Supplementary Table S1

Recursive Self-Improvement in Healthcare AI Agents
Study-level extraction for 85 studies. The unified four-source search of 4 November 2026 re-confirmed31 previously charted studies and added52 (34 from the initial three sources,18 from OpenAlex); two comparators were retained purposively.

Publication status, mechanism evidence and clinical evaluation are separate fields. “Not reported” refers to the examined source version. Clinical-setting classification follows the reported study evaluation design. Coach Iris is retrospective; its operating patient service is documented separately. Source versions and section/table locators are given below. Author spotchecks remain pending.

1. Agent Hospital / MedAgent-Zero

Key
medhospital2024
Name
Agent Hospital / MedAgent-Zero
Title
Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents
Authors
Junkai Li; Yunghwei Lai; Weitao Li; Jingyi Ren; Meng Zhang; Xinhui Kang; Siyu Wang; Peng Li; Ya-Qin Zhang; Weizhi Ma; Yang Liu
Year
2024
Publication status
Preprint examined; no formal publication verified in this extraction
Version
v3
Source url
https://arxiv.org/abs/2405.02957v3
DOI
Update target
Medical case database and natural-language experience principles.
Fixed components
LLM weights; prompting, retrieval and reflection/validation design.
Feedback
Synthetic patient reference decisions; correctness on medical-document exemplar questions.
Timing
Across simulated patient encounters; no parameter training.
Retained artifact
Successful case-answer pairs and validated/refined principles.
Update generator
Frozen doctor LLM reflects on incorrect decisions.
Selection and release
Correct cases retained; principles tested and rejected when incorrect; exemplar validation described.
Evaluation separate from updating
Separate simulated test cases and MedQA; memory-component ablations.
Fixed-procedure control
No matched changing-versus-fixed improver comparison reported.
Evidence concerning improver change
Memory contents change; reflection/validation algorithm not reported as evolving.
Clinical setting
Virtual hospital and exam-question benchmarks.
Evidence boundary
The paper calls MedQA real-world alignment; this is not a prospective patient-care study.
Locators
Agent Evolution
Appendix B.3: reflection, validation and refinement
Appendix C.4: separate simulated training/test cases
Appendix C.7: ablations

2. ClinicalAgents

Key
clinicalagents2026
Name
ClinicalAgents
Title
ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory
Authors
Zhuohan Ge; Haoyang Li; Yubo Wang; Nicole Hu; Chen Jason Zhang; Qing Li
Year
2026
Publication status
Published: KDD 2026
Version
v2
Source url
https://arxiv.org/abs/2603.26182v2
DOI
10.1145/3770855.3818931
Update target
Current-patient evidence, hypotheses, action trajectory and search state.
Fixed components
Static guideline/case experience store; orchestration equations and prompts.
Feedback
Retrieved evidence, missing-information checks, hypothesis confidence and search rewards.
Timing
Within a case; multi-turn inference.
Retained artifact
Working state during consultation; no newly accumulated cross-case skill reported.
Update generator
LLM orchestrator with MCTS-style selection and backtracking.
Selection and release
Inference actions selected by search; no learned-artifact release process described.
Evaluation separate from updating
Test cases excluded from experience store; separate-model judge validated against clinicians.
Fixed-procedure control
Orchestration/memory ablations; no changing-improver control.
Evidence concerning improver change
State/search changes, not a reported learning-procedure update.
Clinical setting
MedChain, dialogue and medical QA benchmarks.
Evidence boundary
Static experience retrieval does not establish newly learned cross-case memory or skill synthesis.
Locators
Section 3.3.2: explicitly static experience memory
Section 3.3.1: working memory
Section 4.1.1: test exclusion
Section 4.2: clinician validation of judge

3. TissueLab

Key
coevolving2025
Name
TissueLab
Title
A Co-Evolving Agentic AI System for Medical Imaging Analysis
Authors
Songhao Li; Jonathan Xu; Tiancheng Bao; Yuxuan Liu; Yuchen Liu; Yihang Liu; Lilin Wang; Wenhui Lei; Sheng Wang; Yinuo Xu; Yan Cui; Jialu Yao; Shunsuke Koga; Zhi Huang
Year
2025
Publication status
Preprint examined; no formal publication verified in this extraction
Version
v1
Source url
https://arxiv.org/abs/2509.20279
DOI
Update target
Editable stored outputs; lightweight tool classifiers; model-ranking preferences.
Fixed components
Foundation models and orchestration framework; human-designed feedback mechanisms.
Feedback
Clinician annotations and accept/reject feedback.
Timing
Interactive refinement and reuse in subsequent executions; community contributions.
Retained artifact
HDF5 outputs/annotations/logs, adapted classifiers, model preferences and contributed tools.
Update generator
LLM workflow/code generation; clinician-guided active learning and model reselection.
Selection and release
Human output review and feedback guide switching; uniform held-out release gate not specified.
Evaluation separate from updating
Task benchmarks; separate pathology hold-out and patient-level fatty-liver split described.
Fixed-procedure control
No isolated changing-versus-fixed improver comparison reported.
Evidence concerning improver change
Ranking preferences adapt; later-learning benefit of changing the improvement procedure is not isolated.
Clinical setting
Retrospective imaging/omics datasets with interactive expert annotation.
Evidence boundary
Human-guided tool adaptation is demonstrated; autonomous experience-to-skill synthesis is not clearly specified.
Locators
Methods: Community-driven co-evolution
Methods: Editable memory layer
Section 2.5: user-created/adapted tools
Methods: benchmarking and separation
Discussion: human supervision and clinical integration limitations

4. MACRO

Key
fan2026macro
Name
MACRO
Title
Evolving Medical Imaging Agents via Experience-driven Self-skill Discovery
Authors
Lin Fan; Pengyu Dai; Zhipeng Deng; Haolin Wang; Xun Gong; Yefeng Zheng; Yafei Ou
Year
2026
Publication status
Preprint examined; no formal publication verified in this extraction
Version
v1
Source url
https://arxiv.org/abs/2603.05860
DOI
Update target
Experience memory, composite-tool registry and agent policy parameters.
Fixed components
Frozen visual encoder; mining criterion and SFT/GRPO update rules.
Feedback
Training-label correctness gates memory; composite-tool invocation yields RL reward.
Timing
Across training episodes, during supervised warm-start and GRPO.
Retained artifact
Successful trajectory fragments, reusable tool sequences and LoRA policy.
Update generator
Trajectory extraction, frequent-subsequence mining and policy optimization.
Selection and release
Correct episodes update memory; frequency threshold admits composites; no separate clinical release gate specified.
Evaluation separate from updating
Imaging benchmarks and component ablations; RAM-W600 test-set results.
Fixed-procedure control
Component ablations share hyperparameters; no isolated improver-change control.
Evidence concerning improver change
Composite registry changes actions/reward eligibility; mining/update rules stay specified; improver gain not isolated.
Clinical setting
Public imaging datasets; offline training/evaluation.
Evidence boundary
Training-time co-evolution does not establish ongoing clinical deployment adaptation.
Locators
Section 3.2: correct training episodes enter memory
Section 3.3: composite registration
Section 3.4: SFT/GRPO and reward
Section 4.3: ablations
Section 5: prospective studies remain future work

5. DiagAgent / DiagGym

Key
qiu2025diaggym
Name
DiagAgent / DiagGym
Title
Evolving Interactive Diagnostic Agents in a Virtual Clinical Environment
Authors
Pengcheng Qiu; Chaoyi Wu; Junwei Liu; Qiaoyu Zheng; Yusheng Liao; Haowen Wang; Yun Yue; Qianrui Fan; Shuai Zhen; Jian Wang; Jinjie Gu; Yanfeng Wang; Ya Zhang; Weidi Xie
Year
2025
Publication status
Preprint examined; no formal publication verified in this extraction
Version
v2
Source url
https://arxiv.org/abs/2510.24654v2
DOI
Update target
Diagnostic agent parameters and examination-selection policy.
Fixed components
Previously trained DiagGym environment, reward specification and GRPO protocol.
Feedback
Simulated examination results; diagnosis match, examination alignment and turn-budget rewards.
Timing
Offline cold-start training followed by interactive simulated RL.
Retained artifact
Trained diagnostic-agent weights.
Update generator
SFT then GRPO using generated diagnostic trajectories.
Selection and release
Reward-based training updates; separate deployment release gate not described.
Evaluation separate from updating
Held-out and out-of-domain DiagBench cases; physician-written rubrics and judge validation.
Fixed-procedure control
SFT/reward ablations; no changing-versus-fixed improver comparison.
Evidence concerning improver change
Diagnostic policy learns; reward/update-generator changes are not evaluated.
Clinical setting
Retrospective EHR/case-report-derived virtual diagnostic trajectories.
Evidence boundary
Physician validation and clinical source records do not establish prospective clinical utility.
Locators
Section 5.2: training and fixed rewards
Section 5.3: benchmark construction
Appendix A.1.3: physician validation of metrics
Section 3.3: ablations
Section 4: virtual environment limitations

6. DoctorAgent-RL

Key
feng2025doctoragentrl
Name
DoctorAgent-RL
Title
DoctorAgent-RL: A Multi-Agent Collaborative Reinforcement Learning System for Multi-Turn Clinical Dialogue
Authors
Yichun Feng; Jiawei Wang; Lu Zhou; Zhen Lei; Yixue Li
Year
2026
Publication status
verified conference publication; extended journal version not verified
Version
ICASSP 2026 metadata; mechanism extraction from matching-title/five-author arXiv v3 (2025-10-14), not extended v4
Source url
https://arxiv.org/html/2505.19630v3
DOI
10.1109/ICASSP55912.2026.11460976
Update target
Doctor-agent questioning/diagnosis policy parameters.
Fixed components
Patient simulator, consultation reward components and prescribed training algorithm.
Feedback
Simulated patient responses; diagnosis/treatment word-F1, information gain and protocol rewards.
Timing
Offline SFT followed by simulated multi-turn GRPO training.
Retained artifact
Fine-tuned doctor-agent weights.
Update generator
Behavior cloning and GRPO under sampled dialogue-turn budgets.
Selection and release
Training objectives update policy; separate clinical release gate not reported.
Evaluation separate from updating
Held-out MTMedDialog; model judge and blinded human assessment of sampled dialogues.
Fixed-procedure control
SFT/RL/turn-budget ablations; no evolving-improver control.
Evidence concerning improver change
Questioning strategy learns; training/evaluation procedure not reported as self-modified.
Clinical setting
Simulated consultations from historical dialogue data.
Evidence boundary
Do not attribute extended v4 clinical experiments to the conference paper.
Locators
Official conference title/authors
Training Procedure
Consultation Evaluator
Blinded human evaluation
Training-component ablations

7. OAIMS

Key
xu2026oaims
Name
OAIMS
Title
You Point, I Learn: Online Adaptation of Interactive Segmentation Models for Handling Distribution Shifts in Medical Imaging
Authors
Wentian Xu; Ziyun Liang; Harry Anthony; Yasin Ibrahim; Felix Cohen; Guang Yang; Konstantinos Kamnitsas
Year
2026
Publication status
Published: ICLR 2026
Version
v4
Source url
https://arxiv.org/abs/2503.06717v4
DOI
Update target
Interactive segmentation model parameters.
Fixed components
Architecture, adaptation losses and prescribed post-/mid-interaction update rules.
Feedback
Correction clicks and click-refined pseudo-label masks.
Timing
After each click and/or image; updated weights persist across image sequence.
Retained artifact
Adapted segmentation weights.
Update generator
Gradient updates with Dice-Focal and click-centered Gaussian losses.
Selection and release
Direct prescribed online updates; no separate release-validation cohort described.
Evaluation separate from updating
Ground-truth segmentation metrics on shifted target databases; baseline/ablation/robustness comparisons.
Fixed-procedure control
Frozen and other fixed adaptation baselines; no changing-improver comparison.
Evidence concerning improver change
Model adapts; update rule is not reported as self-modified.
Clinical setting
Retrospective imaging streams with automatically simulated correction clicks.
Evidence boundary
Non-LLM segmentation comparator; simulated clicks are not prospective clinician-use evidence.
Locators
Section 2.4: update timing
Section 3.2: online evaluation
Section 3.3: robustness
Appendix A.2: clicks simulated using ground truth
Official ICLR publication

8. MedRSI

Key
wu2026medrsi
Name
MedRSI
Title
MedRSI: Recursive Self-Improvement for Medical Agents via Clinically Aligned Self-Evolution
Authors
Junde Wu; Jiayuan Zhu; Minghao Hu; Fenglin Liu; Jiazhen Pan
Year
2026
Publication status
arXiv preprint; no peer-reviewed publication verified in this extraction
Version
v1 (2026-09-21)
Source url
https://arxiv.org/html/2609.24838v1
DOI
10.48550/arXiv.2609.24838
Update target
Tool registry, generated code, trained tool models and derived training resources.
Fixed components
Reasoning model, execution rules, judge/rubric and registration protocol.
Feedback
Diagnostic errors weighted by clinical consequence; paired candidate evaluations.
Timing
Successive retrospective patient batches and candidate-trial cohorts.
Retained artifact
Registered tool packages, dependencies, checkpoints and evaluation histories.
Update generator
Failure clustering/reflection; code generation, composition and model training.
Selection and release
Discovery screening then frozen-candidate trials; sustained benefit without increased clinical cost.
Evaluation separate from updating
Disjoint patient partitions; locked test evaluation after trajectory freezing.
Fixed-procedure control
Matched clinical-cost/registration ablations; no isolated evolving-improver control.
Evidence concerning improver change
New capabilities enable later development; improvement-procedure change/gain not independently isolated.
Clinical setting
Public imaging benchmarks and private retrospective cohorts.
Evidence boundary
Broad recursive capability growth; prospective workflow/patient-outcome validation remains future work.
Locators
Section 2.1: fixed M/P and changing registry
Section 2.2: frozen clinical evaluator
Section 2.3: discovery and registration
Section 3.1: separation and retrospective checkpoint testing
Table 6: matched mechanism ablations
Section 4: prospective validation needed

9. Agent Mental Clinic (AMC)

Key
lan2024amc
Name
Agent Mental Clinic (AMC)
Title
Depression Diagnosis Dialogue Simulation: Self-improving Psychiatrist with Tertiary Memory
Authors
Kunyao Lan; Bingrui Jin; Zichen Zhu; Siyuan Chen; Shu Zhang; Kenny Q. Zhu; Mengyue Wu
Year
2024
Publication status
arXiv preprint; no later journal/conference record identified in the current arXiv metadata and exact-title search (checked 2026-10-19)
Version
Submission history From: Kunyao Lan [ view email ] [v1] Fri, 20 Sep 2024 14:25:08 UTC (1,994 KB) [v2] Wed, 9 Oct 2024 04:37:29 UTC (1,994 KB)
Source url
https://arxiv.org/html/2409.15084v2
DOI
10.48550/arXiv.2409.15084
Update target
Diagnostic skills, encounter/EMR memory, and outcome-dependent memory-importance scores.
Fixed components
Frozen GPT-3.5; constant dialogue, diagnosis and skill-reflection prompts; fixed supervisor stages and relevance/importance sampling formula.
Feedback
Professional D4 risk labels, revealed after diagnosis, guide supervisor reflection; diagnostic correctness increments or decrements retrieved-memory importance.
Timing
After simulated training consultations (Quiz); accumulated memory is reused on subsequent patients and in the single-attempt Exam setting.
Retained artifact
Conversation records, summarized EMRs, reusable diagnostic skills, and memory-importance scores.
Update generator
LLM supervisor compares predicted and reference depression/suicide risk and produces corrective skills; a fixed rule updates importance.
Selection and release
Supervisor-generated skills are written to memory; later retrieval samples ten nodes using fixed relevance/importance weighting. The described workflow does not impose an independent clinical release gate.
Evaluation separate from updating
Separate D4 test cases are evaluated in a single-attempt Exam, using professional reference labels; original versus simulated dialogues and memory/supervisor ablations are reported.
Fixed-procedure control
Memory-off, memory-layer and supervisor ablations test component value; they do not compare an evolved skill-generation procedure with its fixed predecessor.
Evidence concerning improver change
not established: skills and retrieval priorities change under constant reflection prompts and fixed importance-update/sampling rules.
Clinical setting
Chinese depression/suicide-risk classification using D4 dialogue-derived patient simulations and original dialogues; 100 selected training cases and 132 original test cases are described.
Evidence boundary
Retrospective dialogue/simulation evidence, with limited four-class accuracy and poorer simulated dialogue performance. The study explicitly lacks real-time expert feedback and prospective patient assessment.
Locators
patient cohort, professional labels and supervisor role
six-stage loop; constant-prompt Eqs. 1–6; tertiary memory; retrieval Eq. 7
Quiz/Exam distinction, Table 1, memory and supervisor ablations
role-play, cultural coverage and absent real-time expert feedback
intended-use and privacy boundaries

10. MedAgentSim

Key
almansoori2025medagentsim
Name
MedAgentSim
Title
MedAgentSim: Self-Evolving Multi-Agent Simulations for Realistic Clinical Interactions
Authors
Mohammad Almansoori; Komal Kumar; Hisham Cholakkal
Year
2025
Publication status
Peer-reviewed MICCAI 2025 proceedings paper, LNCS 15968, pp. 362–372; publication date 2025-09-21 verified in Crossref.
Version
MICCAI 2025 accepted/published version (official OA PDF); latest arXiv counterpart v2, 2025-10-01.
Source url
https://papers.miccai.org/miccai-2025/paper/2575_paper.pdf
DOI
10.1007/978-3-032-05114-1_35
Update target
Medical-record and reflection-experience buffers used for later diagnostic retrieval.
Fixed components
Doctor/patient/measurement roles; KNN retrieval, chain-of-thought prompting, majority-vote ensemble, and correct-diagnosis storage rule; no model-weight update is part of the described method.
Feedback
Reference diagnosis determines whether the initial or reflected diagnosis is correct; benchmark outputs are matched to labels by an LLM evaluator with manual review of accuracy logs.
Timing
After each simulated consultation; stored records are retrieved during later consultations before ensemble diagnosis.
Retained artifact
CLIP-embedded successful case records and reflection insights from initially incorrect cases corrected on a second attempt.
Update generator
The doctor reflects on an incorrect diagnosis and retries; the record-store module embeds and retains eligible results.
Selection and release
Correct initial diagnoses enter the medical buffer; only successful corrected diagnoses contribute reflection insights to the experience buffer; unsuccessful retries are discarded.
Evaluation separate from updating
NEJM, MedQA and MIMIC-IV cases are converted to interactive simulations; LLM diagnosis scoring and data conversions are manually reviewed. A separate protected memory-development/test split is not specified in the full published Methods/Experiments.
Fixed-procedure control
Incremental measurement, memory, chain-of-thought and ensembling ablations compare task pipelines, not evolved versus unchanged experience-generation procedures.
Evidence concerning improver change
not established: experience buffers grow under an unchanged retrieval, reflection, voting and admission pipeline.
Clinical setting
Interactive doctor/patient/measurement simulations derived from public clinical cases and exam data; limited human-role identification assesses perceived realism, not clinical efficacy.
Evidence boundary
Benchmark/simulation results support memory-assisted diagnosis. The human-role study is not prospective care, and cumulative additions of memory/reasoning/ensembling cannot isolate recursive improver gains.
Locators
official publication status, author list, DOI, pages, and OA version identity
pp. 3–5, §2 and Fig. 2: two buffers, reflection and correctness admission
pp. 5–7, §3: benchmark conversion, diagnosis scoring, Table 1 and incremental Table 2
matching methodology in current arXiv counterpart
publisher metadata and 2025-09-21 publication date

11. GSEM

Key
han2026gsem
Name
GSEM
Title
GSEM: Graph-based Self-Evolving Memory for Experience Augmented Clinical Reasoning
Authors
Xiao Han; Yuzheng Fan; Sendong Zhao; Haochun Wang; Bing Qin
Year
2026
Publication status
Published: Proceedings of EMNLP 2026 (main), pp. 11637-11660; extraction based on arXiv v1
Version
Submission history From: Xiao Han [ view email ] [v1] Mon, 23 Mar 2026 15:24:15 UTC (5,083 KB)
Source url
https://arxiv.org/html/2603.22096v1
DOI
10.18653/2026.emnlp-main.507
Update target
Experience-node quality and inter-experience edge weights in a dual-layer graph; optional insertion of newly extracted experiences.
Fixed components
Backbone models; typed experience schema; graph traversal actions; fixed rank-credit, node/edge calibration equations, learning rates and construction-time connection rule.
Feedback
Scalar task-outcome feedback in [-1,1] recalibrates activated nodes and edges; held-out reliability trials initialize experience quality.
Timing
After each case, with reported snapshots after 50, 150 and 250 online updates.
Retained artifact
Indication/contraindication experiences, entity/experience graph, reliability scores, relation weights and accumulated edge-credit state.
Update generator
LLM extraction of successful strategies and failure divergences; fixed credit-assignment equations update graph weights from feedback.
Selection and release
Experience reliability validation initializes quality; fixed thresholds construct edges; online feedback reweights activated memories for subsequent traversal. No clinician release gate is part of the reported benchmark pipeline.
Evaluation separate from updating
Held-out MedR-Bench diagnosis/treatment cases and five MedAgentsBench subsets; DeepSeek-V3.2 judges reasoning and is also one generator. The relation of online-evolution samples to snapshot test cases is not fully specified in §5.3/Appendices.
Fixed-procedure control
Static GSEM is compared with evolving-memory snapshots, plus retrieval ablations; this isolates graph-state calibration, not a self-revised calibration procedure.
Evidence concerning improver change
not established: feedback changes retrieval-affecting state under fixed update equations; the extraction, traversal and calibration procedures are not shown to revise themselves.
Clinical setting
Public retrospective clinical-reasoning and medical-QA benchmarks; automatic evaluation rather than live clinical interaction.
Evidence boundary
Supports outcome-informed memory selection and graph calibration. It does not establish prospective benefit, clinician-validated updates, or improvement of the memory-update algorithm itself.
Locators
task/experience definitions and dual-layer graph
extraction, held-out Experience Reliability Validation and graph construction
update targets and fixed Eqs. 7–10
benchmark setup, shared backbone/judging and evolution Table 2
retrieval ablations and reasoning-quality analysis
explicit lack of real workflow, longitudinal and expert evaluation
fixed traversal and graph hyperparameters

12. Evo-MedAgent

Key
shen2026evomedagent
Name
Evo-MedAgent
Title
Evo-MedAgent: Beyond One-Shot Diagnosis with Agents That Remember, Reflect, and Improve
Authors
Weixiang Shen; Bailiang Jian; Jun Li; Che Liu; Johannes Moll; Xiaobin Hu; Daniel Rueckert; Hongwei Bran Li; Jiazhen Pan
Year
2026
Publication status
arXiv preprint; no later journal/conference record identified in the current arXiv metadata and exact-title search (checked 2026-10-19)
Version
Submission history From: Bailiang Jian [ view email ] [v1] Wed, 15 Apr 2026 23:12:02 UTC (508 KB)
Source url
https://arxiv.org/html/2604.14475v1
DOI
10.48550/arXiv.2604.14475
Update target
Episodic cases, diagnostic procedural rules/utility statistics and tool-trust state.
Fixed components
Frozen VLM and specialist tools; fixed base protocol, memory factorization, retrieval/utility selection and tool-trust thresholds.
Feedback
The benchmark ground-truth answer is revealed after each episode; multimodal reflection uses images, predictions, traces and that answer.
Timing
Test-time, between sequential chest-X-ray cases; later cases inherit earlier-case memory.
Retained artifact
Compressed episodes, priority-tagged procedural rules and usage statistics, plus per-tool helpful/harmful/misuse counts and trust labels.
Update generator
One post-case multimodal reflection call proposes procedural patches and governance updates; fixed accounting rules maintain utility and trust.
Selection and release
Relevant episodes and utility-ranked rules enter later prompts; tool labels use preset count/rate thresholds. The published experiment does not independently validate each proposed memory patch before reuse.
Evaluation separate from updating
ChestAgentBench answers score sequential predictions; three case-order permutations are reported for GPT-5-mini. This is online benchmark evaluation with prior-answer feedback, not a separately frozen-memory external clinical cohort.
Fixed-procedure control
No-memory and episodic/procedural ablations; fixed versus evolved reflection or memory-update procedures are not compared. Tool-governance quantitative ablation is explicitly deferred.
Evidence concerning improver change
not established: diagnostic procedures and tool-use preferences change; the mechanism generating and admitting future memory changes remains specified by the fixed reflection pipeline.
Clinical setting
Sequential retrospective chest-X-ray questions on ChestAgentBench, including MCQ and open-ended tasks; no model retraining.
Evidence boundary
Initial single-modality benchmark evidence with immediate ground truth. Tool-governance is described but not quantitatively established; delayed/noisy clinical feedback and prospective care remain untested.
Locators
sequential ground-truth reveal and fixed-parameter objective
three memory stores, utility selection, fixed tool-trust thresholds and deferred governance ablation
fixed base protocol and reflection-based patch updates
Table 1, sequential trajectories, three permutations and memory ablations
immediate-feedback, CXR-only and governance limitations

