitations).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Toward Vibe Medicine: A Self-Evolving Multi-Agent Framework for Clinical Decision Support." pith.science (2026). https://pith.science/paper/BHU4YM4U

@misc{pith2026260615504,
  author       = {Pith},
  title        = {Pith review of: Toward Vibe Medicine: A Self-Evolving Multi-Agent Framework for Clinical Decision Support},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHU4YM4U}},
  note         = {Machine review of arXiv:2606.15504}
}

read the original abstract

In recent years, the advances of large language models and autonomous agents have revolutionized the healthcare field, facilitating diagnosis and improving treatment results. However, most existing AI systems rely on pre-trained knowledge and predefined pipelines, which struggle to learn dynamically from the interactive chat session history that contains patient outcomes and past failures. To address this limitation, we propose VIBEMed, a multi-agent framework with a built-in self-evolution mechanism and architecture-level safety sandbox for robust clinical decision support. The system integrates three specialized agents, including a Clinical Diagnostic Agent (CDA) for hypothesis generation, a Therapeutic Execution Agent (TEA) for treatment planning, and a Clinical Evolution Manager Agent (CEMA) that distills longitudinal clinical feedback into reusable knowledge, transforming multimodal patient information into personalized medical decisions. Through self-evolution mechanism, the framework enables iterative updates across memory, model behavior, and decision strategies, allowing the system to improve over time. Experimental results show that VIBEMed demonstrates superior performance through its evolving mechanism in complex clinical cases, particularly in tasks that require integrated decision-making and longitudinal planning. The framework also supports reliable end-to-end decisions in challenging scenarios such as oncology treatment planning, highlighting its feasibility in real-world clinical contexts. Overall, VIBEMed provides a practical path beyond static AI systems toward adaptive, experience-driven clinical decision support, demonstrating the value of combining multi-agent collaboration with continuous evolution for advancing precision medicine.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 linked inside Pith

[1]

X. Hou, Y. Zhao, Y. Liu, Z. Yang, K. Wang, L. Li, X. Luo, D. Lo, J.Grundy,H.Wang,Largelanguagemodelsforsoftwareengineering: A systematic literature review, ACM Transactions on Software Engineering and Methodology 33 (2024) 1–79

2024
[2]

A.J.Thirunavukarasu,D.S.J.Ting,K.Elangovan,L.Gutierrez,T.F. Tan, D. S. W. Ting, Large language models in medicine, Nature medicine 29 (2023) 1930–1940

2023
[3]

R. Li, X. Wang, D. Berlowitz, J. Mez, H. Lin, H. Yu, Care-ad: a multi-agent large language model framework for alzheimer’s disease prediction using longitudinal clinical notes, npj Digital Medicine 8 (2025) 541

2025
[4]

Q. Peng, Y. Cai, J. Liu, Q. Zou, X. Chen, Z. Zhong, Z. Wang, J. Xie, Q.Li, Integrationofmulti-sourcemedicaldataformedicaldiagnosis question answering, IEEE Transactions on Medical Imaging 44 (2024) 1373–1385

2024
[5]

Y.Zheng,W.Gan,Z.Chen,Z.Qi,Q.Liang,P.S.Yu, Largelanguage models for medicine: a survey, International Journal of Machine Learning and Cybernetics 16 (2025) 1015–1040

2025
[6]

C.J.Kelly,A.Karthikesalingam,M.Suleyman,G.Corrado,D.King, Key challenges for delivering clinical impact with artificial intelli- gence, BMC medicine 17 (2019) 195

2019
[7]

H. Shi, Z. Xu, H. Wang, W. Qin, W. Wang, Y. Wang, Z. Wang, S. Ebrahimi, H. Wang, Continual learning of large language models: A comprehensive survey, ACM Computing Surveys 58 (2025) 1–42

2025
[8]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,Ł.Kaiser,I.Polosukhin, Attentionisallyouneed, Advances in neural information processing systems 30 (2017)

2017

Show all 29 references

[9]

Lewis, E

P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, et al., Retrieval- augmented generation for knowledge-intensive nlp tasks, Advances in neural information processing systems 33 (2020) 9459–9474

2020
[10]

Singhal, S

K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, N.Scales,A.Tanwani,H.Cole-Lewis,S.Pfohl,etal., Largelanguage models encode clinical knowledge, Nature 620 (2023) 172–180

2023
[11]

B. Peng, Y. Zhu, Y. Liu, X. Bo, H. Shi, C. Hong, Y. Zhang, S. Tang, Graph retrieval-augmented generation: A survey, ACM Transactions on Information Systems 44 (2025) 1–52

2025
[12]

Y.Shen,L.Heacock,J.Elias,K.D.Hentel,B.Reig,G.Shih,L.Moy, Chatgpt and other large language models are double-edged swords, Radiology 307 (2023) e230163

2023
[13]

J.Haltaufderheide,R.Ranisch, Theethicsofchatgptinmedicineand healthcare:asystematicreviewonlargelanguagemodels(llms), NPJ digital medicine 7 (2024) 183

2024
[14]

Zhong, S

X. Zhong, S. Li, Z. Chen, L. Ge, D. Yu, S. Wang, L. You, H. Shang, Considerationsforpatientprivacyoflargelanguagemodelsinhealth care:scopingreview, JournalofMedicalInternetResearch27(2025) e76571

2025
[15]

C.Chen,X.Feng,Y.Li,L.Lyu,J.Zhou,X.Zheng,J.Yin, Integration of large language models and federated learning, Patterns 5 (2024)

2024
[16]

Meske, T

C. Meske, T. Hermanns, E. Von der Weiden, K.-U. Loser, T. Berger, Vibe coding as a reconfiguration of intent mediation in software development: Definition, implications, and research agenda, IEEE Access 13 (2025) 213242–213259

2025
[17]

Y. Lyu, X. Zhang, X. Yi, Y. Zhao, S. Guo, W. Hu, J. Piotrowski, J. Kaliski, J. Urbani, Z. Meng, et al., Evoscientist: Towards multi- agent evolving ai scientists for end-to-end scientific discovery, arXiv preprint arXiv:2603.08127 (2026)

2026
[18]

J.Fang,Y.Peng,X.Zhang,Y.Wang,X.Yi,G.Zhang,Y.Xu,B.Wu, S.Liu,Z.Li,etal.,Acomprehensivesurveyofself-evolvingaiagents: A new paradigm bridging foundation models and lifelong agentic systems, arXiv preprint arXiv:2508.07407 (2025)

2025 arXiv
[19]

Tavakoli, Z

N. Tavakoli, Z. Shakeri, V. Gowda, K. Samsel, A. Bedayat, A. Ghasemiesfe, U. Bagci, A. Hsiao, T. Leiner, J. Carr, et al., Gener- ativeaiandfoundationmodelsinradiology:Applications,opportuni- ties, and potential challenges, Radiology 317 (2025) e242961

2025
[20]

Z. Wang, Q. Grundy, L. Parker, L. Bero, Variations in processes for guideline adaptation: a qualitative study of world health organization staffexperiencesinimplementingguidelines, BMCPublicHealth20 (2020) 1758

2020
[21]

Newton, A

N. Newton, A. Bamgboje-Ayodele, R. Forsyth, A. Tariq, M. T. Baysari, A systematic review of clinicians’ acceptance and use of clinical decision support systems over time, npj Digital Medicine 8 (2025) 309

2025
[22]

X.Liu,H.Liu,G.Yang,Z.Jiang,S.Cui,Z.Zhang,H.Wang,L.Tao, Y. Sun, Z. Song, et al., A generalist medical language model for disease diagnosis assistance, Nature medicine 31 (2025) 932–942

2025
[23]

Liu, et al., A survey on llm-as-a-judge, The Innovation (2024)

J.Gu,X.Jiang,Z.Shi,H.Tan,X.Zhai,C.Xu,W.Li,Y.Shen,S.Ma, H. Liu, et al., A survey on llm-as-a-judge, The Innovation (2024). :Preprint submitted to Elsevier Page 10 of 11

2024
[24]

Jarrassier, F

A. Jarrassier, F. B. de Saint-Léger, G. de Rocquigny, P. Ariès, A. Joseph, Y. Gojon, C. Jacques-Sebastien, J.-C. Riff, F. Klack, A.Noel,etal., Canaigeneratesafeanaesthesiaplans?acomparative evaluation of three large language models on 100 synthetic cases, Anaesthesia Critical ...