13. DxEvolve

Key
ren2026dxevolve
Name
DxEvolve
Title
Emulating Clinician Cognition via Self-Evolving Deep Clinical Research
Authors
Ruiyang Ren; Yuhao Wang; Yunsen Liang; Lan Luo; Jing Liu; Haifeng Wang; Cong Feng; Yinan Zhang; Chunyan Miao; Ji-Rong Wen; Wayne Xin Zhao
Year
2026
Publication status
arXiv preprint; no later journal/conference record identified in the current arXiv metadata and exact-title search (checked 2026-10-19)
Version
Submission history From: Ruiyang Ren [ view email ] [v1] Wed, 11 Mar 2026 11:41:51 UTC (1,166 KB)
Source url
https://arxiv.org/html/2603.10677v1
DOI
10.48550/arXiv.2603.10677
Update target
A growing repository of diagnostic cognition primitives (DCPs) that guides later test ordering and diagnostic decisions.
Fixed components
Frozen base LLM; shared action/prompt contract, DCP-construction template, embedding retriever, evaluator and bounded diagnostic action space.
Feedback
Accrual-case reference diagnoses, correctness flags, rule-based process feedback and charted clinician test orders guide DCP construction.
Timing
Post-encounter consolidation in a dedicated accrual pool; nested exposure-indexed repositories are tested on fixed, non-overlapping evaluation cohorts.
Retained artifact
DCPs with an experience pattern, test-ordering guidance, diagnostic-decision guidance and provenance including exposure index, pathology and source outcome.
Update generator
The same backbone uses a fixed reflection template to distill completed encounters into reusable DCPs.
Selection and release
DCPs are stored and retrieved by dense pattern similarity; the agent is instructed to use them only when compatible with acquired evidence. Clinician ratings assess sampled DCPs after construction, rather than serving as per-update release approval.
Evaluation separate from updating
Non-overlapping 400-case MIMIC-CDM evaluation; a separate 80-case reader-study split; a 293-case external hospital cohort excluded from accrual; two blinded physicians rate 40 sampled DCPs.
Fixed-procedure control
Backbone-matched DCP-free and retrieval ablations plus exposure curves isolate experiential reuse; no changed-versus-fixed DCP-construction procedure is evaluated.
Evidence concerning improver change
not established: later DCPs receive stronger clinician ratings and improve reuse, but maturation of stored guidance is not evidence that the fixed consolidation/retrieval procedure modifies itself.
Clinical setting
Retrospective EHR-based interactive workups for acute abdominal conditions, plus cross-institution/cross-language external validation; imaging is represented by narrative reports.
Evidence boundary
Stronger external retrospective evidence than simulation-only studies; no prospective patient-outcome trial. The published clinician reference uses full information, so it is not a matched-information superiority comparison.
Locators
workflow, three-part DCP, provenance and evidence-compatible reuse
accrual/evaluation separation, reader-study cohort and external hospital cohort
frozen backbones, common prompt contract, fixed dense retrieval and baselines
nested accrual pools, blinded clinician ratings and process metrics
fixed experience-construction prompt, ground truth and rule-based feedback
unmatched information access for published clinician reference
independent external validation
retrospective scope and prospective-study priorities

14. EvoClinician

Key
he2026evoclinician
Name
EvoClinician
Title
EvoClinician: A Self-Evolving Agent for Multi-Turn Medical Diagnosis via Test-Time Evolutionary Learning
Authors
Yufei He; Juncheng Liu; Zhiyuan Hu; Yulin Chen; Yue Liu; Yuan Sui; Yibo Li; Nuo Chen; Jun Hu; Bryan Hooi; Xinxing Xu; Jiang Bian
Year
2026
Publication status
Preprint; no formal publication identified in arXiv record or author repository
Version
arXiv v1, 2026-01-30
Source url
https://arxiv.org/html/2601.22964v1
DOI
10.48550/arXiv.2601.22964
Update target
Actor prompt rules and action-conditioned external memories.
Fixed components
Base-model weights; Process Grader and Evolver instructions; preset memory-deletion coefficients and budgets.
Feedback
Diagnosis Judge grade, encounter costs, and action-level yield/efficiency critiques.
Timing
Between diagnostic episodes; memories retrieved during subsequent encounters.
Retained artifact
Bounded operational rules and context-action-outcome memories, including error examples.
Update generator
Fixed-instruction LLM Evolver edits Actor state from graded trajectories.
Selection and release
Budgeted add/delete/merge operations; generalize identifying details; no independent candidate-release test specified.
Evaluation separate from updating
Separate diagnosis Judge, but its score feeds adaptation; the reported fixed-order case stream is not an independent frozen holdout.
Fixed-procedure control
Static, memory-only, prompt-only and transcript-feedback controls; these isolate components, not evolving versus fixed Evolver.
Evidence concerning improver change
Not established: Actor strategy changes; the described Evolver/Grader procedure is not revised.
Clinical setting
Med-Inquire simulation over 915 published clinical cases; generated missing details are not clinical ground truth.
Evidence boundary
Supports sequential prompt/memory adaptation and resource-aware diagnosis in simulation, not patient-care safety or improved update-generation ability.
Locators
Sections 4.1–4.5; update access and mechanism
Tables 1–3; fixed-order evaluation and component ablations
fixed Evolver prompt and memory deletion
simulation and safety limitations
reference implementation README; no acceptance claim

15. EvoMDT

Key
liu2026evomdt
Name
EvoMDT
Title
EvoMDT: a self-evolving multi-agent system for structured clinical decision-making in multi-cancer
Authors
Qicai Liu; Zhichao Hu; Tao Huang; Yupeng Niu; Xinche Zhang; Shanwu Ma; Chutong Lin; Goh Kim Huat; Hyeokkoo Eric Kwon; Feng Gao; Xianfu Sun; Zhitao Ying; Guangliang Qiang
Year
2026
Publication status
Peer-reviewed journal article; npj Digital Medicine 9, 124
Version
Published 2026-01-07; version of record 2026-02-04
Source url
https://www.nature.com/articles/s41746-025-02304-8
DOI
10.1038/s41746-025-02304-8
Update target
Prompts, consensus weights and retrieval scope (Abstract); post-task agent/workflow structure (Methods).
Fixed components
Clinical role/schema design, guideline governance and assessment rubrics; no weight-training step specified for self-evolution.
Feedback
Expert development-case feedback, performance metrics and reported outcome signals.
Timing
Prompt development plus reported post-task instruction/workflow revision.
Retained artifact
Revised prompts and coordination/retrieval configurations; versioned guideline knowledge base.
Update generator
Reported automatic prompt/workflow optimizer with expert-validated initial role prompts.
Selection and release
Development prompts receive oncologist review; automatic candidate acceptance and rollback criteria are not detailed in Methods or supplement.
Evaluation separate from updating
Six QA datasets, retrospective hospital records and single-blind physician scoring; these assess recommendations, not improvement cycles.
Fixed-procedure control
No evolution-disabled or frozen-optimizer comparison identified in the main results and 16-page supplement.
Evidence concerning improver change
Reported configuration adaptation; improved ability of the updating procedure itself is not established.
Clinical setting
405 retrospective oncology cases from four hospitals; structured case dossiers and blinded physician assessment.
Evidence boundary
Supports evaluated final-system recommendation quality; does not isolate self-evolution gains or establish prospective patient benefit.
Locators
Multi-agent EvoMDT; prompt optimization and self-evolution
four-center, 405-case retrospective design
single-blind physician protocol
Discussion and limitations
16-page supplement; guideline catalog, cohort figures and evaluation rubric

16. PsychAgent

Key
yang2026psychagent
Name
PsychAgent
Title
PsychAgent: An Experience-Driven Lifelong Learning Agent for Self-Evolving Psychological Counselor
Authors
Yutao Yang; Junsong Li; Qianjun Pan; Jie Zhou; Kai Chen; Qin Chen; Jingyuan Zhao; Ningning Zhou; Xin Li; Liang He
Year
2026
Publication status
Preprint; no formal publication identified in arXiv record or author repository
Version
arXiv v3, 2026-04-28; v1 2026-04-01
Source url
https://arxiv.org/html/2604.00931v3
DOI
10.48550/arXiv.2604.00931
Update target
Client memory, therapeutic skill tree and shared counseling-model parameters.
Fixed components
Therapeutic hierarchy, reward-based session selection and rejection-fine-tuning objective.
Feedback
LLM-rated counseling trajectories; expert-annotated tasks supervise initial memory, planning and skill curation.
Timing
Memory/skill revision across sessions; parameter internalization after selected multi-session rollouts.
Retained artifact
Client profiles, summaries, atomic skills and fine-tuned model weights.
Update generator
Learned skill extractor/manager appends or merges skills; best-of-eight rollouts supply RFT data.
Selection and release
Highest-reward session continues the trajectory; selected trajectories train the model. No independent skill-release test described.
Evaluation separate from updating
Separate evaluation profiles plus human ratings of 522 dialogues by two psychology-trained graduate raters; simulated clients.
Fixed-procedure control
Memory/planning, skill-evolution and internalization ablations; no dedicated control isolating shared-model improvements as an updater.
Evidence concerning improver change
Reported shared-model learning can affect skill extraction/management; improved update-generation ability is not independently established.
Clinical setting
PsychEval simulated multi-session counseling; 120 training and 20 evaluation profiles per therapeutic school.
Evidence boundary
Response-quality and memory/skill benefits; not measured therapeutic outcomes. Public release lacks the complete post-session skill-evolution pipeline.
Locators
three engines; equations 4–6
profile split and implementation
Table 3 ablations
human-rater qualifications and matched dialogue sample
README: incomplete public skill-evolution pipeline

17. SkeMex

Key
sun2026skemex
Name
SkeMex
Title
Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory
Authors
Haoran Sun; Wenjie Li; Yujie Zhang; Zekai Lin; Fanrui Zhang; Kaitao Chen; Xingqi He; Yichen Li; Mianxin Liu; Lei Liu; Yankai Jiang
Year
2026
Publication status
Accepted NeurIPS 2026 poster, officially listed; proceedings publication pending as of 2026-10-19
Version
Audited arXiv v4, 2026-10-09; v1 2026-06-08
Source url
https://arxiv.org/html/2606.09365v4
DOI
10.48550/arXiv.2606.09365
Update target
Skill contents, context-conditioned utility values and repository membership/status.
Fixed components
Backbone weights; retrieval formula, buffer gates, utility update schedule and governance rules.
Feedback
Benchmark returns, skill-adoption judgments and clinical-category reward baselines.
Timing
Retrieve once per episode; write/value after 30-trajectory windows; governance every two windows.
Retained artifact
General, task-level and action-level skills with utility and lifecycle metadata.
Update generator
Fixed-backbone two-pass skill analysis/mutation plus prescribed valuation/governance operators.
Selection and release
Gated trajectory buffer; novelty/quality checks; utility-based promotion, merging, deprecation and capacity limits.
Evaluation separate from updating
Offline held-out in-domain and unseen-benchmark-family tests; cross-backbone skill transfer; separate online repeated-stream experiments.
Fixed-procedure control
Matched-tool/prompt/inference-budget ablations freeze writing, valuation, governance or retrieval components; no evolving-operator comparator.
Evidence concerning improver change
Not established: learned utility/state changes under specified update equations; operator rules remain prescribed.
Clinical setting
Nine medical benchmark families, including multimodal and simulated clinical interaction; not prospective patient care.
Evidence boundary
Supports utility-guided skill memory and transfer; online repeated-epoch gains are distinct from independent clinical validation.
Locators
equations 5–9; writing, valuation and governance
Table 2 lifecycle ablations; Table 3 online epochs
Table 10 split and fixed hyperparameters
OOD and cross-backbone transfer
limitations
official NeurIPS 2026 Poster listing
linked from official conference listing

18. HealthFlow

Key
zhu2026healthflow
Name
HealthFlow
Title
HealthFlow: automating electronic health record analysis via a strategically self-evolving multi-agent framework
Authors
Yinghao Zhu; Zixiang Wang; Yifan Qi; Lei Gu; Dehao Sui; Haoran Hu; Xichen Zhang; Ziyi He; Yasha Wang; Junjun He; Liantao Ma; Lequan Yu
Year
2026
Publication status
Peer-reviewed journal article; npj Digital Medicine 9, 660
Version
Published 2026-08-17; version of record 2026-08-27; supersedes arXiv:2508.02621 title
Source url
https://www.nature.com/articles/s41746-026-03097-0
DOI
10.1038/s41746-026-03097-0
Update target
Persistent safeguards, workflows, dataset anchors, code snippets and memory confidence/status.
Fixed components
Four-agent orchestration, API-model weights, outcome-dependent writeback rules and type-specific retrieval gates.
Feedback
Execution artifacts/traces, methodological-validity verdicts and recovered/failed/successful task outcomes.
Timing
Evaluator-guided within-task retries; reflector-driven memory updates after each task.
Retained artifact
Typed, scoped experience records with provenance, applicability and confidence.
Update generator
Fixed-model reflector synthesizes guidance and validates, weakens or retires previous memory.
Selection and release
Outcome-aware writeback restricts failed runs to safeguards; scope/compatibility gates bound later retrieval.
Evaluation separate from updating
External Gemini judge with expert-built references; 12 blinded experts assess 20 reports. Online memory updates continue within each benchmark.
Fixed-procedure control
Progressive addition of planner/evaluator, reflector-memory and tools; no separately evolving-reflector versus fixed-reflector experiment.
Evidence concerning improver change
Not established: task procedures and memories evolve within prescribed reflection/retrieval/writeback rules.
Clinical setting
Public de-identified EHR analysis and reasoning benchmarks; qualitative offline hospital-pilot feedback.
Evidence boundary
Supports governed analytical reuse and artifact quality. The pilot does not establish prospective clinical outcome benefit.
Locators
Algorithm 1; within-task versus cross-task indices
equations 2–4; retrieval and outcome-specific writeback
independent judging, sequential online evaluation and human review
progressive ablation
limitations
supplementary judge agreement and repair dynamics

19. EvoPatient

Key
du2025evopatient
Name
EvoPatient
Title
LLMs Can Simulate Standardized Patients via Agent Coevolution
Authors
Zhuoyun Du; Lujie Zheng; Renjun Hu; Yuyang Xu; Xiawei Li; Ying Sun; Wei Chen; Jian Wu; Haolei Cai; Haochao Ying
Year
2025
Publication status
Peer-reviewed ACL 2025 main-conference long paper
Version
ACL Anthology proceedings version, 2025-07; pages 17278–17306
Source url
https://aclanthology.org/2025.acl-long.846.pdf
DOI
10.18653/v1/2025.acl-long.846
Update target
Patient demonstration/requirement library and doctor questioning-trajectory library.
Fixed components
Model weights, initial standardized-patient requirements, quality criteria and designed recruitment topology.
Feedback
Validated question/answer quality and subsequent usefulness of retrieved library entries.
Timing
Dialogue-level storage and reuse during simulated encounters; persist across cases.
Retained artifact
Attention library of question-record-answer-requirement tuples and doctor transition trajectories.
Update generator
Evolving doctor questioning generates patient-learning interactions; attention agent extracts requirements.
Selection and release
Quality validation before insertion; keep stronger duplicates; remove entries after two harmful uses.
Evaluation separate from updating
Separate disease-transfer assessment and human/GPT-4 preferences; no prospective human-trainee clinical-outcome test.
Fixed-procedure control
Tables 4–5 test doctor evolution and recruitment; patient-impact evolved-doctor condition bundles trajectory, question-pool and profile components.
Evidence concerning improver change
Demonstrated at interaction-generator level: evolved doctors alter patient-training experience. Recursive improvement of the whole coevolution procedure is not established.
Clinical setting
Simulated standardized patients grounded in hospital records and public data; medical-education task.
Evidence boundary
Supports co-adaptive training interactions. Bundled ablations and fixed evaluation/coevolution rules preclude an unqualified strict-RSI claim.
Locators
Section 3.3, pp. 17281–17282; two evolving libraries
Section 4.3, p. 17285; transfer from one disease to five
Section 4.4, Tables 4–5, pp. 17285–17286; generator and patient-impact ablations
Appendix E, p. 17293; correction/deletion rules
Algorithm 1, p. 17294; quality validation
complete authors, venue, pages and DOI

20. HealthClaw

Key
li2026healthclaw
Name
HealthClaw
Title
A Self-Evolving Agent for Longitudinal Personal Health Management
Authors
Haoran Li; Jiebi Deng; Tong Jin; Jinghong Han; Yuxin Wang; Zexin Wang; Qingyi Si; Weikang Gong; Xiahai Zhuang; Jia You; Wei Cheng; Jianfeng Feng; Hongcheng Guo
Year
2026
Publication status
Preprint
Version
v1
Source url
https://arxiv.org/abs/2607.13940v1
DOI
Update target
User profile, reusable procedures and episodic records.
Fixed components
Shared safety rules and domain knowledge; memory-induction and routing design.
Feedback
Completed episode content, evolving user state and memory-curation rules.
Timing
Writeback after episodes; writeback disabled during evaluation probes.
Retained artifact
User-specific profile, standard operating procedures and episodic memory.
Update generator
LLM induction routes information to profile, procedure, episode or exclusion.
Selection and release
Memory routing and privacy controls; no separate validated improvement-procedure release reported.
Evaluation separate from updating
Hidden-ledger probes on a synthetic year; nine biomedical task datasets.
Fixed-procedure control
Memory/context baselines; biomedical comparisons also differ in backbone and pipeline.
Evidence concerning improver change
Memory contents evolve under a fixed induction design.
Clinical setting
Synthetic 365-day histories for 20 users; offline biomedical benchmarks.
Evidence boundary
Longitudinal simulation is not prospective patient follow-up; task gains do not isolate memory causally.
Locators
Introduction
Results
Discussion
Methods

21. EMR

Key
shi2026emr
Name
EMR
Title
EMR: Self-Evolving Medical Multi-Agent System via Experience Mining and Reuse
Authors
Dongsheng Shi; Yue Li; Xin Yi; Linlin Wang
Year
2026
Publication status
Preprint
Version
v1
Source url
https://arxiv.org/abs/2609.15161v1
DOI
Update target
Clinical principles, diagnostic patterns and representative cases.
Fixed components
LLM weights, role prompts, embedding retrieval and similarity-threshold update rule.
Feedback
Ground-truth answers applied to generated multi-agent trajectories.
Timing
Offline iterative experience mining before benchmark evaluation.
Retained artifact
Hierarchical experience library containing positive insights and warnings.
Update generator
LLM extracts experience, then adds, merges or discards entries.
Selection and release
Similarity threshold and LLM consolidation govern library updates.
Evaluation separate from updating
Mining and testing subsets separated; most datasets evenly resplit, PubMedQA uses official test set.
Fixed-procedure control
Experience/hierarchy ablations and cross-backbone transfer; no evolving-versus-fixed updater test.
Evidence concerning improver change
Experience contents change; update algorithm remains fixed.
Clinical setting
Offline medical question-answering benchmarks.
Evidence boundary
Custom halves of several datasets do not match all published official test protocols; no clinical outcomes.
Locators
Section 3.2: mining
Section 3.3: update and retrieval
Appendix A: mining/test split
Limitations

22. MedForge-RSI

Key
chen2026medforge
Name
MedForge-RSI
Title
MedForge-RSI: Medical Deepfake Detection via Recursive Self-Improvement
Authors
Zhihui Chen; Mengling Feng
Year
2026
Publication status
Preprint
Version
v1
Source url
https://arxiv.org/abs/2609.36549v1
DOI
Update target
Error-derived experience, image-analysis tools and optional directive rules.
Fixed components
Detector weights, official system prompt, scorer, evaluation sets and learning-stream schedule.
Feedback
Verified labels on each learning batch; replay and validation performance.
Timing
Twenty simulated deployment rounds.
Retained artifact
Accepted memory entries, executable tools and reversible directive edits.
Update generator
Detector or external/ensemble reflector proposes structured edits.
Selection and release
Replay plus disjoint validation A/B checks; rollback variants; verifier-failure fallback also reported.
Evaluation separate from updating
Final 4,000-image held-out evaluation, separate from validation and learning.
Fixed-procedure control
Fixed configurations and mechanism ablations; the improvement mechanism itself is protected.
Evidence concerning improver change
Authors explicitly claim bounded self-refinement, not modification of their own improvement mechanism.
Clinical setting
Retrospective medical-image forgery detection benchmark.
Evidence boundary
One resample set is contaminated; final held-out test is distinct. Some configurations/seeds were incomplete.
Locators
Section 3.4: protected components and gates
Section 4: data splits and contaminated resample
Section 7: limitations and bounded claim

23. VIBEMed

Key
zhang2026vibemed
Name
VIBEMed
Title
Toward Vibe Medicine: A Self-Evolving Multi-Agent Framework for Clinical Decision Support
Authors
Qianxue Zhang; Yiming Ren; Shihuan Qin; Xiao Zhang; Liao Zhang; Jinyang Huang; Zhengliang Liu; Chenbin Liu; 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
Year
2026
Publication status
Published: Meta-Radiology 4(2):100223 (2026)
Version
v2
Source url
https://arxiv.org/abs/2606.15504v2
DOI
10.1016/j.metrad.2026.100223
Update target
Memory, model parameters and generated code/routing described at different timescales.
Fixed components
Three-role framework and prescribed evaluation/deployment workflow.
Feedback
Completed cases, reference treatment samples and model judges.
Timing
Post-encounter memory; scheduled parameter adaptation and on-demand code extension described.
Retained artifact
Memory, adapted model and validated code; quantitative evolution test concerns parameter tuning.
Update generator
Clinical Evolution Manager and backbone LLM; SFT/DPO and code repair described.
Selection and release
Model-judge screening and rollback; sandboxed code tests followed by user confirmation.
Evaluation separate from updating
MedBench base-versus-tuned comparison and clinical case demonstration.
Fixed-procedure control
No component-isolating memory/model/code evolution ablation; explicitly deferred.
Evidence concerning improver change
Code evolution is described; improved effectiveness of the update procedure is not isolated.
Clinical setting
Benchmark evaluation with a 1.5B model and an illustrative oncology case.
Evidence boundary
MedAnalysis and MedDiag decline after tuning; a case demonstration does not establish prospective utility.
Locators
Section 2.3: three-level mechanism
Section 3.2 and Table 2: evolution evaluation
Section 3.3: case demonstration
Conclusions: untested component effects
Published metadata; extraction uses arXiv v2

24. ECG-RSI

Key
wang2026ecgrsi
Name
ECG-RSI
Title
ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution
Authors
Xiaoliang Wang; Zhuo Chen; Chuanyang Zheng; Jiangwei Lao; Quanbin Wang; Honglin Lin; Jiajia Liu; Zhihong Lu; Hao Jin; Jian Wang; Jinjie Gu; Lijun Wu
Year
2026
Publication status
arXiv preprint; publication not verified
Version
v1 (2026-10-13)
Source url
https://arxiv.org/html/2610.17718v1
DOI
10.48550/arXiv.2610.17718
Update target
Learner weights; synthesis operators, prompts, reusable skills and data recipes.
Fixed components
Acceptance gates; vision encoder/aligner; selection scoring and SFT protocol.
Feedback
Waveform remeasurement/construction failures; learner errors on separate feedback data.
Timing
Offline synthesis-repair cycles and three model-training rounds.
Retained artifact
Versioned generation rules, accepted examples, selected checkpoints and lineage.
Update generator
Prompted local repairs and feedback-conditioned same-parent recipe branches.
Selection and release
Independent operator/question regression checks; selection-set score and regression gates.
Evaluation separate from updating
Patient-disjoint sources/development/tests; two physicians review synthetic samples, not clinical outcomes.
Fixed-procedure control
Yes: frozen-operator/skill ablations; equal accepted-data volume and matched training, unequal generation counts.
Evidence concerning improver change
Generator-rule changes supported by synthesis-yield and downstream-performance comparisons.
Clinical setting
Retrospective ECG datasets, synthesis and benchmark evaluation.
Evidence boundary
V1 reports code/data unreleased. Total-search compute parity not established; physician explanation support remains imperfect.
Locators
Section 3.1: separate learner, training pool and construction policy; data boundaries
Section 3.2.2: old/new operator and prompt repair comparisons
Appendix A.2: exact rule changes, proposer and independent regression checks
Prompt P5: repair generator component while acceptance criteria remain fixed
Table 10: 10,000 accepted samples; candidates 16,193 full, 36,723 frozen operators, 22,398 frozen skills
Main Table 15: R1 S7 full 43.22±1.40; frozen operators 39.29±1.30; frozen skills 40.15±1.30
Section 5.6/Table 11: same-parent fixed-mixture control
Section 5.7/Table 12: per-branch 10,000 additions and matched replay/training; not proof of total tree/chain compute equality
Table 7: 240 generated samples, two physicians; answer correctness 0.88, explanation support 0.76
Conclusion: unreleased implementation/data and diminishing external gains

25. MedZERO

Key
dang2026medzero
Name
MedZERO
Title
MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation
Authors
Xilin Dang; Weilin Ruan; Xue Yang; Jinghao Wang; Xiaowei Hu; Jinpeng Li; Pheng-Ann Heng
Year
2026
Publication status
ArXiv author-reported NeurIPS 2026 acceptance; proceedings not independently verified
Version
v1 (2026-10-06)
Source url
https://arxiv.org/html/2610.08327v1
DOI
10.48550/arXiv.2610.08327
Update target
Examiner generator parameters, Reasoner parameters and persistent generated knowledge graph.
Fixed components
Reward definitions, frontier-admission thresholds, causal visibility and curation protocol.
Feedback
Current Reasoner agreement/frontier scores, diversity and retrieved-evidence grounding.
Timing
Offline alternating GRPO rounds; counterpart frozen during each stage.
Retained artifact
Updated Examiner/Reasoner checkpoints, retained curriculum and filtered knowledge triplets.
Update generator
Trainable Examiner generates curriculum; Reasoner performance guides Examiner optimization.
Selection and release
Discard exploratory triplets; regenerate, filter valid/frontier-range samples, deduplicate and retain.
Evaluation separate from updating
Public benchmark gold labels; separate option-mapping judge. Curation is a separate stage, not independent truth adjudication.
Fixed-procedure control
Curriculum-source/curation/retrieval ablations; matched frozen-Examiner comparison not reported.
Evidence concerning improver change
Yes: the curriculum generator is optimized as the learner changes; isolated generator-change benefit remains unresolved.
Clinical setting
Open-ended academic medical reasoning benchmarks.
Evidence boundary
Incomplete grounding/consistency verification; benchmark gains do not demonstrate clinical utility.
Locators
ArXiv comment: accepted by NIPS 2026; date/authors
Section 3.1: alternating Examiner and Reasoner updates; counterpart frozen
Section 3.2: Examiner GRPO and fixed frontier/diversity/grounding reward
Section 3.3: causal visibility; temporary buffer discarded; separate curation pass
Section 3.4: fixed Reasoner consistency/grounding reward
Section 3.5: open-ended responses and separate mapping judge
Figure 3 inspected visually: panel c varies KG-only/Web-only/generated-triplets+Web curriculum source, not a frozen-Examiner control
Original Figure 3 SVG rendered for inspection
Section 5.3: three-round evolution
Appendix B: verification, persistent-error and clinical-utility limitations
Appendix F: fixed Reasoner, variable retrieval/prompt/observation budget; within-table comparison only

26. ReclAIm

Key
tzanis2026reclaim
Name
ReclAIm
Title
ReclAIm: A Multiagent Framework for Monitoring and Correcting Performance Decline in Medical Imaging AI
Authors
Eleftherios Tzanis; Michail E. Klontzas
Year
2026
Publication status
Peer-reviewed Technical Development article, Radiology: Artificial Intelligence 8(4):e250923; online 2026-06-03, issue 2026-07-01.
Version
Full methods and supplements read from arXiv:2510.17004v2 (2026-06-05); journal title, publication status, authors, DOI and abstract cross-checked with the publisher and Crossref.
Source url
https://arxiv.org/pdf/2510.17004v2
DOI
10.1148/ryai.250923
Update target
Downstream medical-image classifier weights, through full/partial/head-only fine-tuning or gradual unfreezing; the language-model agent itself is not the trained target.
Fixed components
GPT-4.1 reasoning engine; permanent system prompts, agent roles and predefined Python tools; fixed global decline threshold and parameter-anchoring objective. User prompts may specify the fine-tuning recipe.
Feedback
Ground-truth labels for monitoring images/scans produce accuracy, precision, recall and F1 metrics; relative decline against development-test metrics triggers intervention.
Timing
User-prompted monitoring cycles compare a separate inference cohort with development-test performance, then conditionally fine-tune on a distinct cohort and re-evaluate the same inference cohort.
Retained artifact
Trained/fine-tuned classifier checkpoints, configurations, evaluation/comparison JSON/CSV files, execution traces and conversational task context.
Update generator
A master LLM agent routes comparison and fine-tuning tasks to specialist agents, which invoke predefined training tools using user constraints or configured defaults.
Selection and release
A global 5% relative-decline heuristic flags any deteriorating metric; the fine-tuning tool selects checkpoints using validation performance and the workflow rechecks inference metrics. Clinical approval is recommended in discussion, not evaluated as a release gate.
Evaluation separate from updating
Patient-separated development, monitoring and adaptation cohorts are reported. Monitoring labels select adaptation and the same monitoring cohort measures its outcome, so post-update results are not from an untouched final holdout.
Fixed-procedure control
Before/after classifier performance is compared; the paper explicitly lacks scripted or single-agent baselines and direct catastrophic-forgetting assessment. It does not compare changed versus fixed improver procedures.
Evidence concerning improver change
not established: classifier parameters adapt while the LLM, permanent prompts, comparison rule and fine-tuning toolkit remain fixed; selecting a configured training strategy is not evidence of revising the improver.
Clinical setting
Retrospective public MRI, CT and chest-radiography classification datasets; one MIMIC-CXR temporal experiment, with the remaining datasets randomly partitioned from the same distribution.
Evidence boundary
A labeled monitoring-and-retraining prototype, not label-free shift detection or prospective deployment. Reused monitoring data, a single temporal dataset and inconsistent temporal/general split descriptions limit stronger generalization claims.
Locators
title, authors, volume 8/issue 4, online date and abstract
article e250923 and issue publication metadata
Experimental setup steps I–V: label-based monitoring, separate fine-tuning data, re-evaluation on inference data
5% trigger and eight affected model–dataset combinations
label availability, single temporal dataset, random partitions, no forgetting/control-pipeline tests
Supplement A: agent roles and fine-tuning strategies
fixed GPT-4.1, permanent prompt, parameter-anchoring loss
relative metric-change formula and threshold documentation
tool validation accuracy, early stopping and best checkpoint selection
Supplement B: general 60/20/20 with inner 70/15/15, alongside MIMIC chronological 40/gap20/recent40
Supplement C DenseNet121 prompts explicitly use distinct fine-tuning data and reuse the same inference file for post-update evaluation
example human-specified head-only recipe for cardiomegaly InceptionV3

27. PrivMeSA

Key
wang2026privmesa
Name
PrivMeSA
Title
PrivMeSA: Privacy-Aware Self-Evolving Multi-Agent System for Medicine via Local-Remote LLM Collaboration
Authors
Dannong Wang; Yuran Zhang; Bian Sun; Alex Stinard; Yuzhang Shang; Song Wang; Yu Tian
Year
2026
Publication status
P: arXiv preprint (v1, 29 Sep 2026) with no arXiv comment or journal_ref; listed on OpenReview as an ICLR 2027 Conference Submission (forum Ch62pc2CZs), no acceptance found; Crossref title search returned no record (checked 2026-11-03).
Version
v1
Source url
https://arxiv.org/abs/2609.38458v1
DOI
Update target
Two components change: a local lesson memory of distilled specialist answers, and the local policy's LoRA parameters trained by GRPO to decide when to consult, what to disclose, and whether to accept a retrieved lesson.
Fixed components
Remote GPT-5.6-Luna specialists, the frozen Gemma 4 12B distiller, judge and training attacker, the bge-small-en-v1.5 retriever with fixed thresholds (0.775 retrieval, 0.90 replacement), the GRPO recipe and reward weights (lambda_d=0.5, lambda_a=0.4) stay fixed.
Feedback
RL reward combines task accuracy against recorded MIMIC-IV-ED labels (diagnosis graded by CCS category) with discretion over injected synthetic personal details and registry-based log k-anonymity from an LLM attacker; lesson distillation uses no labels or reward.
Timing
Policy updates occur offline over 200 (Gemma) or 120 (Granite) GRPO steps; memory is updated after each completed episode (one sampled rollout per GRPO group during training, every case during test-time learning), taking effect only for later episodes.
Retained artifact
A LoRA adapter for the local policy and a lesson store of task, generalized question key, specialist-derived rule and provenance (2,026 lessons for Gemma and 1,470 for Granite at end of training).
Update generator
GRPO with group-mean-centred advantages updates the policy; a frozen Gemma 4 12B distiller converts each completed consultation exchange into a lesson or discards it, and near-duplicate same-task lessons are replaced.
Selection and release
Distilled lessons enter memory without outcome checking; at retrieval an LLM judge selects at most one of ten nearest lessons and verifies applicability in a second call, and the trained policy decides whether to use it; no clinical review or release gate is described.
Evaluation separate from updating
Patient-disjoint held-out MIMIC-IV-ED patients (100 for main results, 200 patients/897 encounters for test-time learning) with an evaluation attacker (GPT-5.6-Luna) different from the training attacker; no seeds, confidence intervals or significance tests are reported.
Fixed-procedure control
Ablations (untrained policy, no-probing training, local-only with and without RL, reward-term variants) and a paired memory-on versus memory-off test-time run; no comparison of a revised versus fixed distiller, judge, reward or training procedure.
Evidence concerning improver change
Not established: memory and policy parameters change, but the distiller, judge, retriever, GRPO recipe and reward are fixed throughout; the trained policy only indirectly shapes which exchanges are distilled, and this effect on later learning is not tested.
Clinical setting
Retrospective emergency-department replay of MIMIC-IV-ED encounters (disposition, primary ICD diagnosis by CCS category, coded procedure for admitted patients), with synthetic personal identifiers injected and re-identification measured against the de-identified 205,504-patient MIMIC-IV-ED registry.
Evidence boundary
Retrospective replay on de-identified records with synthetic personal details and a registry-conditional k measure; no clinician-facing, prospective or patient-outcome evaluation, and the test-time memory gain is mainly fewer remote messages at near-equal accuracy.
Locators
Section 3.4: lesson distillation, post-episode memory update, fixed retriever/judge/distiller and memory updates during RL training
Section 3.3: episode reward (accuracy, discretion, anonymity), GRPO optimization, frozen specialists/attacker/distiller/judge/retriever
Section 4.1: patient-disjoint split, registry, tasks, evaluation attacker, paired test-time learning protocol, baselines and implementation
Table 1: main accuracy, leak and k<=10 results for Gemma 4 12B and Granite 4.2 8B versus Delegation, PAPILLON and Remote only
Appendix E.1 (with Table 3): test-time memory growth, share answered from memory, messages per case, memory on versus off

28. TCM-SEA / TCM-MisDx

Key
wang2026tcmsea
Name
TCM-SEA / TCM-MisDx
Title
Learning from Expert-Corrected Misdiagnoses: A Self-Evolving Consultation Agent for Traditional Chinese Medicine after Failed Treatment
Authors
Ye Wang; Zhikang Lin; Jia Liu; Tianle Zhang; Nanyue Wang
Year
2026
Publication status
P: arXiv preprint (q-bio.OT), v1 only, submitted 2026-10-09; no comments or journal_ref on arXiv, and no Crossref, OpenReview or OpenAlex venue record found (OpenAlex lists only the arXiv preprint; checked 2026-11-03).
Version
v1
Source url
https://arxiv.org/abs/2610.40663v1
DOI
Update target
A free-text clinical skill document in the system prompt, which starts empty and is edited by add, replace or delete operations.
Fixed components
Frozen DeepSeek-Flash backbone, fixed task description, fixed Self-Harness loop and acceptance rule, fixed LLM judge, deterministic lookup patient simulator and 10-action budget.
Feedback
For each failed held-in consultation, per-metric outcomes plus the reference syndrome, treatment principle, error rubric and the monograph's original expert analysis with full prescription; the ablation sees only per-metric right/wrong.
Timing
Offline batch rounds (9 rounds, one proposal per round, seed 42) on fixed development sets before sealed test evaluation; no updating during test or deployment.
Retained artifact
The accepted skill-document text; two edits were accepted (rounds 1 and 6), both short pre-submission checklists.
Update generator
The same LLM that consults diagnoses failed held-in cases into signatures, clusters and ranks them into a brief with successful cases as regression checks, and proposes one minimal edit or declines.
Selection and release
An edit is accepted only if the summed composite score does not drop on either the 80-episode held-in set (drawn from training) or the 153-episode validation (held-out) set and rises on at least one; no clinician review before release is reported.
Evaluation separate from updating
Sealed 141-episode test split from the same three monographs, not used in selection, each final skill document run three times and averaged, scored by a fixed DeepSeek-Flash judge (temperature 0) plus exact-match and set-based metrics.
Fixed-procedure control
Base harness (empty document) and a learning-signal ablation (outcome-only feedback) under the same loop; no comparison of an evolved versus fixed edit-generating or gating procedure.
Evidence concerning improver change
Not established: the diagnose-propose-gate procedure, prompts and acceptance rule stay fixed across rounds, and only the skill document changes.
Clinical setting
TCM consultation after failed treatment (syndrome differentiation, recognition of the prior physician's error, corrective treatment principle), using 820 episodes digitized from three published misdiagnosis monographs with a deterministic simulator rather than live patients.
Evidence boundary
Retrospective, simulated benchmark from published case books with one backbone LLM (also used as judge), a single evolution seed and no reported confidence intervals, significance tests or physician assessment of the learned guidance.
Locators
Section 3, Source and episodes, and Table 1: three published monographs, 820 episodes, Train/Val/Test 526/153/141, hidden manager view with expert analysis
Section 4, Editable surface and Evolution loop: frozen LLM, skill document only, 80 held-in and 153 held-out episodes, acceptance gate rule, same LLM for consultation, diagnosis and proposal
Section 4, Learning from expert cases: feedback contents for TCM-SEA versus the outcome-only ablation
Section 5, Setup and Table 2: DeepSeek-Flash for agent and judge, 9 rounds, seed 42, sealed test set, mean of three runs
Section 5, The gate selects edits that transfer, with Tables 3 and 4: accepted edits at rounds 1 and 6, held-in/held-out/test trajectories, per-source results, edit text

29. Asclepius

Key
yuan2026asclepius
Name
Asclepius
Title
Asclepius: An Adaptive Harness for Long-Horizon Clinical Agents
Authors
Grace Chang Yuan; Xiaoman Zhang; Sung Eun Kim; Luyang Luo; Pranav Rajpurkar
Year
2026
Publication status
C: published in Findings of the Association for Computational Linguistics: EMNLP 2026, pp. 44611-44629 (ACL Anthology 2026.findings-emnlp.2094, DOI 10.18653/2026.findings-emnlp.2094); arXiv v1 comment also says EMNLP 2026 Findings.
Version
v1
Source url
https://arxiv.org/abs/2609.13543v1
DOI
10.18653/2026.findings-emnlp.2094
Update target
The natural-language operating manual (main-agent prompt with triage, treatment-completeness and disposition rules and a presentation-to-management table).
Fixed components
Claude Opus 4.6 backbone, MCP clinical tool interface, hand-curated skills library, subagent prompts extracted from the initial manual, proposer prompt, GPT-4.1 judge and CES simulator.
Feedback
Trace-level feedback from completed rollouts: aggregate, per-batch and per-patient LLM-judge scores on four dimensions (graded against gold answer keys), per-turn action records and LLM call audit logs.
Timing
Offline, between iterations; each of 10 proposer steps is followed by a full rollout over the same six search batches, and the selected manual is frozen at inference.
Retained artifact
The selected evolved manual M* (v7), plus a filesystem of all prior candidate manuals with their scores and traces that the proposer reads.
Update generator
A separate Claude Code proposer (Opus 4.6) with a fixed system prompt requiring edits grounded in failure patterns seen in at least 3 patients across batches, no patient- or batch-specific knowledge, and substantive edits to the current best manual.
Selection and release
After 10 iterations the manual with the best overall search-batch score (v7) is selected and reused unchanged on held-out batches.
Evaluation separate from updating
Four held-out batches (48 patients) from the same distribution, never shown to the proposer, with one rollout per configuration; baseline and full system re-graded by four further LLM judges.
Fixed-procedure control
Add-one component ablations and an evolved-versus-initial manual contrast holding skills and subagents fixed; no evolving-versus-fixed proposer test.
Evidence concerning improver change
Not reported; the proposer model, prompt and best-score selection rule are fixed across all iterations.
Clinical setting
Simulated six-hour emergency-department shift in CES (12 patients per batch, 4 beds, 6 nurses, one agent physician) with patient cards from de-identified ED records and GPT-4.1-generated physiology.
Evidence boundary
Simulator-only, LLM-graded, single rollout per batch and single backbone, no expert agreement on Asclepius traces; the evolved manual's marginal held-out gain over skills plus subagents is +0.02 (p=0.883).
Locators
Section 4.1: proposer reads candidate manuals, scores and traces, rewrites the manual, best search-batch manual (v7) selected after N=10 iterations; search/held-out split.
Table 2: held-out (batches 7-10, n=48) mixed-effects differences, critical actions +0.625 (p=0.024), timeliness and overall ns.
Table 5 / Appendix E.1: search vs held-out means per configuration; harness-only variant 4.08 search vs 3.82 held-out against baseline 3.81/3.78.
Table 9 / Appendix E.4: evolved manual vs initial manual with skills and subagents fixed, +0.12 (p=0.364) all batches, +0.02 (p=0.883) held-out.
Section 8 Limitations: CES is a simulator with LLM-generated physiology, single ED configuration and single backbone; performance is a proxy for clinical competence.
Published metadata; extraction uses arXiv v1

30. ReVision3D

Key
lee2026revision3d
Name
ReVision3D
Title
ReVision3D: Attribution-Guided Recursive Self-Improvement for 3D Medical Perception
Authors
Ho Hin Lee; Yuyin Zhou; Yannan Yu; Shi Gu; Yifan Wu
Year
2026
Publication status
P: arXiv preprint (v1 26 Sep 2026, v2 2 Oct 2026; comment '28 pages', no journal_ref or DOI); OpenReview lists it only as an ICLR 2027 Conference Submission and Crossref has no record (checked 2026-11-03).
Version
v2
Source url
https://arxiv.org/abs/2609.32984v2
DOI
Update target
Reader perception configuration (vision-encoder LoRA blocks, rank, input scale, epochs), the RL training recipe that produces reader weights, the 3D candidate-scoring inference program, and a view-selection program.
Fixed components
Frozen LLM designer, frozen pretrained Qwen3-VL-4B weights (only adapters train), mask-grounded verifier with 5 mm tolerance, system objective J, acceptance rule and four predefined intervention levels.
Feedback
Reference-mask verification of localized findings, a mask-grounded training reward, J over false-positive budgets on development patients, and replay-based loss attribution; the designer sees only aggregate trial statistics, attribution summaries and trial histories.
Timing
Offline revision rounds on development patients, each accepted change triggering renewed attribution; held-out test patients are read once after all choices are frozen.
Retained artifact
The accepted incumbent system (adapter weights, recipe, inference program) plus an evidence-backed playbook of successful and failed hypotheses and trial registries.
Update generator
A frozen language-model designer (Claude Opus 5 by default) proposes revisions within the level indicated by attribution, screened by replay where possible; perception changes require actual training.
Selection and release
Training revisions are accepted only if mean dev gain exceeds the incumbent's seed range and every matched seed improves; inference revisions need P(dJ>0) >= 0.9 under paired bootstrap.
Evaluation separate from updating
One-time held-out test on 31 liver and 70 kidney patients with patient-level bootstrap, though Appendix B reports that the 31 liver test patients were also used in earlier liver-study comparisons and only 40 of 70 kidney test patients were unused.
Fixed-procedure control
Equal-compute longer-training control for recipe revisions, frozen-encoder and hand-chosen perception comparisons, and a four-designer swap at the inference level; no evolving-versus-fixed improver test.
Evidence concerning improver change
The training recipe that generates reader weights is revised (one accepted learnability recipe, later recipe proposals rejected), while the designer, attribution, verifier, objective and acceptance rule stay fixed and the playbook's effect on later rounds is not tested.
Clinical setting
Retrospective lesion localization on public de-identified abdominal CT (LiTS liver, KiTS23 kidney); authors state it is not a clinical device and makes no diagnostic claim.
Evidence boundary
Benchmark-only, mask-verifiable lesion localization; routing rule formulated retrospectively and did not schedule the experiments; acquisition never passed the promotion gate; no clinician or patient-outcome evaluation.
Locators
Section 3, Motivation: system components (view selection, perception, weights, inference), frozen designer, and fixed verifier, masks, objective and acceptance rule.
Section 3, Recursive improvement: seed-range and bootstrap acceptance rules, playbook carried across rounds, routing rule formulated retrospectively.
Appendix E.1, Table A9: liver J and max recall after each accepted revision and who made each change.
Section 5.3, Table 2: development (selection) and single-run test J for perception and inference revisions.
Appendix G.1, Table A16: inference-level search repeated with four designers from three model families.

31. TEMPLAR

Key
hu2026templar
Name
TEMPLAR
Title
TEMPLAR: Agentic Induction and Evolution of Standardized Radiology Reporting Templates from Large-Scale Clinical Corpora
Authors
Xiaotian Hu; Mingxuan Liu; Zhonghan Wang; Xinfeng Zhang; Yiming Huang; Ziang Wang; Kasidit Anmahaepong; Yijin Li; Yifei Chen; Hongjia Yang; Zihan Li; Qiyuan Tian
Year
2026
Publication status
P: arXiv preprint (cs.MA, cs.CL) submitted 2026-10-04, single version v1; no arXiv comments or journal_ref, and Crossref/OpenAlex searches (2026-11-03) returned only the arXiv record.
Version
v1
Source url
https://arxiv.org/abs/2610.05247v1
DOI
Update target
A persistent hierarchical reporting template (slot keys, value types, candidate values, slot-level evidence notes, slot-specific extraction patterns) plus anatomical and diagnostic knowledge graphs built from the training corpus.
Fixed components
Backbone LLMs (Qwen3.6-Flash for induction/evolution, Gemini-3.5-Flash for structuring, reconstruction and diagnosis), text-embedding-3-small, prompts, clustering hyperparameters, the deterministic reducer and the evaluation judges are fixed; no parameters are trained.
Feedback
Training-corpus statistics, structural/unit consistency checks, retrieved PubMed/Europe PMC/Google evidence, and one audit of the Clinical Agent's structuring of sampled training findings for collapse to generic values.
Timing
Offline, before evaluation: a single construction pass per training split (one round of feedback-guided granularity calibration); the frozen template and graphs are then applied to held-out reports.
Retained artifact
Final template T* with reference notes and extraction patterns, the anatomical graph, the frozen diagnostic graph, and an indexed training-sentence corpus, all with source provenance and audit logs of retrieval queries and rationales.
Update generator
LLM Evolution Agent under fixed prompts proposes assignments, merges, fine revisions, evidence notes and slot-specific patterns P_s.
Selection and release
A deterministic reducer rejects revisions failing structural or unit checks, slots need more than 3 supporting atoms, vocabularies are supplemented only with training-sentence support, and calibration adds guidance without changing allowed values; no human or clinical release gate is described.
Evaluation separate from updating
Held-out test splits of four datasets scored for LLM-matched template coverage, reference-based reconstruction metrics, LLM-judged diagnostic fidelity (PDA, KFP, CA), and expert plus LLM template-quality ratings; no statistical tests or intervals are reported.
Fixed-procedure control
Cumulative component ablation from ASTAR (M1) to full system (M6), LLM-judged pairwise Revision/Note ablation, and cross-dataset transfer without re-induction; the feedback calibration step is not ablated separately and there is no evolving-versus-fixed updater test.
Evidence concerning improver change
not established: the template and graphs change once under constant prompts, thresholds and reducer rules, and the construction procedure itself is not revised.
Clinical setting
Retrospective English radiology report corpora in whole-body 18F-FDG PET/CT (PETWB-REP, ViMed from a Vietnamese tertiary hospital, machine-translated) and abdominal CT (MERLIN, RadGPT-generated BDMAP reports); no prospective or workflow deployment.
Evidence boundary
Offline corpus-to-template construction evaluated with largely LLM-based metrics on small test splits (n=49 to 408); expert raters are not described, and the authors state that metrics and expert assessments do not establish clinical safety or deployment readiness.
Locators
Section 3.2: template evolution as assembly, revision, external evidence grounding and one feedback-guided granularity calibration.
Appendix B.2.3: calibration audits sampled training structurings and adds slot-specific patterns without changing allowed values.
Section 4.1: four datasets, train/test splits, ASTAR and nine LLM baselines.
Section 4.4: cumulative ablation M1 (ASTAR) to M6 on PETCT, ViMed and PETCT-to-ViMed transfer.
Ethics Statement: research use only; metrics and expert assessments do not establish clinical safety or deployment readiness.

32. AgentClinic

Key
schmidgall2026agentclinic
Name
AgentClinic
Title
AgentClinic: a multimodal benchmark for tool-using clinical AI agents
Authors
Samuel Schmidgall; Rojin Ziaei; Carl Harris; Ji Woong Kim; Eduardo Pontes Reis; Jeffrey Jopling; Michael Moor
Year
2026
Publication status
J/C: npj Digital Medicine 9, Article 499 (2026), open access; received 19 April 2025, accepted 14 April 2026, published 27 April 2026, version of record 02 July 2026.
Version
Published version of record (nature.com HTML, fetched 2026-11-03); arXiv 2405.07960v5 (2025-05-25) used only for the tool table and notebook prompt, which the published HTML does not contain.
Source url
https://www.nature.com/articles/s41746-026-02674-7
DOI
10.1038/s41746-026-02674-7
Update target
A doctor-agent memory notebook of general diagnostic tips that persists across patient cases (in-context text, no weight updates); it is one of six Agent Toolbox tools in a benchmark paper.
Fixed components
Doctor LLM backbone weights, patient, measurement and moderator agents (GPT-4 patient/measurement), prompts, the N = 20 interaction budget and the case set are fixed.
Feedback
Per arXiv v5 Appendix M.2, after each case the notebook-update prompt shows the conversation, the correct diagnosis and the agent's own diagnosis; the published Methods describe ground-truth verification of the final answer.
Timing
Online and sequential: the notebook is rewritten after every case and used on the following AgentClinic-MedQA cases.
Retained artifact
A single notebook capped at 1000 characters that the agent overwrites each time and must re-integrate prior notes into (v5 Appendix M.2); the published Methods additionally describe a success-only vector store of reflective takeaways.
Update generator
The doctor LLM itself under a fixed notebook-writing meta-prompt, instructed to write general, non-patient-specific tips from successes and failures.
Selection and release
No acceptance test, validation gate or human review of notes is described; whatever the agent writes is truncated to the character limit and used on the next case.
Evaluation separate from updating
Not reported as a held-out split: the notebook gain is measured as AgentClinic-MedQA diagnostic accuracy over the same sequential stream in which notes are written (each case is seen before its own note update), with no seeds or intervals for the tool table.
Fixed-procedure control
Compared against a no-tool baseline per backbone and against five other tools (zero-shot, one-shot and reflection CoT, adaptive RAG book and web) on AgentClinic-MedQA, for six backbones.
Evidence concerning improver change
Not reported: the notebook-writing prompt and update procedure stay constant; only the notebook contents change.
Clinical setting
Simulated diagnostic dialogues built from USMLE-style MedQA vignettes (tool experiment), with LLM patient, measurement and moderator agents; no real patients or deployment.
Evidence boundary
Benchmark sub-experiment in simulation with an LLM moderator as judge; the notebook effect is backbone dependent (positive for four models, negative for GPT-3.5 and GPT-4o-mini) and is not tested on held-out cases or in clinical use.
Locators
Abstract: Llama-3 shows up to 92% relative improvement with a notebook that persists across cases.
Results, Comparing tools from the Agent Toolbox: six tools, Claude 3.5 56.1% with Notebook, Llama3-70b 41.1%, GPT-3.5 declines.
Methods, Tool descriptions: notebook persists across patients; experiential learning module stores takeaways from correct cases.
arXiv v5 Table 6: per-tool accuracy with change from no-tool baseline for six backbones.
arXiv v5 Appendix M.2: notebook prompt reveals correct diagnosis, 1000-character overwrite limit.