2026
[25]

Kaplan, S

J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, D. Amodei, Scaling laws for neural language models, arXiv preprint arXiv:2001.08361 (2020)

2001 arXiv
[26]

M.Liu,W.Hu,J.Ding,J.Xu,X.Li,L.Zhu,Z.Bai,X.Shi,B.Wang, H. Song, et al., Medbench: A comprehensive, standardized, and reliable benchmarking system for evaluating chinese medical large language models, Big Data Mining and Analytics 7 (2024) 1116– 1128

2024
[27]

J.Ding,L.Lu,C.Ding,M.Bian,J.Chen,W.Pang,R.Chen,X.Peng, R.Lu,S.Ren,etal., Medbenchv4:Arobustandscalablebenchmark forevaluatingchinesemedicallanguagemodels,multimodalmodels, and intelligent agents, arXiv preprint arXiv:2511.14439 (2025)

2025
[28]

S. Zhou, W. Xie, J. Li, Z. Zhan, M. Song, H. Yang, C. Espinoza, L. Welton, X. Mai, Y. Jin, et al., Automating expert-level medical reasoning evaluation of large language models, npj Digital Medicine (2025)

2025
[29]

S. F. Chen, A. Alyakin, A. Seas, E. Yang, J. J. Choi, J. V. Lee, A. L. Chen, P. I. Warman, R. T. Bitolas, R. J. Steele, et al., Llm-assisted systematic review of large language models in clinical medicine, Nature medicine (2026) 1–8. :Preprint submitted to Elsevier Page 11 of 11

2026

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.

[1] [1]

X. Hou, Y. Zhao, Y. Liu, Z. Yang, K. Wang, L. Li, X. Luo, D. Lo, J.Grundy,H.Wang,Largelanguagemodelsforsoftwareengineering: A systematic literature review, ACM Transactions on Software Engineering and Methodology 33 (2024) 1–79

2024

[2] [2]

A.J.Thirunavukarasu,D.S.J.Ting,K.Elangovan,L.Gutierrez,T.F. Tan, D. S. W. Ting, Large language models in medicine, Nature medicine 29 (2023) 1930–1940

2023

[3] [3]

R. Li, X. Wang, D. Berlowitz, J. Mez, H. Lin, H. Yu, Care-ad: a multi-agent large language model framework for alzheimer’s disease prediction using longitudinal clinical notes, npj Digital Medicine 8 (2025) 541

2025

[4] [4]

Q. Peng, Y. Cai, J. Liu, Q. Zou, X. Chen, Z. Zhong, Z. Wang, J. Xie, Q.Li, Integrationofmulti-sourcemedicaldataformedicaldiagnosis question answering, IEEE Transactions on Medical Imaging 44 (2024) 1373–1385

2024

[5] [5]

Y.Zheng,W.Gan,Z.Chen,Z.Qi,Q.Liang,P.S.Yu, Largelanguage models for medicine: a survey, International Journal of Machine Learning and Cybernetics 16 (2025) 1015–1040

2025

[6] [6]

C.J.Kelly,A.Karthikesalingam,M.Suleyman,G.Corrado,D.King, Key challenges for delivering clinical impact with artificial intelli- gence, BMC medicine 17 (2019) 195

2019

[7] [7]

H. Shi, Z. Xu, H. Wang, W. Qin, W. Wang, Y. Wang, Z. Wang, S. Ebrahimi, H. Wang, Continual learning of large language models: A comprehensive survey, ACM Computing Surveys 58 (2025) 1–42

2025

[8] [8]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,Ł.Kaiser,I.Polosukhin, Attentionisallyouneed, Advances in neural information processing systems 30 (2017)

2017