33. MedExpMem

Key
feng2026medexpmem
Name
MedExpMem
Title
MedExpMem: Adapting Experience Memory for Differential Diagnosis
Authors
Qianhan Feng; Zhongzhen Huang; Yakun Zhu; Yannian Gu; Winnie Chiu Wing Chu; Xiaofan Zhang; Qi Dou
Year
2026
Publication status
J/C: MICCAI 2026 book chapter in LNCS, pp. 430-440, Springer, published 2026-09-25 (Crossref confirmed); extraction uses arXiv 2605.22872v1, which its comment labels 'MICCAI 2026 Early Accept. Submission Version'.
Version
arXiv 2605.22872v1 (2026-05-20, only version listed by the arXiv API on 2026-11-03); the LNCS chapter text was not examined.
Source url
https://arxiv.org/html/2605.22872v1
DOI
10.1007/978-3-032-38098-2_41
Update target
An external, per-model experience memory of structured pairwise differential-diagnosis notes (differentials, confusions, discriminators, decision_rule, error_analysis), indexed by 11 departments and 118 organ/region paths.
Fixed components
VLM parameters are not updated (Qwen3-VL 2B/8B/30B, InternVL3.5 8B/30B, Lingshu 7B/32B); the S-PubMedBert-MS-MARCO retrieval encoder, threshold tau=0.9, top-K=10 and the prompts are fixed.
Feedback
On construction cases the agent's diagnosis is compared with the Eurorad ground-truth diagnosis, and the expert discussion (imaging findings and diagnostic reasoning) is supplied when extracting a note from each error.
Timing
Offline construction in two passes over 6,228 pre-2025 Eurorad cases (Phase I zero-shot, Phase II re-diagnosis with memory), then inference on 227 cases from 2025; updating memory during the test set is not reported.
Retained artifact
The pairwise differential notes, merged per differential pair, persist as a separate memory library built independently for each VLM.
Update generator
The same diagnostic agent (VLM) extracts each note from its own misdiagnosis, instructed to identify confusion points, discriminators, bidirectional decision rules and its reasoning failure.
Selection and release
Notes are written only for errors (d != d*); Phase II adds notes for stochastic errors and note deficiencies, refines notes that caused correct-to-incorrect transitions and merges insights into existing pairs; no separate validation gate or human review of notes is reported, and at test time retrieved notes are filtered by threshold, top-K and an agent relevance score.
Evaluation separate from updating
Accuracy on a temporally held-out set of 227 Eurorad 2025 cases versus zero-shot baseline, with retrieval recall/precision and beneficial/harmful case rates; the answer-matching method, number of runs, confidence intervals and significance tests are not reported.
Fixed-procedure control
Zero-shot baseline for all seven models, a PubMed-RAG (top-3 articles) comparison and construction/retrieval ablations, all on Qwen3-VL-8B only for the comparison and ablations.
Evidence concerning improver change
Not reported: the extraction, Phase II refinement and retrieval procedures are fixed by design; the one-round versus two-round construction ablation compares preset procedure variants rather than a procedure that revises itself.
Clinical setting
Retrospective evaluation on Eurorad radiology teaching-file cases (CT, MRI, X-ray, ultrasound with clinical history); no prospective, reader-study or deployment evaluation is reported.
Evidence boundary
Gains are single-benchmark accuracy changes of +1.7 to +7.0 points without reported variance or tests, with harmful (correct to wrong) rates of 1.3% to 5.8% and ablations on one model.
Locators
Section 2.2, Table 1: note schema, and 11 departments / 118 organ-region taxonomy.
Section 2.3: Phase I error-driven extraction with expert discussion; Phase II re-diagnosis, refinement and merging.
Section 3.1: 6,228 pre-2025 construction cases, 227 test cases from 2025, tau=0.9, K=10, per-agent memory.
Table 2: baseline vs w/Exp accuracy, recall, precision, beneficial and harmful rates for seven VLMs.
Table 3: Qwen3-VL-8B PubMed-RAG comparison and construction/retrieval ablations.

34. EHRAgent

Key
shi2024ehragent
Name
EHRAgent
Title
EHRAgent: Code Empowers Large Language Models for Few-shot Complex Tabular Reasoning on Electronic Health Records
Authors
Wenqi Shi; Ran Xu; Yuchen Zhuang; Yue Yu; Jieyu Zhang; Hang Wu; Yuanda Zhu; Joyce Ho; Carl Yang; May D. Wang
Year
2024
Publication status
C (EMNLP 2024 main conference, pages 22315-22339)
Version
arXiv v3 (2024-10-04) read in full; numbers checked against the EMNLP 2024 Anthology PDF
Source url
https://arxiv.org/html/2401.07128v3
DOI
10.18653/v1/2024.emnlp-main.1245
Update target
A long-term memory of past successful question and code solutions, retrieved as the K=4 few-shot demonstrations (§3.2, Eq. 1).
Fixed components
GPT-4 (gpt-4-0613, temperature 0) backbone, tool definitions, EHR metadata, prompts and Levenshtein-similarity retrieval stay fixed (§3.2, §4.1).
Feedback
Ground-truth success of each answer decides memory admission; execution error messages drive within-case debugging only (§3.2, App. F.2).
Timing
Online across the evaluated question stream; no separate training split for memory is reported (§3.2, App. F.2).
Retained artifact
Successful past cases (question plus code solution) that replace the original hand-written demonstrations (§3.2, App. H.2 Fig. 9).
Update generator
The agent's own successful solving runs; no separate generator component.
Selection and release
A case enters memory only when it succeeds against the ground-truth answer; a variant admitting all completed cases is also tested (Table 8).
Evaluation separate from updating
Not reported: success rate is measured on the same question stream from which memory accumulates (§4.1, App. F.2).
Fixed-procedure control
Ablation without long-term memory under the same backbone and prompts on MIMIC-III and eICU (Table 2, Table 7); shuffled-order runs reported as almost no influence (App. F.2).
Evidence concerning improver change
None; the storage rule and retrieval procedure are fixed.
Clinical setting
Multi-hop question answering over structured EHR tables of MIMIC-III, eICU and TREQS (§4.1).
Evidence boundary
Offline benchmark with ground-truth labels used at test time for memory admission; no clinician-facing or prospective evaluation (§4.1, Limitations).
Locators
§3.2 Demonstration Optimization through Long-Term Memory
Table 2, ablation on MIMIC-III
Table 8, LTM with success vs completion vs none
Table 7, eICU ablation
Published version, Algorithm 1 and §3.2

35. ReflecTool

Key
liao2025reflectool
Name
ReflecTool
Title
ReflecTool: Towards Reflection-Aware Tool-Augmented Clinical Agents
Authors
Yusheng Liao; Shuyang Jiang; Yanfeng Wang; Yu Wang
Year
2025
Publication status
C (ACL 2025 main, Volume 1 Long Papers, pages 13507-13531)
Version
arXiv v3 (2025-06-16) read in full; key table values match the ACL 2025 Anthology PDF
Source url
https://arxiv.org/html/2410.17657v3
DOI
10.18653/v1/2025.acl-long.663
Update target
A long-term memory of successful reflective trajectories and a per-tool store of tool-wise experience (§3.2, Eq. 3-4).
Fixed components
Qwen2-7B/72B backbone (72B with 4-bit GPTQ), 15-type clinical toolbox, prompts, BM25 retrieval and verifier procedures stay fixed (§2.2, §3.3, §4.2).
Feedback
Ground-truth answers on the optimization subset; the agent reflects on its first trajectory against the answer and retries (§3.2).
Timing
Offline optimization stage before inference; memory and experience are then frozen and only read at inference (§3.2-3.3).
Retained artifact
Successful refined trajectories as memory items, plus merged action-wise suggestions for each tool and for Plan, Think and Finish (Algorithm 1).
Update generator
The same LLM agent writes reflections and tool-wise suggestions by comparing failed and successful trajectories (Eq. 4, App. D.2 Fig. 8-9).
Selection and release
A trajectory is stored only when its refined answer equals the ground truth; tool experience is merged without further validation (Eq. 3).
Evaluation separate from updating
Test samples of CAB distinct from the optimization subset; ablations on 200 random samples per dimension; optimization subset size not reported (§3.2, §4.3).
Fixed-procedure control
Ablations removing reflective memory or tool-wise reflection (Table 3); baselines given the same long-term memory (Table 4); performance versus optimization steps (Fig. 3).
Evidence concerning improver change
None; the reflection and storage procedure is fixed.
Clinical setting
ClinicalAgent Bench: 18 public tasks over knowledge, multimodal, numerical, data understanding and trustworthiness dimensions (§2.1, App. B).
Evidence boundary
Offline benchmark; authors note correct answers do not guarantee correct trajectories (Limitations); no clinician-facing or prospective evaluation.
Locators
§3.2 Optimization Stage
§3.3 Inference Stage
Table 3, ablation of memory and tool-wise reflection
Table 4, long-term memory given to baselines
Figure 3, effect of optimization step

36. MDTeamGPT

Key
chen2026mdteamgpt
Name
MDTeamGPT
Title
MDTeamGPT: Mitigating Context Collapse and Enabling Self-Evolution in Medical Multi-Agent Reasoning
Authors
Kai Chen; Xinfeng Li; Tianpei Yang; Hewei Wang; Guang Yang; Jing Huo; Yang Gao
Year
2026
Publication status
C (Findings of ACL 2026, pages 28578-28606); retitled from arXiv 2503.13856
Version
Findings of ACL 2026 version read; arXiv v1 (2025-03-18) also read and differs (600 vs 900 accumulation rounds, text-only, average 87.0)
Source url
https://aclanthology.org/2026.findings-acl.1427.pdf
DOI
10.18653/v1/2026.findings-acl.1427
Update target
Two knowledge bases: CorrectKB for correct consultations and ChainKB for reflective analyses of incorrect ones (§3.3).
Fixed components
gpt-4-turbo backbone in the main text (gpt-5 for multimodal), agent roles, Lead Physician and Residual Context protocol, and text-embedding-3-small retrieval (§3, §4.2, App. C.3).
Feedback
The final consultation answer is validated against the ground truth, which routes the record to CorrectKB or ChainKB (§3.3).
Timing
Accumulated over 900 consultation rounds on training sets before testing; test-time KBs contain only training-set experience (§4.1-4.2).
Retained artifact
Structured summaries: question, answer and final-round statements (CorrectKB); hypothesis, analysis, conclusion and error reflection (ChainKB) (§3.3).
Update generator
A Chain-of-Thought Reviewer agent summarizes each finished consultation into the KB format (§3.3).
Selection and release
Routing by ground-truth correctness only; no further validation, pruning or release gate is reported (§3.3).
Evaluation separate from updating
Official MedQA and PubMedQA test sets, five runs with SD; cross-dataset KB transfer (Table 3); VQA-RAD and SLAKE official test sets (App. C.2).
Fixed-procedure control
Ablation adding CorrectKB, ChainKB or both with the discussion protocol fixed (Table 2); accuracy versus accumulated cases for eight backbones (Fig. 4).
Evidence concerning improver change
None; the reviewer and routing rule are fixed.
Clinical setting
Simulated multi-disciplinary team consultation on exam and research-abstract QA, radiology VQA and self-collected complex cases (§4.1, App. C).
Evidence boundary
Offline benchmarks with labels used for KB admission; no deployment, authors plan shadow-system validation (§6 Limitations); human evaluation on 20 samples (App. C.8).
Locators
§3.3 Summary and Output Stage, KB construction
Table 2 ablation and §4.5 Self-Evolving Capability
Fig. 4 and Table 3 cross-dataset generalization
§6 Limitations
App. C.2-C.3, Table 6 multimodal results

37. Learning to Be A Doctor (medical agent workflow search)

Key
zhuang2025doctor
Name
Learning to Be A Doctor (medical agent workflow search)
Title
Learning to Be A Doctor: Searching for Effective Medical Agent Architectures
Authors
Yangyang Zhuang; Wenjia Jiang; Jiayu Zhang; Ze Yang; Joey Tianyi Zhou; Chi Zhang
Year
2025
Publication status
C (ACM MM 2025, Proceedings of the 33rd ACM International Conference on Multimedia)
Version
arXiv v2 (2025-08-15, comment: Accepted at ACM MM 2025); published version not read
Source url
https://arxiv.org/html/2504.11301v2
DOI
10.1145/3746027.3755559
Update target
The agent workflow graph (nodes, edges, conditional, loop and parallel structures) and node system and human prompts (§3.1-3.3).
Fixed components
Backbone LLM (GPT-4o, GPT-4o-mini or Claude 3.5 Sonnet, temperature 1.0), search-space operations, CLIP image-search tool and reference database, evolution prompts (§3.2-3.3, §4.1).
Feedback
Diagnostic errors against ground-truth labels, which the LLM classifies as image-understanding or diagnostic errors (§3.3).
Timing
Offline batched iterations over the training split, validated after each batch until validation accuracy converges; final workflow frozen for testing (§4.1).
Retained artifact
The final evolved workflow, e.g. 11 nodes and 5 branches at test time (Table 2).
Update generator
The LLM analyses failures and emits structural and prompt suggestions, reformulated into structured modifications via templates (§3.3).
Selection and release
Infeasible suggestions are filtered, structural edits pass graph-validity checks, and each updated workflow is checked on the 50-image validation set (§3.3, §4.1).
Evaluation separate from updating
Held-out test splits: 224 Skin Concepts images (before excluding 100 images for moderation or broken links) and 120 Skin Conditions images (§4.1).
Fixed-procedure control
Ablations disabling individual operations (Table 3) and operation levels (Table 4); static IO, CoT and Round Table baselines (Table 1); single seed 42.
Evidence concerning improver change
None; the analysis, suggestion and validation procedure is fixed.
Clinical setting
Dermatology image diagnosis on Skin Concepts (Fitzpatrick17k extension) and the synthetically augmented Skin Conditions dataset (§4.1).
Evidence boundary
Small 50/50 training and validation splits, one seed, one synthetic dataset; offline benchmarks only, no clinician or prospective evaluation (§4.1).
Locators
§3.3 Workflow Evolution
§4.1 datasets, splits, implementation details
Table 1, top-k accuracy vs baselines
Table 3, operation ablation
Table 4, operation-level ablation

38. MACD

Key
li2025macd
Name
MACD
Title
MACD: Multi-Agent Clinical Diagnosis with Self-Learned Knowledge for LLM
Authors
Wenliang Li; Rui Yan; Xu Zhang; Li Chen; Hongji Zhu; Jing Zhao; Junjun Li; Mengru Li; Wei Cao; Zihang Jiang; Wei Wei; Kun Zhang; Shaohua Kevin Zhou
Year
2025
Publication status
P: arXiv preprint (cs.AI), v1 2025-09-24, v5 2026-08-24; no comments or journal_ref; Semantic Scholar lists only arXiv (checked 2026-11-04). v5 declares an npj Digital Medicine editor conflict, suggesting submission there; no acceptance found.
Version
v5
Source url
https://arxiv.org/abs/2509.20067v5
DOI
Update target
Disease-specific Self-Learned Knowledge (general and rare diagnostic concepts) injected into the diagnostician prompt.
Fixed components
Base LLMs (Llama-3.1-8B/70B, DeepSeek-R1-Distill-Llama-70B), prompt templates, summarizer and refiner procedures (BioBERT MMR de-duplication to top-7, concept ablation), decoding settings, matching rules.
Feedback
Ground-truth diagnosis correctness on learning cases; refiner uses change in learning-set diagnostic accuracy when each concept is ablated.
Timing
Offline accumulation over the learning set (K_t to K_{t+1} iterations; number of iterations not reported), then frozen knowledge applied to the test set.
Retained artifact
Per-disease, per-model Self-Learned Knowledge text (Supplementary Tables 10-12), also transferred to other LLMs.
Update generator
Diagnostician diagnoses learning cases with current knowledge; Summarizer agent abstracts insights from correct diagnoses; Refiner agent merges, de-duplicates and removes concepts whose ablation does not lower accuracy.
Selection and release
Learning cases sampled from cases correctly diagnosed by at least one base agent (90 per disease per agent; all 23 for pericarditis); concepts kept by the refiner's redundancy and causal-ablation filters; no clinician release gate.
Evaluation separate from updating
Held-out MIMIC-MACD test cases never used for learning; comparisons with Mayo Clinic and guideline-derived knowledge, CoT, few-shot and LoRA fine-tuning; cross-model transfer; clinician concept-relevance ratings and a 140-case physician reader study.
Fixed-procedure control
Summarizer-only versus summarizer plus refiner ablation and knowledge-composition ablations under the same pipeline; no evolving-versus-fixed improver comparison.
Evidence concerning improver change
Not established: summarizer, refiner and selection rules stay fixed; only the knowledge text changes.
Clinical setting
Primary diagnosis of seven acute abdominal and chest diseases from MIMIC-IV emergency/inpatient text (history, examination, laboratory results, radiology reports).
Evidence boundary
Single-institution retrospective text vignettes; test set is the non-random residue after success-based learning-case selection; physician comparison on text-only vignettes; number of repeated runs not reported.
Locators
Methods, Dataset construction: 4,390 cases, success-driven learning-set selection
Methods, Multi-agent clinical diagnosis framework: Eqs. 1-4, summarizer, refiner, causal concept ablation
Results: SLK vs authoritative knowledge, CoT, few-shot, LoRA, SOTA models (Fig. 2)
Table 2: ablations including Sum. Only vs Sum. + Ref.
Results: MACD-human workflow and physician comparison (Fig. 4)

39. MedDCR

Key
zheng2026meddcr
Name
MedDCR
Title
MedDCR: Learning to Design Agentic Workflows for Medical Coding
Authors
Jiyang Zheng; Islam Nassar; Thanh Vu; Xu Zhong; Yang Lin; Tongliang Liu; Long Duong; Yuan-Fang Li
Year
2026
Publication status
J/C: Findings of the Association for Computational Linguistics: ACL 2026, pp. 12878–12893 (Crossref).
Version
arXiv v1 (17 Nov 2025) read in full; ACL Anthology PDF checked and Tables 1–4 and Algorithm 1 match.
Source url
https://arxiv.org/html/2511.13361v1
DOI
10.18653/v1/2026.findings-acl.627
Update target
Executable medical-coding workflow programs (tool and strategy composition) and the archive of scored past designs.
Fixed components
GPT-4o/GPT-5 backbones, Designer/Coder/Reflector meta-prompts, ICD-10 tool library, objective G and the top-5 plus recent-3 exemplar rule.
Feedback
Validation micro-F1-based score with guideline-compliance and cost terms, plus Reflector textual critique of intermediate LLM outputs.
Timing
Offline search for 100 iterations before use; no parameter training.
Retained artifact
Archive tuples of plan, compiled workflow, score and reflection; the best-scoring workflow is the released output.
Update generator
LLM Designer proposes plans conditioned on elite and recent archive entries; Coder compiles them with a self-fixing retry loop.
Selection and release
Every candidate is appended to the archive; the final workflow is the argmax validation score over the archive (Alg. 1, line 18).
Evaluation separate from updating
Not reported; optimisation uses D_val and the paper does not describe a separate test split for Table 1.
Fixed-procedure control
Ablations removing top-k, recent-n, score feedback or text feedback (Table 4); no changing-improver comparison.
Evidence concerning improver change
None reported; Designer, Reflector and archive-selection rule stay fixed while workflows change (archived workflows do serve as Designer exemplars).
Clinical setting
ICD-10 coding benchmarks: MDACE (MIMIC-III charts incl. discharge summaries and radiology reports) and synthetic ACI-BENCH notes.
Evidence boundary
Benchmark workflow search; overlap between selection and reporting data cannot be excluded from the text, and no clinical deployment.
Locators
§3.2 framework overview and Algorithm 1
§3.4 memory archive and plug-and-play
§4.1 datasets and 100-iteration implementation
§4.3 ablations, Table 4
§6 limitations

40. KEA (on PG-Bench)

Key
lu2025kea
Name
KEA (on PG-Bench)
Title
Advancing Chinese Conversation-based Patient Guidance with a Benchmark and Knowledge-Evolvable Assistant
Authors
Wenpeng Lu; Kangjun Liu; Jianlei Wang; Xueping Peng; Tao Shen; Fa Zhu; Weiyu Zhang; Jiabing Zhu; Tao Xin; Athanasios V. Vasilakos
Year
2025
Publication status
J/C: IEEE Journal of Biomedical and Health Informatics, 2025, pp. 1–12 (early access per Crossref; PubMed 41336169).
Version
Author manuscript in IEEE template from the UTS OPUS repository (open-access copy listed by Semantic Scholar).
Source url
https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf
DOI
10.1109/JBHI.2025.3639805
Update target
Experience bank of correct case–answer pairs and reflection bank of corrected cases with self-summarised error analyses.
Fixed components
Backbone LLM weights (GPT-4o, DeepSeek-V3, BianCang-7B-Instruct), BGE retriever, static medical knowledge base and prompt templates.
Feedback
Gold department labels on training dialogues; incorrect answers trigger LLM reflection and re-prediction.
Timing
Offline evolution rounds (1–3, extended to 7 in Table IX) over the 70% training split before test inference.
Retained artifact
Prompt–answer pairs in the experience bank; prompt, both answers and reflection stored only if the revised answer matches the label.
Update generator
The same frozen task LLM writes reflections on its own errors.
Selection and release
Label correctness gates storage; no other release criterion reported.
Evaluation separate from updating
Held-out 30% test split of PG-Bench; separate high- and low-frequency test sets (Tables V–VI); human evaluation assesses dataset realism only.
Fixed-procedure control
Component ablations (Table VIII) and round-count comparisons; no changing-improver comparison.
Evidence concerning improver change
None; reflection prompt and storage rule fixed across rounds.
Clinical setting
LLM-generated, expert-filtered Chinese patient-guidance (triage) dialogues across 98 departments.
Evidence boundary
Synthetic-dialogue benchmark; best GPT-4o accuracy is 37.91% and gains plateau after round 5; no deployment study.
Locators
§III-B PG-Bench construction and 7:3 split (pp. 3–4)
§IV-A evolving knowledge pool
§IV-B dynamic evolution rule
Tables II–IV round-wise results
§VII-E/F ablation and extended rounds, Tables VIII–IX (pp. 9–10)

41. SpAgents

Key
ji2026spagents
Name
SpAgents
Title
Early diagnosis of axial spondyloarthritis in primary care using multi-agent systems
Authors
Xiaojian Ji; Zhuofeng Li; Lulu Zeng; Lidong Hu; Yanyan Wang; Kui Zhang; Lianjie Shi; Meng Wei; Lifeng Chen; Lin Guo; Jing Dong; An’an Wang; Lei Sun; Yimin Song; Huatao Wang; Jingming Wang; Ying Lei; Wenqian Yue; Zheng Zhao; Jian Zhu; Feng Huang; Jing Zhang; Tao Li; Kunpeng Li
Year
2026
Publication status
J/C: npj Digital Medicine 9, 185 (2026).
Version
Published HTML article and Supplementary Information PDF (algorithms, Fig. S3).
Source url
https://www.nature.com/articles/s41746-026-02372-4
DOI
10.1038/s41746-026-02372-4
Update target
Long-term case memory of ClinicalBERT-encoded patient data with physician diagnostic labels.
Fixed components
DeepSeek-Chat Planner/Data/Tool agents, DeepSeek-Reasoner DoctorAgent, SPARCC MRI tool, prompts and top-3 FAISS retrieval.
Feedback
Physician diagnostic labels submitted through a 'learning' task.
Timing
Memory updated immediately on feedback; when updates occurred in the reported experiments is not reported.
Retained artifact
Patient case text plus diagnostic label vectors, with create/read/update/delete management.
Update generator
None; labelled cases are stored directly.
Selection and release
Not reported beyond storing physician-labelled cases.
Evaluation separate from updating
Single-centre validation set (n=186) and external test set (n=51) from other hospitals; with/without-memory comparison.
Fixed-procedure control
With versus without memory (Table 3); no changing-improver comparison.
Evidence concerning improver change
None.
Clinical setting
Retrospective axSpA diagnosis from records, laboratory data and sacroiliac MRI; reader study with seven physicians.
Evidence boundary
Memory ablation gain shown, but memory contents during testing unreported; external hospitals counted as five (Abstract) and six (Methods).
Locators
Results: Impact of long-term memory, Table 3
Methods: data preparation and cohorts
Methods: Design of DoctorAgent (diagnosis and learning tasks)
Methods: Evaluation of the long-term memory
Supplementary Note 1.4 and Fig. S3 (memory retrieval/update)

42. FHIR-AgentEval memory configurations

Key
mokssit2026fhiragenteval
Name
FHIR-AgentEval memory configurations
Title
FHIR-AgentEval: A Modular Sandbox for Benchmarking Clinical LLM Agents with an Evaluation of Memory-Augmented Configurations
Authors
Youssef Mokssit; Kamalakkannan Ravi; Mengshu Nie; Junyoung Kim; Cong Liu
Year
2026
Publication status
J/C: npj Digital Medicine, accepted article preview published 10 Oct 2026; full text read from the Research Square preprint.
Version
Research Square preprint v1 (DOI 10.21203/rs.3.rs-8746188/v1); published preview page shows abstract only.
Source url
https://www.researchsquare.com/article/rs-8746188/v1
DOI
10.1038/s41746-026-03348-0
Update target
Dual FAISS memory of micro reflections (FHIR resource–operation tips) and macro reflections (task strategy).
Fixed components
GPT-4.1-mini ReAct evaluation agent, GPT-4.1 training agent, evaluator and reflector prompts/models, MCP tools, deterministic validators.
Feedback
Deterministic validator outcomes, execution logs and a GPT-4.1 evaluator critique.
Timing
Separate offline training phase (up to 2 variations and 2 trials per task, early stop on success); memory then frozen.
Retained artifact
Reflection lessons indexed for semantic and exact-key retrieval.
Update generator
GPT-o4-mini reflector, optionally grounded in the FHIR R4 specification server.
Selection and release
Not reported; generated reflections appear to be indexed without a gate.
Evaluation separate from updating
Unseen task variations (Experiments 2–3) and 22 held-out tasks (Experiment 4), three runs each, deterministic validation.
Fixed-procedure control
Five configurations varying memory and specification access; no changing-improver comparison.
Evidence concerning improver change
None.
Clinical setting
Resettable HAPI FHIR sandbox with synthetic appointment-management and genetic-testing workflow tasks.
Evidence boundary
Sandbox administrative workflows; held-out transfer gain is smaller than within-task gains; no patient-care evidence.
Locators
Methodology: Reflexion long-term memory MCP server
Methodology: Reflexion training process (micro and macro reflections)
Methodology: four experimental settings, Experiment 4 held-out tasks
Results: Figure 5a success rates
Discussion: memory effects and specification grounding

43. ClinicalReTrial

Key
xing2026clinicalretrial
Name
ClinicalReTrial
Title
ClinicalReTrial: Clinical Trial Redesign with Self-Evolving Agents
Authors
Sixue Xing; Kerui Wu; Xuanye Xia; Haoyu He; Meng Jiang; Jintai Chen; Tianfan Fu
Year
2026
Publication status
J/C: Findings of the Association for Computational Linguistics: EMNLP 2026, pp. 16296–16330 (Crossref).
Version
arXiv v3 (13 Sep 2026) read in full; ACL Anthology PDF checked for Table 3 and Appendix C.2.
Source url
https://arxiv.org/html/2601.00290v3
DOI
10.18653/2026.findings-emnlp.772
Update target
Global memory of edit-category success rates and mean rewards, plus within-trial local memory and redesign pool.
Fixed components
GPT-4o-mini agents, LightGBM outcome predictor used as reward, RAG safety validator, beam search and prompts.
Feedback
Change in predicted success probability, attributed to individual augmentations as marginal rewards.
Timing
Within-trial over 5 iterations; global memory updated after each trial converges.
Retained artifact
Edit-category priors (success rate, mean reward) used for confidence calibration and generation-budget allocation.
Update generator
LLM synthesis distils each trial's reward distribution into the global pool.
Selection and release
Only safety-validated augmentations receive rewards; no further release gate reported.
Evaluation separate from updating
60 failed trials from the held-out TrialBench test split; HINT re-scoring never used in optimisation (App. C.2); 8 retrospective case studies.
Fixed-procedure control
Paired ablations without memory, pool or analysis (Table 3); no changing-improver comparison.
Evidence concerning improver change
None.
Clinical setting
In-silico redesign of failed Phase I–IV clinical-trial protocols.
Evidence boundary
Gains are predicted probabilities; order of global-memory accumulation and separation of global from local memory are not reported.
Locators
§3.3.3 cross-trial global memory
Appendix A Algorithm 1
§4.3 Tables 2–3 and ablations
Appendix C.2 cross-predictor HINT evaluation
Limitations

44. R4Agent

Key
sayeedi2026r4
Name
R4Agent
Title
Route, Retrieve, Reflect, Repair: Self-Improving Agentic Framework for Visual Detection and Linguistic Reasoning in Medical Imaging
Authors
Md. Faiyaz Abdullah Sayeedi; Rashedur Rahman; Siam Tahsin Bhuiyan; Sefatul Wasi; Ashraful Islam; Saadia Binte Alam; AKM Mahbubur Rahman
Year
2026
Publication status
P: arXiv preprint with no comment or journal_ref; no Crossref publication found.
Version
arXiv v2 (8 Mar 2026).
Source url
https://arxiv.org/html/2601.08192v2
DOI
Update target
Exemplar memory of task type, specialization, cue, final report and tags from processed cases.
Fixed components
Frozen VLM backbones (Gemini-2.5-Flash, MedGemma-4B, Qwen2.5-VL-7B), router rules, reflector/repairer prompts, lexical-overlap retrieval.
Feedback
The system's own Reflector critique and pass@k draft scoring; no ground-truth feedback described for memory writes.
Timing
Memory updated after each processed case during inference.
Retained artifact
Self-generated final reports with cues and tags, reused as few-shot exemplars.
Update generator
Cue and tag functions applied to the final repaired output.
Selection and release
Outputs judged not to be a system error are stored; the criterion is not further specified.
Evaluation separate from updating
IU Chest X-ray report generation and VinBigData localization; GPT-4o-mini reference-based judge, overlap metrics and mAP50; test sizes and case order not reported.
Fixed-procedure control
Not reported; only a pass@k ablation (Fig. 2).
Evidence concerning improver change
None.
Clinical setting
Public retrospective chest X-ray datasets.
Evidence boundary
Reported gains bundle routing, reflection, repair and memory; stored exemplars are unverified self-generated outputs.
Locators
§3.2 Exemplar Memory
§3.3 Self-Improvement via Exemplar Curation
§4.1 Datasets
§5 Results, Table 1 and Fig. 2 pass@k ablation

45. AgentEHR / RetroSum

Key
liao2026agentehr
Name
AgentEHR / RetroSum
Title
AgentEHR: Advancing Autonomous Clinical Decision-Making via Retrospective Summarization
Authors
Yusheng Liao; Chuan Xuan; Yutong Cai; Lina Yang; Zhe Chen; Yanfeng Wang; Yu Wang
Year
2026
Publication status
C: EMNLP 2026 main conference (ACL Anthology 2026.emnlp-main.1529); arXiv v1 examined, key numbers (0.2880, 0.2117, 100 training examples, top-1) also present in the published PDF.
Version
arXiv v1 (2026-01-20), cross-checked against EMNLP 2026 PDF
Source url
https://arxiv.org/html/2601.13918v1
DOI
10.18653/2026.emnlp-main.1529
Update target
External Experience Memory Bank of actor heuristics and summarizer salience guidelines keyed by EHR embeddings.
Fixed components
LLM backbone weights (no parameter updates), RetroSum actor and summarizer prompts, MCP toolbox of 19+ tools, bge-m3 retriever with top-1 retrieval.
Feedback
Ground-truth answer sets of training instances compared with the agent's trajectory, summary and predictions by a reflection prompt.
Timing
Offline before evaluation on 100 MIMIC-IV-Common training examples per task; the same accumulated memory is then reused on all three test subsets.
Retained artifact
Triplets (EHR embedding, actor experience, summarizer experience), one per training instance.
Update generator
Reflection module that shares the agent's frozen LLM, prompted to extract procedural tool-use heuristics and information-salience guidelines.
Selection and release
No filtering, validation or pruning of generated experiences is reported.
Evaluation separate from updating
Held-out MIMIC-IV-Common cohorts plus MIMIC-IV-Rare (label shift) and MIMIC-III (schema shift) with the same memory; evolved versus non-evolved ablations; ReflecTool and ReasoningBank comparators across five backbones.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
None; only memory contents change.
Clinical setting
Retrospective hospital and ICU EHR databases (MIMIC-IV, MIMIC-III) queried through tools for six prediction tasks such as diagnoses and prescriptions.
Evidence boundary
Single-center MIMIC data; memory built once offline; disjointness of the 100 experience examples from test cohorts not stated explicitly; significance tests compare against ReSum, not against non-evolved RetroSum.
Locators
Sec. 3.3 Evolving Strategy: experience generation, memory construction, evolved inference
App. D.2: 100 training examples per task, same memory for all subsets, top-1 bge-m3 retrieval
Table 2: MIMIC-IV-Common results, evolved vs non-evolved, five backbones
Table 3: OOD subsets MIMIC-IV-Rare and MIMIC-III
Table 4: ablation of retrospect and evolved components

46. UltrasoundAgents

Key
zhu2026ultrasoundagents
Name
UltrasoundAgents
Title
UltrasoundAgents: Hierarchical Multi-Agent Evidence-Chain Reasoning for Breast Ultrasound Diagnosis
Authors
Yali Zhu; Kang Zhou; Dingbang Wu; Gaofeng Meng
Year
2026
Publication status
P: arXiv preprint; no journal_ref or acceptance comment; no published version found in Crossref or OpenAlex.
Version
arXiv v1 (2026-03-11)
Source url
https://arxiv.org/html/2603.10852v1
DOI
Update target
Weights of the attribute sub-agent and the main diagnostic agent (Qwen2.5-VL-3B).
Fixed components
Backbone architecture, crop-and-zoom tool, BUS-CoT attribute taxonomy, reward definitions, GRPO and SFT hyperparameters.
Feedback
Ground-truth attributes, boxes, malignancy and BI-RADS labels: attribute accuracy plus format reward (Stage 1), diagnosis correctness with oracle attributes (Stage 2), refined trajectories for SFT (Stage 3).
Timing
Three sequential offline training stages before testing; agents fixed at inference.
Retained artifact
Trained sub-agent weights, distilled main-agent weights, and the refined trajectory set used for SFT.
Update generator
Stage-2 policy rollouts; the Stage-2 model rewrites rationales of misdiagnosed samples conditioned on ground truth; predicted boxes replaced by ground-truth boxes.
Selection and release
All sampled trajectories are refined and used; no further selection or validation gate is reported.
Evaluation separate from updating
Held-out BUSBRA, BUSI and BUDIAT splits (BUS-CoT protocol) and external BrEaST OOD set; ablations without oracle training and without self-distillation; GT-box and GT-attribute upper bounds.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
The trained Stage-2 policy also rewrites its successor's SFT data; this role is not isolated from box replacement in the ablation.
Clinical setting
Breast ultrasound image diagnosis (malignancy and BI-RADS) on public datasets.
Evidence boundary
Single 3B backbone and single runs without variance; BI-RADS agreement low (overall kappa 0.224, OOD 0.037); no learning after deployment.
Locators
Sec. 2.2: three-stage decoupled training and corrective self-distillation
Sec. 3.1: datasets, splits, implementation and baselines
Table 1: main diagnostic performance, in-domain and BrEaST OOD
Table 2: ablation and upper-bound analysis

47. TheraAgent

Key
chen2026theraagent
Name
TheraAgent
Title
TheraAgent: Multi-Agent Framework with Self-Evolving Memory and Evidence-Calibrated Reasoning for PET Theranostics
Authors
Zhihao Chen; Jiahui Wang; Yizhou Chen; Xiaozhong Ji; Xiaobin Hu; Jimin Hong; Wolfram Andreas Bosbach; Axel Rominger; Ali Afshar-Oromieh; Hongming Shan; Kuangyu Shi
Year
2026
Publication status
P: arXiv preprint; code 'released upon acceptance'; no published version found in Crossref or OpenAlex.
Version
arXiv v1 (2026-03-14)
Source url
https://arxiv.org/html/2603.13676v1
DOI
Update target
SEA-Mem case memory (extracted feature profile and outcome per patient) and subgroup outcome-rate patterns.
Fixed components
GPT-4o radiologist, biochemist, oncologist, integrator and reasoning agents and prompts; 23-publication trial knowledge base; feature-matching retrieval.
Feedback
Documented treatment outcomes (PSA response, overall survival beyond 12 months).
Timing
Described as updating when outcomes become available; the order and timing of memory population in the reported experiments are not reported.
Retained artifact
Patient feature profiles with outcomes; empirical outcome rates for feature combinations shared by three or more cases.
Update generator
Rule-based: cases appended and subgroup rates recomputed; no LLM-generated memory content.
Selection and release
Patterns formed only when at least three cases share a feature combination; no other gate reported.
Evaluation separate from updating
35 real patients and 400 synthetic cases; single-LLM, MedAgents, MDAgents and MedAgent-Pro baselines; w/o SEA-Mem ablation on real data.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
None; the update rule is fixed.
Clinical setting
Pre-therapy outcome prediction for 177Lu-PSMA radioligand therapy in mCRPC from PSMA PET/CT reports, laboratory tests and clinical notes.
Evidence boundary
Single hospital, n=35; exclusion of the query patient from memory (leave-one-out or sequential order) not reported; synthetic cases are LLM-generated; no variance or significance tests.
Locators
Sec. 2.3: SEA-Mem structure, pattern learning, retrieval, update
Sec. 3.1: real and synthetic data, baselines
Table 1: real-patient results
Table 2: synthetic results by complexity
Table 3: ablation including w/o SEA-Mem

48. OpenHospital

Key
liu2026openhospital
Name
OpenHospital
Title
OpenHospital: A Thing-in-itself Arena for Evolving and Benchmarking LLM-based Collective Intelligence
Authors
Peigen Liu; Rui Ding; Yuren Mao; Ziyan Jiang; Yuxiang Ye; Yunjun Gao; Ying Zhang; Renjie Sun; Longbin Lai; Zhengping Qian
Year
2026
Publication status
P: arXiv preprint (v3); no journal_ref or acceptance comment; no published version found in Crossref or OpenAlex.
Version
arXiv v3 (2026-05-31)
Source url
https://arxiv.org/html/2603.14771v3
DOI
Update target
Accumulated clinical experience of 38 physician agents produced by post-case reflection; storage form not specified.
Fixed components
Qwen3-Next-80B-A3B weights for physician and patient agents; Agent-Kernel action space; synthetic patient records and personas.
Feedback
Ground-truth diagnosis, gold examination list and guideline-based treatment comparison used in post-case self-critique.
Timing
After each case, across 22 sequential training batches of about 500 cases.
Retained artifact
Diagnostic, investigative and treatment reflections described as accumulated clinical experience; format and retrieval not described.
Update generator
Physician agents' own multi-dimensional self-critique.
Selection and release
Not reported.
Evaluation separate from updating
A 9:1 train/test split of 12,000 synthetic records is stated, but reported trajectories are plotted across training batches; no ablation or comparator system.
Fixed-procedure control
Not reported.
Evidence concerning improver change
None reported.
Clinical setting
Simulated multi-department hospital with LLM patient agents built from synthetic records (583 diseases, 19 departments).
Evidence boundary
Fully synthetic, text-only arena; mechanism under-specified; no no-reflection control or repeated runs; authors state it is not for clinical decision support.
Locators
Sec. 4.2: dataset split, 22 batches, baseline reflection mechanism
Sec. 4.3: evolution results
Figure 4: performance trajectories across batches
Sec. 3.1: synthetic patient data construction
Limitations

49. SkinGPT-X

Key
chen2026skingptx
Name
SkinGPT-X
Title
SkinGPT-X: A Self-Evolving Collaborative Multi-Agent System for Transparent and Trustworthy Dermatological Diagnosis
Authors
Zhangtianyi Chen; Yuhao Shen; Florensia Widjaja; Yan Xu; Liyuan Sun; Zijian Wang; Hongyi Chen; Wufei Dai; Juexiao Zhou
Year
2026
Publication status
P: arXiv preprint; no journal_ref or acceptance comment; no published version found in Crossref or OpenAlex.
Version
arXiv v1 (2026-03-27)
Source url
https://arxiv.org/html/2603.26122v1
DOI
Update target
EvoDerma-Mem graph database of historical cases (image embedding, key findings, expert diagnosis) and versioned per-disease diagnostic guidelines.
Fixed components
Qwen3-VL vision agent, fine-tuned disease-diagnosis model, Qwen3 case-review and summarization agents and prompts, Oxford-handbook RAG store, image feature extractor.
Feedback
Expert diagnosis labels of stored historical (training) cases.
Timing
Guidelines rewritten when new samples reach threshold N_thresh; the N_thresh value and whether test cases are added are not reported.
Retained artifact
Case triplets and guideline versions G_C^t per disease.
Update generator
LLM Summarization Agent that merges the previous guideline with new key findings.
Selection and release
Threshold-triggered update with contradiction resolution inside the summarizer; no validation gate before use reported.
Evaluation separate from updating
Held-out test splits of Dermnet, HAM10000, DDI31, Fitzpatrick-17k, Dermnet498 and RSDD; with/without memory ablation; blinded review of 600 evolutionary cases by two dermatologists.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
None; the summarizer procedure is fixed (each guideline update takes the previous version as input).
Clinical setting
Dermatology image diagnosis on public clinical and dermoscopic datasets and a web-curated rare-disease set.
Evidence boundary
Offline public benchmarks; memory built from labelled training data; RSDD images gathered from websites; no prospective use.
Locators
Sec. IV-C: EvoDerma-Mem storage and guideline evolution
Sec. IV-D3: case review with retrieved cases and evolved guidelines
Sec. II-E: memory ablation and dermatologist review
Figure 5: memory ablation, guideline evolution timeline, physician ratings
Sec. IV-B: datasets and splits

50. SEA

Key
li2026sea
Name
SEA
Title
Joint Optimization of Reasoning and Dual-Memory for Self-Learning Diagnostic Agent
Authors
Bingxuan Li; Simo Du; Yue Guo
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; no Crossref or OpenAlex venue record found 2026-11-04)
Version
arXiv v2, 2026-08-07 (v1 2026-04-08)
Source url
https://arxiv.org/html/2604.07269v2
DOI
Update target
Qwen3-4B/8B policy weights trained by round-wise GRPO to diagnose and issue memory operations; at deployment the short-term case memory (capacity K) and long-term rule memory change while weights stay frozen (Sec 3.1-3.3, App F.2.2).
Fixed components
Closed-set candidate pipeline (ground truth plus 199 GPT-5.2-ranked distractors, fixed seeds), memory tool schema (list/append/pop/consolidate), reward weights and schedule, decoding settings; weights frozen in the long-horizon stream (Sec 4.1.1, App D, App E.1-E.2).
Feedback
Ground-truth diagnosis returned after each case; training reward of +5/-5 for correctness plus a short-term-memory occupancy penalty (alpha = 3), linearly re-weighted across rounds (Sec 3.2-3.3, Sec 4.2.1).
Timing
Offline RL on the MedCaseReasoning training split (about 48 GPU-hours on 8 A100 for Qwen3-8B); online memory updates after every case in the ER-Reason stream (Sec 4.2.1, App E.2).
Retained artifact
Outcome-annotated short-term case records and natural-language diagnostic rules consolidated from evicted cases, carried forward across stream rounds (Sec 3.1).
Update generator
The trained policy itself decides when to append or pop cases and writes the consolidated rules through tool calls (Sec 3.1, App E.1).
Selection and release
No gate on memory writes beyond the agent's own consolidate actions; training checkpoint chosen by best validation accuracy (Sec 3.1, App E.2).
Evaluation separate from updating
MedCaseReasoning official test split (each case independent, 10 training cases as memory context, App F.2.3); ER-Reason never trained on and streamed once with frozen weights (Sec 4.2.1, Table 4).
Fixed-procedure control
Same dual-memory tool and prompt with untrained or SFT writers (Qwen-8b SFT + Dual Memory, GPT-5.2 + Dual Memory) and memoryless SFT/RL baselines (Table 2); component ablations on MedCaseReasoning (Table 3).
Evidence concerning improver change
RL with a memory reward revises the policy that writes memory, and stream gains are compared with untrained writers using the identical memory tool (Table 2); writer and diagnostician share weights, so the improver effect is not separated from reasoning gains.
Clinical setting
Closed-set diagnosis (1 of 200 candidates) from published case reports (MedCaseReasoning) and emergency-room cases (ER-Reason); no deployment (Sec 4, App B).
Evidence boundary
Closed-set labels with immediate ground-truth feedback; one stream order, no seeds or intervals for stream metrics; Table 1 and Table 3 gains come from a setting without cross-case carryover; rule quality rated by one internal medicine resident; preprint.
Locators
Sec 3.1: dual-memory design and memory operations as agent actions
Sec 3.2-3.3: diagnostic and memory rewards, round-wise schedule, GRPO
Table 2: ER-Reason long-horizon final accuracy and Delta Acc@50/@100
App F.2.3 and Table 4: standard vs long-horizon protocols, no test carryover in standard setting
Sec 5.1 and App G: expert evaluation by one physician

51. Healthcare AI GYM (TT-OPD)

Key
jeong2026healthcaregym
Name
Healthcare AI GYM (TT-OPD)
Title
Healthcare AI GYM for Medical Agents
Authors
Minbyul Jeong
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; no Crossref or OpenAlex venue record found 2026-11-04)
Version
arXiv v1, 2026-05-01
Source url
https://arxiv.org/html/2605.02943v1
DOI
Update target
Full parameters of Qwen3.5-9B trained from scratch (no SFT warm-up) by vanilla GRPO or TT-OPD, i.e. GRPO with cosine length reward plus turn-level KL to an EMA teacher, for 60 steps (Sec 4.2, Sec 5.1, App F).
Fixed components
Gym with 10 domains, 3,631 tasks (2,657 train, 307 validation), 135 tools, 828K-passage knowledge base, domain policy prompts, 5D reward weights, AgentRunner; EMA rule alpha = 0.995 every 5 steps with hard copy every 30 steps (Sec 3, Sec 4.2, App A).
Feedback
Episode-level 5D reward (accuracy 0.25, process 0.20, safety 0.20, format 0.10, coherence 0.10) plus cosine length reward; outcome-conditioned hints given only to the teacher (Sec 3, Sec 4.2).
Timing
Online RL during training (60 steps on 8 A100 80GB); no updates at evaluation (Sec 5.1).
Retained artifact
Trained policy checkpoint, evaluated with the same multi-turn AgentRunner, tools and knowledge base (Sec 5.2).
Update generator
Student on-policy rollouts in the gym; an EMA copy of the student, conditioned on outcome hints, supplies per-turn KL targets (Sec 4.2).
Selection and release
Not reported; results given for TT-OPD at step 60 and GRPO peak at step 55 on the validation set (Sec 5.3).
Evaluation separate from updating
307 held-out validation tasks from the training distribution; 18 external benchmarks (MC QA, VQA, EHR, long-form QA) with test-set fingerprinting for contamination (Sec 5.1, Table 1, App G).
Fixed-procedure control
Base+AR (same tools and knowledge base, no RL), vanilla GRPO with identical hyperparameters, four OPD variants adding one component each (Table 1, Sec 6.1, Fig 4).
Evidence concerning improver change
The teacher that shapes each update is itself updated as an EMA of the student; ablations compare teacher rules (periodic reset, EMA, EMA plus hints, full) on validation curves, not the learned teacher's contribution alone (Sec 6.1).
Clinical setting
Simulated tool-based clinical tasks (EHR, triage, drug interaction, visual and clinical diagnosis, psychiatry, obstetrics, pathways) and offline medical QA/VQA/EHR benchmarks; no clinician or patient involvement (Sec 3, App A.2).
Evidence boundary
Training tasks are converted from the same benchmark families used for testing; single run without seeds or intervals; text says best on 10 and on 12 of 18 benchmarks; below text-only base on MMLU-Med, SLAKE, PMC-VQA; preprint.
Locators
Sec 3: environment design, tools, 5D reward
Sec 4.2: TT-OPD, EMA teacher, outcome-conditioned hints
Table 1: 18 benchmarks, Base (text), Base+AR, GRPO, TT-OPD
Sec 6.1 and Fig 4: OPD failure progression ablation
App A.2: task sources, AutoTaskGenerator conversion of MedQA/MedMCQA/MMLU/VQA/MIMIC

52. CodeClinic (clinical autoformalization)

Key
ossowski2026codeclinic
Name
CodeClinic (clinical autoformalization)
Title
CodeClinic: Evaluating Automation of Coding Skills for Clinical Reasoning Agents
Authors
Timothy Ossowski; Xinchi Liu; Danyal Maqbool; Vaibhav Dhanuka; Sheng Zhang; Hoifung Poon; Majid Afshar; Tyler Bradshaw; Junjie Hu
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; no Crossref or OpenAlex venue record found 2026-11-04)
Version
arXiv v1, 2026-05-10
Source url
https://arxiv.org/html/2605.09675v1
DOI
Update target
Shared library of verified Python functions, one per MIMIC-IV clinical concept (63 concepts), built offline and then given to inference agents (Sec 4.1-4.3).
Fixed components
Inference LLM weights, MIMIC-IV schema and query_db tool, LLM-generated guidelines from PubMed abstracts, verification agent, threshold 0.90, up to 100 iterations; the library built with Qwen3.5-27B is reused for every backbone (Sec 4, Sec 5, App D, App F).
Feedback
Accuracy of each candidate function on the 10% verification split (about 100 labeled rows per variant), with per-question error traces fed back to the builder (Sec 4.2, App A).
Timing
Offline construction per concept in topological dependency order (about 20 GPU-hours per concept); library frozen as a static resource at inference (Sec 4.3, App I).
Retained artifact
Library of verified concept functions exposed as signatures; later concept builds can search and load earlier verified functions (Sec 4.1, App G.1).
Update generator
ReAct autoformalization agent (Qwen3.5-27B) that explores the schema, retrieves guidelines and writes code (Sec 4.1).
Selection and release
Function saved to the shared library when verification accuracy reaches 0.90; otherwise the best candidate across iterations is retained (Sec 4.2, App F Algorithm 1).
Evaluation separate from updating
Held-out 90% of compositional queries (56,700 test vs 6,300 train) and the longitudinal ICU surveillance task (2,000 held-out stays, subject-level split) not used to build the library (Sec 3.1-3.2, Sec 5).
Fixed-procedure control
Zero-shot agent with the same schema and query_db tool; GRPO fine-tuning with 50 examples per concept on the same backbone (Table 3, Fig 6-7).
Evidence concerning improver change
Not reported; the build-and-verify procedure stays fixed.
Clinical setting
Retrospective MIMIC-IV ICU EHR: compositional concept extraction (e.g. SOFA, Sepsis-3) and 4-hourly surveillance over 48 h with 13 checkpoints (Sec 3).
Evidence boundary
Single institution; library built by one model; small verification sets; LLM-written guidelines; frontier models run on a 21% subset; longitudinal gains shown only in Fig 6 without a numeric table; preprint.
Locators
Sec 4: autoformalization pipeline (tool generation, verification, inference)
Table 3: zero-shot vs GRPO vs autoformalization on held-out compositional test set
Fig 6: longitudinal step accuracy, zero-shot vs autoformalized library
Sec 3.2: concepts, dependency levels, 10%/90% split
App A: limitations

53. SepsisAgent

Key
wu2026sepsisagent
Name
SepsisAgent
Title
Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model
Authors
Minghao Wu; Yuting Yan; Zhenyang Cai; Ke Ji; Chuangsen Fang; Ziying Sheng; Xidong Wang; Rongsheng Wang; Hejia Zhang; Shuang Li; Benyou Wang; Hongyuan Zha
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; no Crossref or OpenAlex venue record found 2026-11-04)
Version
arXiv v1, 2026-05-14
Source url
https://arxiv.org/html/2605.14723v1
DOI
Update target
Full parameters of Qwen3-4B-Instruct through Stage I SFT (about 100k instances), Stage II behavior cloning (about 1k), Stage III GRPO (about 3k trajectories, 300 steps) in the world model (Sec 4, App G.2).
Fixed components
Clinical World Model trained beforehand on the MIMIC-IV training split, GPT-4.1-synthesized rationales and propose-simulate-refine traces, reward coefficients, guideline constraints, 5x5 fluid-vasopressor action grid (Sec 3, Sec 4.2-4.4, App G.3).
Feedback
Composite reward from world-model rollouts: survival +15 or mortality -15, SOFA and lactate step terms, -10 per guideline violation, -5 for not terminating; scaled by 0.1 and clipped to [-2, 2] (Sec 4.4, App G.3).
Timing
Offline staged training; no updates during evaluation (Sec 4, Sec 5.1).
Retained artifact
Trained agent checkpoint used with the propose-simulate-refine workflow at inference; outcome prediction also tested without simulator access (Sec 2.3, Sec 5.4).
Update generator
GPT-4.1 for Stage I-II supervision; the agent's own multi-turn rollouts inside the world model for Stage III (Sec 4.2-4.4).
Selection and release
Not reported beyond terminating GRPO after 300 steps (App G.2).
Evaluation separate from updating
725-episode held-out MIMIC-IV test set with off-policy estimators (DR, WIS, WPDIS), guideline adherence and unsafe-action rates never optimized in training; 2,862 MIMIC-III episodes (Sec 5.1, App I).
Fixed-procedure control
Same world model and workflow without training (o3, Gemini-3-Flash, GPT-4.1-mini + WM), untrained base model, and cumulative stage ablation (Table 3, Table 4).
Evidence concerning improver change
Not reported.
Clinical setting
Retrospective ICU sepsis fluid and vasopressor dosing at 4-hour steps on 20,092 MIMIC-IV stays; no clinician review or deployment (Sec 3, App C).
Evidence boundary
Off-policy value estimates only; world model and policy trained on the same source; MIMIC-III check comes from the same hospital system; no seeds or intervals; authors state not for clinical use; preprint.
Locators
Sec 4.4: world-model-based agentic RL, rollouts and composite reward
Table 3: OPE, guideline adherence and unsafe actions vs baselines
Table 4: cumulative stage ablation and IHM/VR prediction without simulator
App I, Table 11: MIMIC-III cross-dataset OPE
App G: training data, settings and reward coefficients

54. VDSS

Key
li2026vdss
Name
VDSS
Title
Human-in-the-Loop Multi-Agent Ventilator Decision Support with Contextual Bandit Preference Learning
Authors
Sijia Li; Xiaoyu Tan; Qixing Wang; Weiyi Zhao; Chen Zhan; Teqi Hao; Xuemin Wang; Lei Gu; Roland Eils; Xihe Qiu
Year
2026
Publication status
C (MICCAI 2026, Lecture Notes in Computer Science, pp. 290-300, published 2026-09-25 per Crossref)
Version
arXiv v1, 2026-05-22 (examined); LNCS version not examined
Source url
https://arxiv.org/html/2605.23320v1
DOI
10.1007/978-3-032-38260-3_28
Update target
Per-clinician contextual-bandit preference state over 12 preference categories (arms), plus long-term memory entries holding cycle summaries and preference signals (Sec 2.2-2.3).
Fixed components
LLM/VLM agents (Waveform Analyzer, Detection, Phase Goal Manager, Hold/Adjust Gate, Strategy Selector, Mode Select, Parameter Planner, Reflect, Note Generator), deterministic safety and compatibility checks, backbone weights, round budget K_max (Sec 2.1-2.2).
Feedback
Clinician accept or reject per proposal with structured rejection rationale; final accepted settings and a Note-Generator-extracted preference signal drive the update (Sec 2.2-2.3).
Timing
Online, one bandit update at the close of each adjustment cycle (Algorithm 1).
Retained artifact
Clinician preference state passed to the reasoning engine as preference context that modulates candidate ranking in later cycles; summaries written to long-term memory (Sec 2.3).
Update generator
Note Generator LLM extracts the preference signal; BanditUpdate rule, whose algorithm is not specified (Sec 2.3).
Selection and release
Updates only from clinician-accepted outcomes, not from rejected proposals; no further gating reported (Sec 2.3).
Evaluation separate from updating
Retrospective replay on a multi-center ICU cohort (1,309 records, 7,447 setting entries, 13 modes) for next-setting prediction; expert ratings of 100 cycles (Sec 3, Table 1).
Fixed-procedure control
Single-model end-to-end generation with the same backbones; NoPref, NoImg and NoImgNoPref ablations within VDSS (Table 1).
Evidence concerning improver change
Not reported.
Clinical setting
ICU mechanical-ventilation titration with clinician-in-the-loop review, retrospective trajectory replay and prototype interface (Sec 3-4).
Evidence boundary
Cross-cycle learning shown only by a regret trend over 100 cycles from one clinician (Fig 2) without numbers or statistics; patient span, number of experts, replay feedback source and bandit algorithm not stated.
Locators
Sec 2.2 and Algorithm 1: cycle-end preference update and long-term memory writes
Sec 2.3: contextual bandit over 12 preference categories
Table 1: replay metrics, clinician ratings, NoPref/NoImg ablations
Fig 2: regret over 100 cycles from a single clinician

55. Traj-Evolve

Key
zeng2026trajevolve
Name
Traj-Evolve
Title
Traj-Evolve: A Self-Evolving Multi-Agent System for Patient Trajectory Modeling in Lung Cancer Early Detection
Authors
Sihang Zeng; Matthew Thompson; Ruth Etzioni; Meliha Yetisgen
Year
2026
Publication status
P
Version
arXiv v1 (2026-06-01); no journal_ref or acceptance comment
Source url
https://arxiv.org/html/2606.02812v1
DOI
10.48550/arXiv.2606.02812
Update target
A non-parametric experience pool (ExPool) of rejection-sampled reasoning traces and the QLoRA weights of the worker and manager GPT-OSS-20B agents (multi-agent reward-ranked fine-tuning).
Fixed components
Traj-CoA chain-of-agents architecture and prompts, nomic-embed-text-v1.5 embedder, label-conditioned selection rule, binary reward thresholds (s>=6 for cases, s<=4 for controls), and the test sets.
Feedback
Ground-truth one-year lung cancer labels of training patients, used for best-of-4 label-conditioned selection (ExPool) and a binary acceptance reward (MARL) (Sec 3.3-3.4).
Timing
Offline before testing; a stream of verified patients is simulated by incrementally sampling the training set, MARL is one QLoRA epoch, and inference uses the full training-set ExPool (Sec 3.1, 3.5, 4).
Retained artifact
ExPool vector store of (summary embedding, manager rationale, predicted risk, label) for 13,629 training patients, plus fine-tuned worker and manager adapters.
Update generator
The base system's own rollouts (m=4, temperature 1.5) filtered by fixed rejection sampling; in the combined variant MARL rollouts are conditioned on leave-one-out ExPool retrieval (Sec 3.5).
Selection and release
No accept/reject gate on updates; hyperparameters and thresholds set on a validation split; single training round released (Appendix B).
Evaluation separate from updating
Disjoint held-out test sets from the same institution (overall n=1,000; never-smoker n=835), bootstrap 1,000 resamples (Sec 4).
Fixed-procedure control
Static zero-shot Traj-CoA and RAG with the same backbone, plus ExPool-only, MARL-only and combined variants (Tables 1-2).
Evidence concerning improver change
None; rejection rule, reward and training recipe are fixed.
Clinical setting
One-year incident lung cancer prediction from up to five years of multimodal EHR (codes, labs, notes, radiology reports) in a 1:10 matched case-control cohort from one medical center.
Evidence boundary
Single-institution retrospective case-control design with a simulated patient stream and one MARL round; authors note label delay and need for prospective, multi-institution validation (Limitations).
Locators
Sec 3.3 Evolving Experience Pool
Sec 3.4 Multi-Agent Reinforcement Learning
Table 1 overall population results
Fig 4 and Sec 5.2 ExPool size dynamics
Sec 6 Conclusion and Limitations

56. MedEvoEval

Key
zhang2026medevoeval
Name
MedEvoEval
Title
MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes
Authors
Hui Zhang
Year
2026
Publication status
A*
Version
arXiv v3 (2026-10-18), author comment: camera-ready accepted at NeurIPS 2026 Evaluations and Datasets Track
Source url
https://arxiv.org/html/2606.28900v3
DOI
10.48550/arXiv.2606.28900
Update target
The doctor agent's external memory bank of success, failure-correction and action-rule cards, retrieved by TF-IDF (top K=3) before later decisions.
Fixed components
Doctor model (Qwen-Turbo), patient simulator (Qwen-Turbo), manager scorer (Kimi-K2.5), test catalog, rubrics, case order, action budget of 10.
Feedback
Manager-side rubric scores of diagnosis, evidence and management after FINALIZE, with access to the reference diagnosis (Sec 3.3-3.4).
Timing
Online sequential write-back after each episode over a 500-case stream (passes E1-E3) and continued during held-out evaluation (H3).
Retained artifact
Unbounded memory bank; 4,500 records after E3 (312 success, 1,188 corrective, 3,000 rule cards) (Appendix B.4).
Update generator
The manager LLM writes compact IF-THEN and IF-AVOID cards after scoring.
Selection and release
No acceptance gate or eviction; every card is written back.
Evaluation separate from updating
200 held-out cases sharing no content with the stream, but same source and H3 keeps writing memory during evaluation; two-physician audit of 80 scored outputs.
Fixed-procedure control
No-memory conditions E0 and H0 with identical doctor, simulator, order and scorer; frozen-bank relevant vs case-mismatched retrieval control (Table 27).
Evidence concerning improver change
None.
Clinical setting
Simulated Chinese outpatient first-visit diagnostic episodes converted from MedQA-derived cases, with action-gated inquiry and test requests.
Evidence boundary
Simulated standardized episodes; stream gains E2-E3 combine repeated exposure with memory; held-out clinical score and strict accuracy not significant after Holm (Sec 6, Table 21).
Locators
Sec 3.4 Experience write-back
Table 4 longitudinal, held-out and BWT results
Table 21 paired H3-H0 held-out endpoints
Table 27 frozen-bank retrieval control
Sec 6 Discussion, limitations and ethics

57. TopoAgent

Key
meng2026topoagent
Name
TopoAgent
Title
TopoAgent: An Agentic Framework for Automated Topology Learning in Medical Imaging
Authors
Guangyu Meng; Pengfei Gu; Xueyang Li; Yiyu Shi; Erin Wolf Chambers; Danny Z. Chen
Year
2026
Publication status
P
Version
arXiv v1 (2026-06-29)
Source url
https://arxiv.org/html/2606.29763v1
DOI
10.48550/arXiv.2606.29763
Update target
Long-term memory of structured diagnostic entries (failed descriptor, diagnosed cause, successful correction), accumulated within a dataset and reset between datasets.
Fixed components
GPT-4o backbone at temperature 0, 21 perception and descriptor tools, distilled skill set (properties, tiered rankings, parameters), Perception-Reasoning-Action-Reflection pipeline, classifiers.
Feedback
LLM Reflection verdict on feature-vector quality (sparsity, variance, kurtosis and similar statistics); no ground-truth labels enter the memory (Sec 3.2-3.3).
Timing
Online during runs; an entry is written whenever Reflection fails and triggers a retry.
Retained artifact
Per-dataset long-term memory entries consulted by later Determination and Reflection steps.
Update generator
The LLM Reflect step writes the diagnosis and correction (Algorithm 1, line 17).
Selection and release
No gate; entries are written on failure without verification.
Evaluation separate from updating
26 TopoBenchmark datasets with 5-fold CV on frozen splits, but the skill set was distilled from the same benchmark at object-type level; three external datasets for downstream integration.
Fixed-procedure control
Ablations without dual memory and without long-term memory under the same pipeline (Table 2).
Evidence concerning improver change
None.
Clinical setting
Selecting topological descriptors for 2D medical image classification across 11 modalities (histopathology, retina, X-ray, dermoscopy and others).
Evidence boundary
Memory is per dataset and reset; skill set built from evaluation datasets; endpoint is classifier accuracy, not clinical decisions; about 32 s per image (Sec 5).
Locators
Sec 3.3 Dual Memory and Algorithm 1
Table 1 main results
Table 2 ablation
Sec 5 Conclusions and limitations

58. SAGEAgent

Key
qu2026sageagent
Name
SAGEAgent
Title
SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction
Authors
Chongyu Qu; Can Cui; Zhengyi Lu; Junchao Zhu; Tianyuan Yao; Junlin Guo; Juming Xiong; Yanfan Zhu; Yuechen Yang; Bennett A. Landman; Yuankai Huo
Year
2026
Publication status
C
Version
arXiv v1 (2026-07-10) examined; published in MICCAI 2026, LNCS pp. 536-546 (Crossref)
Source url
https://arxiv.org/html/2607.09521v1
DOI
10.1007/978-3-032-38059-3_51
Update target
Episodic memory of the agent's own stage decisions with rewards, and a semantic memory of up to 10 stage-indexed stop/acquire rules.
Fixed components
Frozen Qwen-2.5-7B-Instruct, frozen Transformer survival predictor and uncertainty head, text tools, FAISS index of training cases, MCDA burdens and reward formulas.
Feedback
Stage reward (-burden + uncertainty reduction) and terminal reward ((1-uncertainty) - total burden); top and bottom reward quartiles label episodes good or poor to score rule effectiveness (Sec 2.3).
Timing
Experience accumulation on training patients, 3 episodes each (about 408 episodes per fold), reflection every 10 patients, 10 cycles per fold (Sec 3.2, 3.5).
Retained artifact
Episode store and active rule set; four rules recurred in at least 4 of 5 folds (Table 3).
Update generator
The frozen LLM's reflection module proposes rules from an effectiveness report.
Selection and release
Proposed rules are checked for direction clarity and novelty; rules with consistently poor outcomes are deprecated; 10.8 candidates and 4.6 survivors per fold (Sec 3.5).
Evaluation separate from updating
Nested 5x5 cross-validation on 170 complete-modality patients with outer-fold test patients; whether memory keeps updating on test patients is not reported.
Fixed-procedure control
Component ablation from base LLM through tools, episodic and semantic memory with the same frozen LLM (Table 2); Reflexion baseline (Table 1).
Evidence concerning improver change
None.
Clinical setting
Glioma survival prediction with ordered acquisition of demographics, radiology, pathology and genomics (TCGA-LGG, TCGA-GBM, BraTS).
Evidence boundary
Small public cohort (170 evaluation patients); rewards use model uncertainty and burden rather than outcomes; C-index slightly below full-workup fusion.
Locators
Sec 2.3 episodic and semantic memory
Table 1 core results
Table 2 component ablation
Table 3 discovered rules and Sec 3.5
MICCAI 2026 LNCS version, pp. 536-546

59. Cura 1T

Key
chen2026cura
Name
Cura 1T
Title
Cura 1T: Healthcare Foundation Model via Recursive Self-Improvement
Authors
Haolin Chen; Leon Qi; Steve Brown; Deon Metelski; Tao Xia; Joonyul Lee; Qixuan Wang; Kevin Riley; Frank Wang; Weiran Yao
Year
2026
Publication status
P
Version
arXiv v3 (2026-10-05); no journal_ref
Source url
https://arxiv.org/html/2607.15314v3
DOI
10.48550/arXiv.2607.15314
Update target
The post-training data mixture and resulting Kimi-K2.6 (1T) weights, trained per round by SFT screen, RL and self-distillation fine-tuning.
Fixed components
Base model, benchmark harnesses (one MedAgentBench harness bug fix counted as a round), data-synthesis skill set, training stack, human review step.
Feedback
Graded benchmark trajectories and scores on MedAgentBench, HealthBench Professional/Hard, MedXpertQA and AgentClinic, categorized into failure modes (Sec 3.2).
Timing
Offline rounds per capability loop (1-5 rounds shown), then consolidation of the accumulated mixture into the released model (Sec 4).
Retained artifact
Validated data refinements accumulated into a consolidated mixture; kept checkpoints serve as the next round's start; released Cura 1T weights.
Update generator
A training agent (model identity not reported) using reasoning-correction (frontier-model explanations), knowledge-injection, behavior-calibration, retention-anchor and mixture-curation skills.
Selection and release
A human approves each plan and makes keep, revert, refine or deploy decisions from benchmark scores; reverted rounds are reported (Fig 1, Tables 2-5).
Evaluation separate from updating
None for healthcare: the selection benchmarks also give the reported scores (authors call them development measurements, Sec 4.6); AIME, GPQA-Diamond and tau2-Bench were not used for selection.
Fixed-procedure control
Base Kimi-K2.6 and frontier comparators only; authors state isolating contributions or further ablations was beyond capability (Limitations).
Evidence concerning improver change
Not tested.
Clinical setting
Patient-care answers, expert medical QA with images, simulated interactive diagnosis (AgentClinic) and FHIR EHR tool use (MedAgentBench).
Evidence boundary
Benchmarks guide keep/revert and report results; human-in-the-loop decisions; no ablations; not evidence of clinical safety (Sec 6).
Locators
Sec 3.2 Recursive self-improvement loop
Table 2 MedAgentBench rounds
Table 4 MedXpertQA rounds
Sec 4.6 Improvement across benchmarks
Sec 6 Conclusion, Limitations, Safety

60. MIRA

Key
wang2026mira
Name
MIRA
Title
MIRA: Medical Image Reflection for Agentic Diagnosis
Authors
Shengzhi Wang; Jun Yang; Kai Wu; Xiaozhong Ji; Yiwen Ye; Ziyang Chen; Mingliang Xiong; Wen Fang; Mingqing Liu; Mengyuan Xu; Miaoxuan Shan; Caiyan Liu; Bin He; Qingwen Liu
Year
2026
Publication status
P
Version
arXiv v1 (2026-08-11), PDF only
Source url
https://arxiv.org/pdf/2608.10827v1
DOI
10.48550/arXiv.2608.10827
Update target
Qwen3-VL-8B policy weights (SFT then GRPO) and a global textual reflection memory of diagnostic principles injected into the system prompt.
Fixed components
Six-tool suite, fixed Seed-series reward judges and prompts, reward coefficients, validation set, static quality gate, edit-budget schedule.
Feedback
Composite reward (answer correctness, consistency judged partly against applicable memory principles, format); rollouts with reward below 0.3 feed memory edits; validation reward gates acceptance (Sec 3.2-3.3).
Timing
Online during RL (750 steps, about one epoch); memory update after rollout rounds with at least 4 unique failed cases.
Retained artifact
Accepted memory saved with each policy checkpoint; released MIRA-VL-8B weights, trajectories and tool environment.
Update generator
A reflection optimizer and editor (model not reported) propose bounded patch edits; an MCTS teacher engine (Seed 1.8, Gemini 2.5 Pro, GPT-5.2) built the SFT data.
Selection and release
Candidate memory accepted only if mean validation reward under the frozen updated policy improves; rejected edits stored to steer later proposals (Appendix E).
Evaluation separate from updating
Test splits of nine medical VQA benchmarks reported as never used for training or selection; whether the final memory is injected at test time is not stated explicitly.
Fixed-procedure control
RL ablation with format+accuracy, plus consistency, plus reflection memory (Table 4); validation curve with and without memory (Fig A7).
Evidence concerning improver change
None; the gate and editor are fixed, only the edit budget follows a preset schedule.
Clinical setting
Medical image question answering across radiology, pathology, dermatology, lesion and expert multimodal benchmarks.
Evidence boundary
Benchmark VQA rather than clinical workflow; single runs; memory's role inside the reward not isolated; authors note limited memory generalization (Sec 5).
Locators
Sec 3.2-3.3 composite reward and on-policy reflection memory
Table 1 main results (p. 10)
Table 4 RL ablation (p. 13)
Appendix E online reflection memory details (pp. A13-A16)
Sec 5 limitations (p. 14)

61. BaT (Benchmark-as-Teacher)

Key
liu2026bat
Name
BaT (Benchmark-as-Teacher)
Title
BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics
Authors
Junqi Liu; Yufan He; Yexiao He; Pengfei Guo; Dong Yang; Andriy Myronenko; Can Zhao; Hanrong Ye; Tianhao Qi; Yuyin Zhou; Daguang Xu; Yucheng Tang
Year
2026
Publication status
P
Version
arXiv v1 (2026-08-17)
Source url
https://arxiv.org/html/2608.16211v1
DOI
10.48550/arXiv.2608.16211
Update target
Qwen3.5-4B/9B policy weights (SFT cold start then multi-round GRPO) and the per-round stage curriculum.
Fixed components
OpenHands execution environment with public stage skills, pool proportions, LLM rubric verifier, controller evaluation, fallback rule.
Feedback
LLM rubric verifier scores rollouts on synthetic sandboxes (binary rubric items times evidence completeness); aggregate stage scores from a fixed held-out controller evaluation route the curriculum (Sec 2.2).
Timing
Offline rounds (about ten) alternating GRPO blocks with re-evaluation (Fig 1c, Sec 4.2).
Retained artifact
Best retained checkpoint, plus per-round records of states, mixtures, reward version and controller scores.
Update generator
Stage Bank Data Factory teacher models synthesize leakage-checked fictional medical-imaging tasks and executable sandboxes (20,299 states); the router picks the weak stage.
Selection and release
Best-so-far checkpoint retained; fallback after three consecutive drops or KL shift above a threshold; best checkpoints auto-selected (Sec 2.2.2).
Evaluation separate from updating
AutoMedBench-Lite scores both steer training and report results (authors: not an untouched final test, Sec 5.1-5.2); ABRA, MedXpertQA-Text and eight external benchmarks were not used for selection.
Fixed-procedure control
Same SFT start trained by plain GRPO with a single final reward, and Stage Bank pool ablations (Table 2, Fig 5).
Evidence concerning improver change
None.
Clinical setting
Long-horizon medical-imaging AI development workflows (plan, setup, validate, inference, submit) and radiology viewer tasks (ABRA).
Evidence boundary
Benchmark participates in adaptation; authors say differences describe observed scores rather than a matched ablation (Sec 4.1); research workflows, not patient care.
Locators
Sec 2.1 Stage Bank
Sec 2.2 BiCuRL outer and inner loops
Table 2 policy-level results
Fig 5 pool ablation
Sec 5.2 Limits

62. AMR

Key
murugesan2026amr
Name
AMR
Title
Adaptive Memory and Reflection Multi-Agent System for Medical Question Answering
Authors
Pradeep Murugesan; Luoxiao Yang; Xueli Chen; Xinqi Fan
Year
2026
Publication status
A* (arXiv comment 'Accepted by IEEE SMC 2026'; no journal_ref and no Crossref record found 2026-11-04)
Version
arXiv v1, 2026-08-19
Source url
https://arxiv.org/html/2608.19029v1
DOI
Update target
Per-agent FAISS memories holding question context, answer, post-hoc notes and, after incorrect predictions, role-specific reflection entries; model weights are not updated (Sec III-C).
Fixed components
GPT-4o for every role, LangGraph routing (Moderator, Recruiter, specialists, Decision Maker, Ethical Overseer), text-embedding-3-large, top-5 retrieval, and a RAG corpus frozen before evaluation (Sec IV-B).
Feedback
Ground-truth multiple-choice label; reflection triggers only when the prediction differs from the label (Sec III-C, Alg. 1).
Timing
Alg. 1 writes feedback after each labelled question; Sec IV-B says test questions, predictions and reflections were never written back, so test-time memory was static; the phase that filled it is not reported.
Retained artifact
Role-specific memory entries and corrective reflection entries in FAISS, retrieved top-5 per agent (Sec III-C, IV-B).
Update generator
The same GPT-4o agents produce answers and reasoning summaries; a reflection step writes corrective feedback against the label (Sec III-C).
Selection and release
No filtering, pruning or validation of memory entries; the authors name pruning and confidence-based retention as future work (Sec V).
Evaluation separate from updating
MedQA official test split and a disjoint MedMCQA subset, evaluated in sequential batches of 50; test-set sizes, seeds and intervals are not reported (Sec IV-A, IV-B).
Fixed-procedure control
Ablation against a multi-agent baseline without memory, reflection or RAG, and feedback-only and memory-only variants (Table IV); no comparison of a revised versus a fixed update procedure.
Evidence concerning improver change
None; the reflection rule and memory format stay fixed.
Clinical setting
USMLE-style (MedQA) and AIIMS/NEET (MedMCQA) multiple-choice exam questions; no patient data.
Evidence boundary
Retrospective exam benchmarks only; Ethical Overseer not clinician-validated; source of test-time memory contents unreported; no uncertainty estimates (Sec IV-B, V).
Locators
Sec III-C: agent-specific memory and reflective update
Algorithm 1: reflective feedback learning loop
Sec IV-B: frozen retrieval corpus, no test write-back
Table IV: memory, feedback and RAG ablation
Table V: comparison with GPT-4, MDAgents, MedAgents

63. MediSkill-Evo

Key
wu2026mediskillevo
Name
MediSkill-Evo
Title
MediSkill-Evo: Process-Constrained Self-Evolution for Evidence-Grounded Clinical Interaction
Authors
Ruoyu Wu; Shenfu Xie; Yinqian Sun; Haibo Tong; Feifei Zhao
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; anonymized code link; no Crossref record found 2026-11-04)
Version
arXiv v2, 2026-08-25
Source url
https://arxiv.org/html/2608.23397v2
DOI
Update target
Four typed external banks: Clinical Skills, Process Rules, Symbolic Schemas and Measurement Skills (visual procedures); backbone weights untouched (Sec 2.1, 2.2).
Fixed components
Hosted Qwen3.6-Flash or DeepSeek-V4-Flash Doctor, Process-Constrained Preference Harness (three candidates, symbolic verifier, Clinical Process Critic, risk audit), Patient and Measurement environment, evaluator, case order and turn ceiling (Sec 3.1).
Feedback
Post-encounter evaluation of each completed training trajectory against reference diagnosis and benchmark targets, then reflection (Sec 2.2, Table S1).
Timing
Offline: proposals after each of 700 training encounters (200 NEJM cases for the Measurement Bank) are merged into snapshots; the final snapshot is frozen and writes are disabled at test (Sec 2.2, 3.1).
Retained artifact
Versioned bank artifacts with content, applicability scope, trajectory provenance and lifecycle status, published as immutable hashed snapshots (Sec 2.2, S1, Table S3).
Update generator
An LLM reflector proposes typed add, merge, patch, deprecate or discard operations after each trajectory (Sec 2.2, S2.1, S2.2).
Selection and release
Typed merge operator and validator check type consistency, provenance, label leakage, controller-defined safety and replay on the proposing trajectories; failing artifacts are withheld or rejected (Sec 2.2).
Evaluation separate from updating
300 held-out MIMIC-IV FullChain encounters, 180 stress conditions from 30 disjoint cases, and 100 NEJM image cases; one rollout per case at temperature 0; LLM judge is the comparison backbone (Sec 3.1-3.3).
Fixed-procedure control
No-memory and single-bank-removal variants on a 100-case subset (Table 7), and six baselines with method-native frozen memory trained on the same 700 encounters (Table 4, Table S3); no revised versus fixed update-procedure test.
Evidence concerning improver change
None; reflector, merge and validation operators stay fixed.
Clinical setting
Simulated doctor-patient-measurement encounters from MIMIC-IV records (history to prescription) and NEJM image-challenge cases under the AgentClinic protocol.
Evidence boundary
Automatic LLM-judged labels without clinician calibration; single run per configuration; the authors say results do not establish clinical safety or causal credit for components (Sec 3.3, 4, 5).
Locators
Sec 2.2: four-bank self-evolution, merge and validation
Table 4: FullChain results on 300 encounters, two backbones
Table 5: controlled-stress aggregates on 180 conditions
Table 7: no-memory and bank-removal ablation
Table S3: comparator ledger and frozen object counts

64. CDEG

Key
dai2026cdeg
Name
CDEG
Title
CDEG: Learning Decision-Critical Evidence for Long-Horizon Diagnostic Agents
Authors
Xiwei Dai; Zijie Meng; Zhiting Fan; Yixuan Tang; Guanyu Jiang; Ziru Niu; Zuozhu Liu
Year
2026
Publication status
P (arXiv preprint; comment '17 pages, 5 figures'; no journal_ref; no Crossref record found 2026-11-04)
Version
arXiv v3, 2026-10-14
Source url
https://arxiv.org/html/2608.22899v3
DOI
Update target
A Counterfactual Diagnostic Evidence Graph of diagnosis, evidence and acquisition-action nodes plus diagnosis-revision edges (Graph Consolidation).
Fixed components
Doctor-agent backbones (Gemini-3-Flash-Preview, Gemini-3.1-Pro, GPT-5.6-Luna, Claude-Sonnet-5), patient environment and tools, 30-turn budget, Qwen3-Embedding and SapBERT matching thresholds (Implementation Details).
Feedback
Correctness of each trajectory's ICD-aligned final diagnosis against the reference, then verifier preference shifts under counterfactual evidence edits (Candidate Evidence Mining; Counterfactual Validation).
Timing
One offline construction pass: K=4 rollouts per construction case from two Gemini agents; the graph is frozen and shared across all evaluated backbones (Implementation Details).
Retained artifact
Frozen graph with 1,156 nodes and 1,426 relations, including 270 diagnosis-revision edges (Table 6).
Update generator
Failed-successful trajectory pairing (cosine > 0.7), UMLS evidence-atom mapping and counterfactual verification by GPT-5.5 (Implementation Details).
Selection and release
A candidate is kept only if the preference shift holds in both diagnosis orders; 692 of 2,271 candidates (30.47%) retained (Counterfactual Validation; Table 7).
Evaluation separate from updating
200 patient-disjoint MIMIC-IDx cases (ID) and 107 AgentClinic-MedQA cases (OOD); accuracy judged by GPT-5.6-Sol; repeated runs or intervals not reported (Benchmarks; Metrics; App A).
Fixed-procedure control
Vanilla agent, self-reflection, Trajectory RAG and Flat Experience baselines from the same construction split, plus a graph without counterfactual validation (Tables 2, 3); no revised versus fixed construction procedure over rounds.
Evidence concerning improver change
None; pairing, mapping and verifier stay fixed; online graph evolution is named as future work (Conclusion).
Clinical setting
Long-horizon interactive diagnosis over MIMIC-IV admissions with notes and matched MIMIC-CXR reports, plus AgentClinic MedQA scenarios.
Evidence boundary
Retrospective records in an interactive environment; LLM-judged accuracy; single construction pass; Claude-Sonnet-5 ID accuracy is below Trajectory RAG (Table 2).
Locators
Methods: candidate mining, counterfactual validation, graph consolidation
Implementation Details: construction agents, frozen shared graph
Table 2: accuracy across four backbones, ID and OOD
Table 3: ablations incl. without counterfactual validation
Tables 6-7: construction and validation statistics

65. LearnActCoder

Key
ghaffari2026learnactcoder
Name
LearnActCoder
Title
LearnActCoder: Role-Aware Error Memory for Adaptive Clinical Coding Agents
Authors
Meysam Ghaffari; Bhaskar Sen; Nasim Sabetpour; Nina Fatehi; Animesh Agarwal; Carlos Morato
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; no Crossref record found 2026-11-04)
Version
arXiv v1, 2026-09-17
Source url
https://arxiv.org/html/2609.19721v1
DOI
Update target
MistakeKDB, a structured JSON error memory; false-negative lessons go to the Coder prompt, false-positive lessons to the Judge prompt (Sec 3.2).
Fixed components
GPT-5 weights, the Coder-Judge workflow with up to three rounds, and CodeLookupDB (192,214 ICD-10-CM and 26,533 CPT/HCPCS entries) (Sec 3.1, 4).
Feedback
Mismatches between predictions and reference administrative codes from MIMIC tables (false positives and false negatives) (Sec 3.2, 4).
Timing
Offline LEARN phase on 25 labelled notes per dataset, then memory reused unchanged in the ACT phase (Sec 3.2, 4, 5.3).
Retained artifact
Lessons with error category, implicated codes, explanation and transferable lesson, consolidated from 1,028 MIMIC-III LEARN mismatches (Sec 3.2, 5.4).
Update generator
A post-hoc LLM MistakeAnalyzer that sees note, predictions and reference codes and assigns one of six error categories (Sec 3.2).
Selection and release
Similar lessons are consolidated and frequent patterns prioritized before prompt injection; no separate validation gate (Sec 3.2).
Evaluation separate from updating
Note-level held-out ACT sets: 150 matched MIMIC-III notes (paired bootstrap, Wilcoxon), 250 matched MIMIC-IV notes, 250 and 1,000 memory-only MIMIC-III notes (Sec 4, 5).
Fixed-procedure control
Same pipeline with and without MistakeKDB on identical notes, plus raw-example and reflection-summary memory formats from the same error pool (Tables 1-3); routing policy not ablated; no revised update-procedure test.
Evidence concerning improver change
None; MistakeAnalyzer and routing rule stay fixed.
Clinical setting
ICD-9/CPT coding of MIMIC-III discharge summaries and ICD-10/HCPCS coding of MIMIC-IV notes.
Evidence boundary
Retrospective single-system study; one model family; administrative labels as proxy; taxonomy not coder-validated; CPT/HCPCS performance remains low (Sec 6).
Locators
Sec 3.2: Learn-Then-Act error memory and role-aware routing
Table 1: matched 150-note MIMIC-III ablation with CIs
Table 2: matched 250-note MIMIC-IV ablation
Table 3: memory-format comparison
Sec 6: limitations

66. CASE

Key
shao2026case
Name
CASE
Title
From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning
Authors
Minye Shao; Chaohui Yu; Yixuan Wu; Fan Wang; Ling Shao; Yang Long; Jing Wang; Hangjie Yuan; Shang Liu
Year
2026
Publication status
P (arXiv preprint; comment lists code and Hugging Face links only; no journal_ref; no Crossref record found 2026-11-04)
Version
arXiv v3, 2026-10-14
Source url
https://arxiv.org/html/2609.39566v3
DOI
Update target
Full weights of one shared Qwen3-VL-8B-Instruct policy acting as Radiologist, Anatomist and Consultant, including the visual pathway (Sec 5).
Fixed components
Tool harness and replayed evidence (Vista3D masks, MedGemma-27B reports), role prompts and tool budgets, Kimi K2.6 rubric judge, a frozen Consultant skill store built from teacher trajectories (Sec 5.1, 5.3, App B).
Feedback
Exact-match answer reward, format reward, Kimi K2.6 rubric score with top-2 linkage, and a Causal-KL penalty from an answer-privileged view of the same policy snapshot (Sec 5.4, 5.5, Eq. 7).
Timing
Offline: RSFT on screened teacher trajectories, then one epoch of on-policy GRPO with G=8 rollouts per prompt; the trained policy is evaluated frozen (Sec 5.2, 5.3, Fig. 3).
Retained artifact
Post-trained CASE-8B policy weights.
Update generator
The learner's own on-policy rollouts; the refreshed policy snapshot also serves as privileged self-teacher for the OPSD penalty (Sec 5.3, 5.4).
Selection and release
Last checkpoint of the RL run; RSFT data screened by tool use and a rubric score of at least 0.4 (Eq. 3); no checkpoint-gating procedure described.
Evaluation separate from updating
4,850 questions from 455 held-out participants (14,135 agent-question evaluations), first-pass exact answer-set accuracy, participant-cluster bootstrap for gaps (Sec 4, App C, App F).
Fixed-procedure control
Component ablation from shared RSFT initialization: RSFT only, GRPO, GRPO+RLAIF, GRPO+OPSD, full (Table 2); the evolving self-teacher is not compared with a frozen teacher.
Evidence concerning improver change
The privileged self-teacher is the evolving policy itself and is refreshed after updates (Sec 5.4); rubric and reward weights stay fixed.
Clinical setting
Longitudinal multi-sequence MRI from UK Biobank baseline and follow-up visits with ICD-10-derived multiple-choice diagnostic questions.
Evidence boundary
Single cohort; labels may be unobservable from images; generated evidence and LLM judges are fallible; utility beyond this cohort untested (Sec 7).
Locators
Sec 5.3: on-policy interaction with replayable environment
Sec 5.5: rubric feedback, composite reward, GRPO
Table 1: accuracy versus medical and frontier models
Sec 6.3 / Table 2: post-training ablations
App B: Consultant skill memory, frozen across runs

67. ClinVerse

Key
chen2026clinverse
Name
ClinVerse
Title
ClinVerse: A Realistic Multi-Agent Diagnostic Environment for Training and Evaluating Medical Agents
Authors
Sirry Chen; Fangke Chen; Xuanwen Ding; Wei Chen; Zhongyu Wei
Year
2026
Publication status
P (arXiv preprint; no comments or journal_ref; no Crossref record found 2026-11-04)
Version
arXiv v1, 2026-10-26
Source url
https://arxiv.org/html/2610.34878v1
DOI
Update target
External doctor-agent artifacts (skills, SkillBank, utility-scored memory, experience bases, cheatsheet, playbook, tool registry) with Qwen3-32B frozen; Appendix C also updates Qwen3-8B weights by SkillRL-GRPO for 30 steps.
Fixed components
Environment roles (dual-layer patient simulator, laboratory physician with ClinLab-2B, environment controller), rule-based judge, ClinPatient cases, and the Qwen3-32B backbone for the nine Table 2 methods (Sec 2.2, App B).
Feedback
Deterministic trajectory-level scores against EHR ground truth: weighted information collection rate, test and abnormal-test coverage, ICD-10 block and three-character accuracy (Sec 2.4).
Timing
L1 offline adaptation from 5,000 trajectories; L2 online and L3 methods over 1,000 optimization trajectories (Sec 4.3); whether online updates ran on the 100 test cases is not reported.
Retained artifact
Method-specific persistent stores: Trace2Skill skill directory, SkillRL SkillBank, MemRL intent-experience-utility bank, MedAgent-Zero record libraries, Dynamic Cheatsheet, ACE playbook, STOP improved improver, MCE skill-context pairs, MedRSI tool registry (App B).
Update generator
Each borrowed method's own LLM distiller, reflector, curator or improver working on doctor-agent trajectories; Table 3 also tests MemRL with Qwen3-32B versus GPT-5.6-Luna trajectories.
Selection and release
Method-specific: MemRL utility re-ranking, ACE curator deduplication, MCE iterative evaluation selection, MedRSI fast-discovery slow-registration over later cohorts; ClinVerse adds no separate release gate (App B).
Evaluation separate from updating
100 cases randomly sampled from a held-out test set, 5 independent runs, mean and SD (Sec 4.1); ClinVerse-adapted agents tested on AIHospital without further adaptation (Table 4).
Fixed-procedure control
Unadapted Qwen3-32B (L0) under the same environment and metrics, plus fixed-procedure L1/L2 methods beside procedure-revising L3 methods (Table 2); no seed-improver versus improved-improver comparison.
Evidence concerning improver change
STOP applies its improver to its own code and MCE evolves its context-engineering skill (App B), but their effect on later learning is reported only as end performance, not isolated (Table 2).
Clinical setting
Simulated diagnostic consultation built from de-identified hospital and ICU EHRs (MIMIC-IV v3.1, MIMIC-IV-Note, SICdb), 150k cases, ICD-10 primary diagnosis as hidden label (Sec 3.1).
Evidence boundary
Simulated environment scored by the authors' own rules; methods reimplemented (SkillRL parameter-free); many gains within one SD over 5 runs; no clinician review of trajectories; preprint.
Locators
Table 2: L0-L3 results for Qwen3-32B with nine methods
Sec 4.3: evolution environment results and trajectory budgets
Appendix B: descriptions of the nine self-evolution methods
Table 4: transfer to AIHospital
Appendix C / Table 7: SkillRL-GRPO weight updates on Qwen3-8B

68. GuideSkill

Key
cao2026guideskill
Name
GuideSkill
Title
GuideSkill: Evolving Executable LLM Agent Skills for Guideline-Grounded Clinical Reasoning
Authors
Lang Cao; Yuhao Shen; Tianyang Luo; Simo Du; Hao Peng; Yue Guo
Year
2026
Publication status
P (arXiv preprint)
Version
arXiv v1, 2026-07-28
Source url
https://arxiv.org/html/2607.26160v1
DOI
10.48550/arXiv.2607.26160
Update target
Guideline-grounded executable ICD-10 scoring skills.
Fixed components
Backbone weights, skill compiler and generator remain fixed.
Feedback
Labelled training cases and recurring diagnostic rationale patterns.
Timing
Offline evolution before disjoint benchmark evaluation.
Retained artifact
Compiled Python scoring programs and diagnostic rules.
Update generator
Claude Sonnet 4.6 distils grouped case rationales.
Selection and release
Syntax and tier-contract checks; no clinical release gate reported.
Evaluation separate from updating
Four benchmarks; MedThinkBench excluded from evolution.
Fixed-procedure control
Skill-component ablations; no changing-improver versus frozen-improver control.
Evidence concerning improver change
Task skills change; update procedure stays fixed.
Clinical setting
Clinical diagnosis benchmarks.
Evidence boundary
No prospective care evaluation; external executable skill learning rather than parameter posttraining.
Locators
§3.3 skill evolution
§4 and Tables 1–3
Appendix H

69. Role-based norm agent

Key
ruan2026normcompliance
Name
Role-based norm agent
Title
A Self-Evolving LLM Agent Framework for Role-Based Norm Compliance in Healthcare
Authors
Haijie Ruan; Xiaowu Jiang; Zhanpeng Li; Wei Jia; Xuanwu Xu; Xiao-Fen Shan; Shujie Chen; Xindong Ye
Year
2026
Publication status
J/C (Findings of ACL 2026, pages 22562–22577)
Version
Official ACL Anthology conference PDF
Source url
https://aclanthology.org/2026.findings-acl.1133.pdf
DOI
10.18653/v1/2026.findings-acl.1133
Update target
Norm memory, system prompts and role identity.
Fixed components
Qwen-plus weights and feedback/update design.
Feedback
Norm trajectories and fixed AI preference feedback.
Timing
Cross-episode prompt edits and periodic identity updates.
Retained artifact
Structured norm memory and revised role prompts.
Update generator
Fixed diagnosis/reflection and identity-update routines.
Selection and release
Constraint checks described; no clinical deployment gate.
Evaluation separate from updating
Three synthetic healthcare communication tasks.
Fixed-procedure control
Component ablations; no isolated changing-improver control.
Evidence concerning improver change
Task role policies change; evaluator/update routine stays fixed.
Clinical setting
Simulated medication review, emergency triage and discharge handoff.
Evidence boundary
Judge count differs between Appendix A.1 and A.3 (5 versus 1); identity consistency is described as embedding distance and LLM scoring. No prospective patient evaluation.
Locators
§3.4–3.7; Algorithm 1
§4; Table 1
Appendix A.1 and A.3 reproducibility

70. MedGuard

Key
shi2026medguard
Name
MedGuard
Title
MedGuard: an LLM-based gatekeeper for detecting clinical risks in Chinese telemedicine consultations
Authors
Yuntao Shi; Qiwen Wang; Shang Gao; Jing Zhou; Jinzhen Lin; Zhenzhe Ying; Xiangxiang Zeng; Chen Lin
Year
2026
Publication status
J/C (npj Digital Medicine; accepted article published online 2026-08-10)
Version
Publisher Article in Press PDF
Source url
https://www.nature.com/articles/s41746-026-03116-0_reference.pdf
DOI
10.1038/s41746-026-03116-0
Update target
Shared Qwen2.5-7B agent policy weights.
Fixed components
Prompts, BM25 knowledge base, synthetic-data generator and GRPO routine.
Feedback
JSON-format and reference risk-label rewards after full agent trajectories.
Timing
Offline GRPO through Agent Lightning before separate consultation evaluation.
Retained artifact
Updated weights reused across different consultations.
Update generator
Fixed DeepSeek-R1 generates risk/no-risk training samples.
Selection and release
Rule-based terminal rewards; no prospective clinical release gate.
Evaluation separate from updating
Separate real/synthetic conversations and retrospective expert assessment.
Fixed-procedure control
Component ablations; no evolving-versus-fixed improver comparison.
Evidence concerning improver change
Policy changes; training-data generator and learning procedure stay fixed.
Clinical setting
Retrospective Chinese telemedicine risk checking.
Evidence boundary
Global memory is patient-specific and planner refinement is within-claim; eligibility rests on trajectory-based policy learning. Expert ratings are not prospective patient outcomes.
Locators
Methods; MedGuard Agent
Domain-Transfer Reinforcement learning, pp.24–26
Doctor Evaluation and retrospective validation, pp.12–15

71. SEMPA

Key
jiang2026sempa
Name
SEMPA
Title
Self-Evolving Dual-Role Deliberation for Medication Recommendation from Clinical Records
Authors
Jiyue Jiang; Yu Lei; Yitao Xu; Zikang Wang; Zhenyu Wang; Licheng Zong; Han Huang; Ziyi Li; Zhiyu Huo; Yuhan Chen; Jiayang Chen; Xiaodan Fan; Yu Li
Year
2026
Publication status
J/C (Findings of EMNLP 2026, pages 20985–21004)
Version
Official ACL Anthology conference PDF
Source url
https://aclanthology.org/2026.findings-emnlp.994.pdf
DOI
10.18653/2026.findings-emnlp.994
Update target
Episodic bank and Proposer/Critic natural-language policies.
Fixed components
GLM-5 weights, retrieval, pool partition and rewrite routine.
Feedback
Reference-drug overlap, DDI, molecular grounding and evolving Critic scores.
Timing
Periodic policy rewrites across encounters; policies freeze before held-out evaluation.
Retained artifact
Past episodes and appended role-policy principles.
Update generator
Fixed LLM meta-reflection contrasts success/failure pools.
Selection and release
Utility-based pools and semantic-anchor checks.
Evaluation separate from updating
Disjoint evolution stream and 100 held-out MIMIC-IV encounters.
Fixed-procedure control
Evolution-module ablation; changing Critic is not isolated.
Evidence concerning improver change
Critic affects later update selection; no tested revision of meta-reflection procedure.
Clinical setting
Retrospective medication recommendation.
Evidence boundary
Small single-centre test; no prospective intervention. Drug-class F1 and exact-name Jaccard use different matching definitions.
Locators
§3.4–3.5; Eqs.5–7
§4.1 evolve–freeze–test
Table 5
Appendix A.7 Algorithm 1

72. Skill-CDPO

Key
li2026skillcdpo
Name
Skill-CDPO
Title
Skill-CDPO: Evolving Agent Tool-Use via Critical Step Preference Optimization
Authors
Yuchen Li; Jinghao Lin; Yuanqing Wang
Year
2026
Publication status
J/C (Findings of EMNLP 2026, pages 42309–42326)
Version
Official ACL Anthology conference PDF
Source url
https://aclanthology.org/2026.findings-emnlp.1985.pdf
DOI
10.18653/2026.findings-emnlp.1985
Update target
Runtime-derived tool-use skills and DR-Tulu-8B weights.
Fixed components
Tools, expert model, critical-step selector and DPO procedure.
Feedback
Invocation failures and expert/local rollout outcome differences.
Timing
Offline skill construction, then single-round CDPO and held-out evaluation.
Retained artifact
Markdown skill cards, generated guards and trained weights.
Update generator
LLM summarises historical failure clusters; fixed expert supplies alternatives.
Selection and release
Rollout divergence selects preference steps; no clinical release gate.
Evaluation separate from updating
CURE-Bench, MedBrowseComp and PubMed Search with contamination filters.
Fixed-procedure control
Skill and preference-learning ablations; no evolving-improver control.
Evidence concerning improver change
Task skills/weights change; update and selection procedure stays fixed.
Clinical setting
Medical information retrieval and question-answering benchmarks.
Evidence boundary
Only single-round CDPO tested; retained experience-based skills establish eligibility independently. No patient-care evaluation.
Locators
§3.1
Table 1
Appendix A.1
Limitations; single-round training

73. Gated adapter / OpenClaw agent

Key
feng2026gatedopenclaw
Name
Gated adapter / OpenClaw agent
Title
Self-Evolving Diagnostic Framework based Gated Residual Adapters and OpenClaw-Based Medical Agents
Authors
Shuai Feng; Pan Su
Year
2026
Publication status
J/C (International Journal of Computer Science and Information Technology 8(4), pages 1–11)
Version
Publisher PDF, 2026
Source url
https://ijocsit.org/pdf/2026/10.62051-ijcsit.v8n4.01.pdf
DOI
10.62051/ijcsit.v8n4.01
Update target
Persistent clinical guidance and final adapter gating parameter.
Fixed components
RetFound/BioClinicalBERT towers, adapter MLPs and identity specification.
Feedback
Original reports simulate physician adjudication of conflicting cases.
Timing
Updates every 100 iterations across repeated test-set passes.
Retained artifact
Special-case MEMORY, distilled SOUL guidance and final gate.
Update generator
Periodic fixed case summarisation and selective gate optimisation.
Selection and release
Conflict cases trigger feedback; no independent clinical release gate.
Evaluation separate from updating
Private fundus-image data; repeated adaptation on test cases.
Fixed-procedure control
Adapter-capacity comparison; no matched evolving-improver control.
Evidence concerning improver change
Task guidance and gate change; update procedure remains fixed.
Clinical setting
Retrospective fundus diagnosis/report generation.
Evidence boundary
Agent LLM backbone unspecified. Evolution evaluated by conflict counts, with report-simulated feedback and reused test data, not independent diagnostic improvement or prospective care.
Locators
§2.3.1–2.3.3
§3.1; Table 1
§3.2; Figure 4

74. EMAI

Key
li2026emai
Name
EMAI
Title
Experience-guided multi-agent interpretable framework for radiology report summarization
Authors
Jia Li; Tong Zhou; Zichun Zhou; Xuan Wei; Hong Song; Zhixiang Wang; Yubo Chen; Han Lv
Year
2026
Publication status
J/C (Computer Methods and Programs in Biomedicine 273, 109078)
Version
Publisher HTML methods/results; first online 2025-09-29, issue January 2026
Source url
https://www.sciencedirect.com/science/article/abs/pii/S016926072500495X
DOI
10.1016/j.cmpb.2025.109078
Update target
Cluster-specific natural-language summarisation rules.
Fixed components
Backbone during rule evolution, clustering and selection routine.
Feedback
Automatic impression scores and difficult training reports.
Timing
Offline iterative evolution before different test reports.
Retained artifact
Selected experience rules and historical reports.
Update generator
LLM induces/refines rules using retained rules and difficult examples.
Selection and release
Discard lower-scoring half; stop at predefined count/performance thresholds.
Evaluation separate from updating
MIMIC-CXR and Open-I; separate training/test comparisons.
Fixed-procedure control
Experience/component ablations; no isolated evolving rule-generator control.
Evidence concerning improver change
Experience rules provide context for their own refinement; the induction/selection procedure stays fixed, as in the existing Level 1 notebook rule.
Clinical setting
Retrospective radiology report summarisation.
Evidence boundary
Text findings-to-impression task; no image interpretation or prospective care. Contextual refinement of task rules does not establish a changed improver.
Locators
§2.2 Self-evolutive nearest-neighbor explicit experience induction
§2.4 Agentic framework
§4.2–4.3; Table 3; Figure 4
§4.5 manual evaluation

75. BridgeGap

Key
li2026bridgegap
Name
BridgeGap
Title
BridgeGap: Belief-Grounded Training for Medical Dialogue Agents
Authors
Weijun Li; Tao Yang; Jiayu Yang; Fei Chao
Year
2026
Publication status
J/C (EMNLP 2026, pages 44932–44947)
Version
Official ACL Anthology conference PDF
Source url
https://aclanthology.org/2026.emnlp-main.1974.pdf
DOI
10.18653/2026.emnlp-main.1974
Update target
Qwen3-8B diagnostic dialogue policy weights.
Fixed components
Patient simulator, profile verifier and reward constants.
Feedback
Environment-verified belief progress and terminal diagnostic correctness.
Timing
SFT followed by online simulated dialogue PPO across profiles.
Retained artifact
Trained policy weights; within-case belief states are temporary.
Update generator
Frozen Qwen2.5-72B patient; fixed rule-based reward computation.
Selection and release
Profile-grounded NLI verifies symptom evidence; no clinical release gate.
Evaluation separate from updating
HealthBench, AgentClinic and static medical QA.
Fixed-procedure control
Training/reward ablations; no changing-improver comparison.
Evidence concerning improver change
Task policy changes; simulator, reward and learning procedure stay fixed.
Clinical setting
Simulated diagnostic consultation.
Evidence boundary
Belief-state persistence alone is within-case; eligibility rests on cross-episode policy learning. No prospective patient-care evaluation.
Locators
§3.1–3.2
§3.4.2 online belief-progress RL
§4.1 and Appendix C PPO details

76. SeM-Agents

Key
chen2025semagents
Name
SeM-Agents
Title
A Self-Evolving Framework for Multi-Agent Medical Consultation Based on Large Language Models
Authors
Kai Chen; Ji Qi; Jing Huo; Pinzhuo Tian; Fanyu Meng; Xi Yang; Yang Gao
Year
2025
Publication status
J/C (ICASSP 2025)
Version
IEEE Xplore full-text HTML, accessed through existing Tohoku University entitlement
Source url
https://ieeexplore.ieee.org/document/10889517
DOI
10.1109/ICASSP49660.2025.10889517
Update target
Correct-case and error-reflection experience databases.
Fixed components
GPT-4-turbo weights, residual consultation protocol and reviewer prompts.
Feedback
Reference correctness routes successful cases or error abstractions.
Timing
Offline training consultations before test-set evaluation.
Retained artifact
CorrectKB cases and ChainKB structured error analyses.
Update generator
Fixed Chain-of-Thought Reviewer abstracts incorrect consultations.
Selection and release
Reference correctness gates storage; safety/ethics reviewer screens answers.
Evaluation separate from updating
MedQA/PubMedQA test sets; experience banks contain training-set cases only.
Fixed-procedure control
Memory/discussion variants; no changing-improver control.
Evidence concerning improver change
Memory grows; abstraction and retrieval procedures stay fixed.
Clinical setting
Offline clinical and biomedical QA.
Evidence boundary
No prospective care. Comparisons mix MedQA option counts; role names do not establish image use. Distinct earlier ICASSP paper, not the retitled ACL MDTeamGPT record.
Locators
§II.C Summary and output; Algorithm 2
§III.A Datasets
§III.B–D; Tables I–II; Figure 2

77. Evolved multi-image decision rules

Key
li2026decisionrules
Name
Evolved multi-image decision rules
Title
Inference-Time Agentic Decision Rules Beat Longer Evolving Search for Multi-Image Medical Reasoning
Authors
Site Li; Jianyi Hao; Xiaofeng Liu
Year
2026
Publication status
P (arXiv; CVPR workshop presentation stated in manuscript, acceptance not independently verified)
Version
arXiv v1, 2026-07-30
Source url
https://arxiv.org/abs/2607.27564v1
DOI
10.48550/arXiv.2607.27564
Update target
Strategy-specific prompt-program configuration for multi-image VQA.
Fixed components
Qwen3-Coder-Next proposer, Qwen3.5-35B-A3B task VLM, ShinkaEvolve archive/selection and evaluation.
Feedback
Clinical VQA accuracy on evolution cases, with larger-pool periodic reevaluation.
Timing
50 offline evolution generations before held-out generation selection and final testing.
Retained artifact
Selected inference policy/configuration reused on different medical VQA cases.
Update generator
Fixed Qwen3-Coder-Next patch proposals through ShinkaEvolve.
Selection and release
Archive/island selection, periodic evolution-pool rescoring and a separate 665-case holdout.
Evaluation separate from updating
855 final-test examples, grouped by video ID apart from 1331 evolution and 665 selection cases; five independent main-method repeats.
Fixed-procedure control
Common search engine and generation budget; no comparison that revises the improvement procedure.
Evidence concerning improver change
Task inference configurations change; proposal, archive and selection pipeline remain fixed.
Clinical setting
MedFrameQA multi-image medical VQA.
Evidence boundary
Internal partition of the public benchmark, not its official hidden test; search-mini feedback differs across strategies (256 versus128). No prospective clinical evaluation.
Locators
§3 and Table1, frozen split
§4 and Figure4, mutable configuration
§5 and Table3, models/budget/repeated results

78. Shennong-Agent

Key
dou2024shennongagent
Name
Shennong-Agent
Title
Autonomous Pharmaceutical Care with Large Language Models
Authors
Yutao Dou; Zike Deng; Tao Xing; Jian Xiao; Shaoliang Peng
Year
2024
Publication status
J/C (IEEE BIBM 2024; added to IEEE Xplore 2025-01-10)
Version
Published IEEE HTML, accessed through existing Tohoku University entitlement
Source url
https://ieeexplore.ieee.org/document/10822302
DOI
10.1109/BIBM62325.2024.10822302
Update target
Experience vector database containing clinician corrections and prompts.
Fixed components
Qwen-7B, five prompted clinical modules, retrieval/tool interfaces and update routine.
Feedback
Doctors add prompts or edit LLM inputs/outputs; all edits and prompts are re-encoded into the vector database.
Timing
Feedback retained after human interventions for later similar clinical situations.
Retained artifact
Indexed clinician corrections, prompts and case-derived experience.
Update generator
Fixed recording, encoding and retrieval procedure.
Selection and release
Clinician intervention; no independent update-release threshold reported.
Evaluation separate from updating
461 retrospectively collected elderly outpatient prescriptions/clinical records from one hospital; two pharmacist raters and a third for disagreement. No separate adaptation/test split reported.
Fixed-procedure control
With/without HITL comparison; no isolated persistent-memory or changing-improver control.
Evidence concerning improver change
Task experience changes; update procedure and model are not shown to change.
Clinical setting
Polypharmacy review, interaction checks, medication schedules and patient education.
Evidence boundary
Abstract calls the mechanism RLHF, but methods document clinician-edit vector-memory reuse rather than parameter RL. HITL also includes current-case corrections, so its comparison does not isolate cross-case learning. Clinical data are retrospective, and deployment effectiveness is unvalidated.
Locators
§III opening paragraphs, correction/prompt recording and similar-situation reuse
§III.B vector database; §IV.C Qwen-7B
§IV.A–B retrospective cohort and raters
§V.B HITL comparison; §VI limitations

79. Living-OEP

Key
rugo2026livingoep
Name
Living-OEP
Title
Development and validation of a human-supervised AI-augmented living oncology evidence platform: a breast cancer pilot study
Authors
H. Rugo; Anna Forsythe; D. Flora; S. Glück; S. Grieve; R. Campden; R. J. Liu; M. Musat; J. M. Rege; P. A. Kaufman; R. L. Coleman; P. Tarantino; J. O’Shaughnessy; L. Schwartzberg
Year
2026
Publication status
J/C (ESMO Real World Data and Digital Oncology 13:100770)
Version
Published 2026-09-01; full PMC article
Source url
https://pmc.ncbi.nlm.nih.gov/articles/PMC13563553/
DOI
10.1016/j.esmorw.2026.100770
Update target
Annotation manual, prompt templates and few-shot examples.
Fixed components
GPT-4.1/o3/Claude Sonnet-4 and human-supervised update procedure.
Feedback
Human corrections of earlier abstract annotations.
Timing
Repeated development iterations before locked, newly arriving validation publications.
Retained artifact
Revised instructions and validated annotation examples.
Update generator
Human scientists revise instructions and agent components.
Selection and release
Variable-specific accuracy targets; unsuccessful variables remain manually annotated.
Evaluation separate from updating
1997 later breast-cancer publications excluded from training; eight clinical evidence scenarios.
Fixed-procedure control
Commercial chatbot comparison; no isolated persistent-update or changing-improver control.
Evidence concerning improver change
Task annotation guidance changes; no revised updater tested.
Clinical setting
Oncology evidence annotation supporting treatment evidence retrieval.
Evidence boundary
Daily literature refresh alone is not the qualifying mechanism. Eligibility rests on feedback-driven instruction refinement. All outputs receive human review; trial-level interpretation remains human. Clinical patient outcomes are untested. Published author initials retained where full names are unverified.
Locators
Materials and methods: agentic AI construction and annotation process
Validation of the agentic AI system: independent locked publications
Results: iterative extraction refinement; chatbot-comparison limitations

80. U-EHR

Key
wu2026uehr
Name
U-EHR
Title
U-EHR: A Self-Evolving EHR Agent with Step-Level Credit Assignment and UCB-Guided Memory Retrieval
Authors
Hao Wu; Zihan Wang; Ziyang Rao; Heyi Lin; Jinjing Zhu; Qianyi Cai; Yi Zhou; An Lin; Hao Wang; Hui Xiong
Year
2026
Publication status
J/C (KDD 2026 V.2, pp12420–12431)
Version
Published 2026-08-08; online publisher PDF read through existing browser
Source url
https://dl.acm.org/doi/pdf/10.1145/3770855.3819038
DOI
10.1145/3770855.3819038
Update target
Positive/negative experience bank, memory quality estimates and retrieval statistics.
Fixed components
Backbone weights, step-scoring/extraction prompts, three-score scheme and UCB update rules.
Feedback
Relevance, progress and correctness scores for prior actions; completed-task success updates memory retrieval counts.
Timing
After each clinical task, memories are extracted and statistics updated before the next task.
Retained artifact
Task summaries, effective action patterns, errors/reflections and historical retrieval/success counts.
Update generator
Fixed LLM scoring, extraction and memory consolidation procedures.
Selection and release
Positive/negative step thresholds 0.7/0.3; UCB retrieval uses fixed quality weights and exploration coefficient.
Evaluation separate from updating
AgentEHR-Bench/MIMIC-IV and private CN-EHR with four backbones; all tasks are processed sequentially with memory from earlier tasks, without separate held-out memory learning.
Fixed-procedure control
Ablations remove UCB, step scoring and negative memory; no changing-improver control.
Evidence concerning improver change
Task memory and retrieval estimates change under fixed scoring/extraction/UCB rules.
Clinical setting
Six EHR prediction tasks and Chinese diagnosis coding.
Evidence boundary
Same-stream sequential adaptation, not an independent post-learning holdout. CN-EHR is externally sourced but also evaluated sequentially. No prospective patient outcomes; memory-on/off and component ablations do not isolate an evolving updater.
Locators
§3 and Figure2, retained bank and next-task reuse
§4.1–4.3, scoring, extraction and UCB
§4.4, Tables1–3
§5.2–5.4 and Table4
AppendixA, sequential task order and fixed prompts

81. Coach Iris / FAST

Key
neary2025fast
Name
Coach Iris / FAST
Title
Think FAST: a novel framework to evaluate fidelity, accuracy, safety, and tone in conversational AI health coach dialogues
Authors
Martha Neary; Emily Fulton; Victoria Rogers; Julia Wilson; Zoe Griffiths; Ram Chuttani; Paul M. Sacher
Year
2025
Publication status
J/C (Frontiers in Digital Health 7:1460236)
Version
published18Jun2025
Source url
https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/full
DOI
10.3389/fdgth.2025.1460236
Update target
Shared agent prompts, response parameters and retrieval data; model refinement is described.
Fixed components
Backbone and optimizer Not reported.
Feedback
Human FAST ratings and SME reviews.
Timing
Continuous monitoring; revision cadence Not reported.
Retained artifact
Adjusted prompts/RAG/settings; fine-tuned model described.
Update generator
Human reviewers, SMEs and ML team.
Selection and release
Unacceptable ratings trigger refinement; release gate Not reported.
Evaluation separate from updating
Five reviewers initially assessed 12 recorded dialogues; the article reports framework development/face validity and ongoing service monitoring, without an independent post-update learning evaluation.
Fixed-procedure control
Not reported.
Evidence concerning improver change
Human rating rubric was revised; no effect of those revisions on subsequent agent learning was tested.
Clinical setting
Retrospective FAST assessment of recorded health-coaching dialogues from an operating patient service; deployment status is documented separately from the evaluation design.
Evidence boundary
The service has active patient users, but this article evaluates a human review framework on recorded dialogues. It does not report a prospective trial of agent adaptation or quantify post-update patient/learning benefits. Human-led model/rubric refinements are distinguished from autonomous self-improvement.
Locators
§2.4 SME-directed refinements; PDFp4
§2.5 updates from monitoring
§3.1.1 unacceptable feedback to ML fine-tuning
§3.4 rubric changes
§4 face validity and absent measure validation

82. MedEnv / Diagnosis-R1

Key
fan2026medenv
Name
MedEnv / Diagnosis-R1
Title
MedEnv: Scaling Multimodal Virtual Medical Environments for Long-Horizon Diagnosis
Authors
Zhiting Fan; Xuxiang Ren; Yuan Wang; Ruizhe Chen; Zijie Meng; Keli Hu; Jiaxiang Liu; Jian Wu; Weiqi Zhai; Hongxia Xu; Zuozhu Liu
Year
2026
Publication status
J/C (MICCAI 2026, LNCS16886, pp420–429)
Version
Public accepted10-page paper; society HTML abstract still calls environment DiagWorld and500K+, final PDF/ Springer useMedEnv496K+.
Source url
https://papers.miccai.org/miccai-2026/paper/5084_paper.pdf
DOI
10.1007/978-3-032-38098-2_40
Update target
Doctor-policy parameters.
Fixed components
Simulator/tool interfaces, disease KB and GRPO/reward formula.
Feedback
Terminal diagnosis correctness plus confidence gain toward correct diagnosis; unsupported-action penalty.
Timing
SFT then stochastic interactive RL rollouts and parameter updates.
Retained artifact
Trained Diagnosis-R1 checkpoint.
Update generator
GRPO under rllm; SFT policy filters cases using8 rollouts, discarding uniformly correct/incorrect cases.
Selection and release
Case filtering described; checkpoint acceptance/release criterion absent.
Evaluation separate from updating
2000 heldout cases separate from RL, per author response; patient-level disjointness and external validation unspecified.
Fixed-procedure control
Table3 matchedSFT, outcome-onlyRL and full-reward ablations; same initialization/data/steps/hyperparameters stated.
Evidence concerning improver change
The process reward uses policy-model diagnostic likelihood, assigning the trained component a scoring role in later policy updates. Whether scoring parameters are synchronized after every update is not specified; the reward formula and GRPO procedure stay fixed.
Clinical setting
MIMIC-grounded virtual consultation.
Evidence boundary
Simulated clinical interactions; no prospective deployment. Author response states 2000 held-out cases differ from RL cases, but patient-level separation is unspecified. Policy likelihood is specified as process reward, not ordinary action-logprob loss; scoring-parameter synchronization and a frozen-scoring-model control are absent.
Locators
§2.1, clinical simulators/tools
§2.2, interactive GRPO and policy-likelihood process reward
§3.1/3.3, training and Table3
Author Feedback, held-out-case separation

83. SR-MAPR

Key
mao2026srmapr
Name
SR-MAPR
Title
A Self-Reflective Multi-Agent Physician Recommendation Framework
Authors
Shuai Mao; Yuqing Yang
Year
2026
Publication status
J/C (IEEE CNML2026)
Version
Published IEEE HTML; conference 2026-01-30–02-01; added2026-03-31
Source url
https://ieeexplore.ieee.org/document/11452303
DOI
10.1109/CNML68938.2026.11452303
Update target
Heuristic rule memory for disease-specific physician assessment.
Fixed components
Specialist/Summary/Reflection prompts, backbone options, clustering, retrieval and expert reference labels.
Feedback
Summary predictions that disagree with expert gold ratings trigger error analysis.
Timing
Error-derived rules are added for future physician-profile evaluations; update/test scheduling not detailed.
Retained artifact
Rule-based memory alongside labelled case memory.
Update generator
Fixed Reflection Agent produces generalized heuristics from rating errors.
Selection and release
New rules stored in memory; acceptance/release criteria not specified.
Evaluation separate from updating
7362 physician profiles with stated80/20 split, four Summary backbones, ranking/classification metrics and component ablations.
Fixed-procedure control
Rules/case/RAG removal ablations; no fixed-versus-changing improver comparison.
Evidence concerning improver change
Task assessment heuristics change; criteria-generation and reflection procedures stay fixed.
Clinical setting
Disease-specific patient-to-physician recommendation.
Evidence boundary
Professional-profile benchmark, not patient outcomes. How rule creation is partitioned from test ratings is not reported. Evaluation heuristics serve the physician-ranking task, not the procedure that evaluates later agent updates.
Locators
§III.A dataset and split
§III.C.3–4 dual memory and self-reflective rule storage
§IV.A fixed backbone/retrieval implementation
§V.C and TableIII ablations

84. MDTace

Key
chen2027mdtace
Name
MDTace
Title
MDTace: Agentic Context Engineering for Multi-disciplinary Team Medical Consultation
Authors
Kai Chen; Yang Gao
Year
2027
Publication status
J/C (ICIC2026, LNCS16649, pp81–92; online2026, imprint2027)
Version
Publisher full HTML; first online2026-07-19; citation/imprint2027
Source url
https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6
DOI
10.1007/978-981-92-3403-5_6
Update target
Global Bullet Playbook of structured consultation experience.
Fixed components
GPT-4o/backbone variants, specialist selection, lead/reflector prompts and update/retrieval procedure.
Feedback
Completed training-case deliberations, disagreements and tool-use feedback are distilled and filtered by the Reflector.
Timing
Bullets accumulated from training cases, then retrieved on separate test cases.
Retained artifact
Up to600 recent training-case bullets; cosine-similar entries can be replaced.
Update generator
Fixed Lead Physician aggregation and Reflector filtering.
Selection and release
High-value reusable bullets retained; capacity/similarity replacement rule and top5 retrieval fixed.
Evaluation separate from updating
MedQA, PubMedQA, VQA-RAD and SLAKE test sets; all stored bullets explicitly training-derived.
Fixed-procedure control
Primary-doctor, context-layer and Playbook removal ablations; no changing-improver control.
Evidence concerning improver change
Task memory changes and conditions subsequent consultation summaries; memory-construction procedure remains fixed.
Clinical setting
Textual and multimodal multidisciplinary diagnostic QA.
Evidence boundary
Benchmark evidence only. Context-length stability and Playbook ablations do not establish an improved updater. SeM-Agents and MDTeamGPT are separately cited predecessors, not publication aliases. Imprint2027 follows first online2026 within cutoff.
Locators
§3.2, Reflector-to-Playbook memory write
§3.3, cross-case bullet aggregation/filtering
§4.1, training-only memory/test evaluation
§4.2, capacity/retrieval, results and ablations
§2.1, distinct SeM-Agents/MDTeamGPT predecessors

85. RHIS

Key
chen2026rhis
Name
RHIS
Title
From API Acknowledgment to Clinical State: Runtime Governance, Controlled Self-Evolution, and Cost in MedAgentBench
Authors
Wenjie Chen
Year
2026
Publication status
P (SSRN working paper)
Version
SSRN Working Paper Version0.1, September2026, posted2026-09-21;13-page online PDF
Source url
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519
DOI
10.2139/ssrn.7482519
Update target
Versioned external action-governance controller rules, payload adapters and terminal-obligation checks.
Fixed components
Kimi kimi-k3 with high reasoning effort; frozen MedAgentBench source and isolated official scorer; declared change-admission record.
Feedback
Visible execution failures, internal runtime-calibration signals or authorized external instructions trigger proposed controller revisions.
Timing
Successor controller versions are released between frozen runs and reused across later benchmark stages; same-stage recoveries and full retests are reported separately.
Retained artifact
Controller, self-evolver, configuration templates and versioned rules; v0.3.25 package reports128 passing unit/regression tests.
Update generator
Self-evolver generates versioned rule changes under a fixed protocol; detailed implementation of its proposal procedure is not reported.
Selection and release
Change record requires frozen trigger evidence, root-cause account, proposed rule, adjudication, successor version, rollback target, regression result and independent-evaluation boundary.
Evaluation separate from updating
300 virtual-EHR tasks in11 first-pass stages; controller versions change across stages. Only named same-stage retests support version contrasts; no separate held-out learning-transfer test.
Fixed-procedure control
Named predecessor/successor controller runs, without a matched frozen-improver control; aggregate across versions is explicitly descriptive.
Evidence concerning improver change
The external controller changes task action admission. No revision and later-learning test of the change-proposal/adjudication procedure itself is demonstrated.
Clinical setting
Virtual-EHR task execution with FHIR-oriented tools, including medication and service writes.
Evidence boundary
Single-creator preprint; same-stage retests, missing scorer outcomes and candidate governance labels are disclosed. Version changes across different stages do not isolate causal gains; no clinical care was delivered.
Locators
PDFp3, §3.1 benchmark source and300-task stages
PDFp4, §3.2–3.4 controller and controlled self-evolution
PDFpp5–6, §5 first-pass versus named post-evolution runs
PDFp7, §6 costs and state-closure boundary
PDFpp10–11, §10 limitations