85-study design matrix

Resource version0.3.0; search cutoff4November2026. Two reporting-quality exclusions approved5November2026. Author verification remains pending.

keynamemechanism_groupmechanism_group_labelevidence_levelclinical_evaluation_categoryupdate_targetfixed_componentsretained_artifactfeedbackupdate_timingupdate_generatorselection_and_releasecontrol_descriptionindependent_test_descriptionmatched_generator_control_in_reviewimprover_change_descriptionevidence_boundarysource_versionsource_urlsource_locatorsclassification_sourceauthor_checkauthor_comment
medhospital2024Agent Hospital / MedAgent-ZeromemoryMemory and experience across encounters1simulatedMedical case database and natural-language experience principles.LLM weights; prompting, retrieval and reflection/validation design.Successful case-answer pairs and validated/refined principles.Synthetic patient reference decisions; correctness on medical-document exemplar questions.Across simulated patient encounters; no parameter training.Frozen doctor LLM reflects on incorrect decisions.Correct cases retained; principles tested and rejected when incorrect; exemplar validation described.No matched changing-versus-fixed improver comparison reported.Separate simulated test cases and MedQA; memory-component ablations.not_identifiedMemory contents change; reflection/validation algorithm not reported as evolving.The paper calls MedQA real-world alignment; this is not a prospective patient-care study.v3https://arxiv.org/abs/2405.02957v3[{'url': 'https://arxiv.org/html/2405.02957v3#Sx3', 'location': 'Agent Evolution'}, {'url': 'https://arxiv.org/html/2405.02957v3#Sx12.SSx3', 'location': 'Appendix B.3: reflection, validation and refinement'}, {'url': 'https://arxiv.org/html/2405.02957v3#Sx13.SSx4', 'location': 'Appendix C.4: separate simulated training/test cases'}, {'url': 'https://arxiv.org/html/2405.02957v3#Sx13.SSx7', 'location': 'Appendix C.7: ablations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
clinicalagents2026ClinicalAgentscomparatorsDiagnostic training and informative adaptation comparators0offline_benchmarkCurrent-patient evidence, hypotheses, action trajectory and search state.Static guideline/case experience store; orchestration equations and prompts.Working state during consultation; no newly accumulated cross-case skill reported.Retrieved evidence, missing-information checks, hypothesis confidence and search rewards.Within a case; multi-turn inference.LLM orchestrator with MCTS-style selection and backtracking.Inference actions selected by search; no learned-artifact release process described.Orchestration/memory ablations; no changing-improver control.Test cases excluded from experience store; separate-model judge validated against clinicians.not_identifiedState/search changes, not a reported learning-procedure update.Static experience retrieval does not establish newly learned cross-case memory or skill synthesis.v2https://arxiv.org/abs/2603.26182v2[{'url': 'https://arxiv.org/html/2603.26182v2#S3.SS3.SSS2', 'location': 'Section 3.3.2: explicitly static experience memory'}, {'url': 'https://arxiv.org/html/2603.26182v2#S3.SS3.SSS1', 'location': 'Section 3.3.1: working memory'}, {'url': 'https://arxiv.org/html/2603.26182v2#S4.SS1.SSS1', 'location': 'Section 4.1.1: test exclusion'}, {'url': 'https://arxiv.org/html/2603.26182v2#S4.SS2', 'location': 'Section 4.2: clinician validation of judge'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
coevolving2025TissueLabgeneratorsParameter updates and changing training generators1retrospectiveEditable stored outputs; lightweight tool classifiers; model-ranking preferences.Foundation models and orchestration framework; human-designed feedback mechanisms.HDF5 outputs/annotations/logs, adapted classifiers, model preferences and contributed tools.Clinician annotations and accept/reject feedback.Interactive refinement and reuse in subsequent executions; community contributions.LLM workflow/code generation; clinician-guided active learning and model reselection.Human output review and feedback guide switching; uniform held-out release gate not specified.No isolated changing-versus-fixed improver comparison reported.Task benchmarks; separate pathology hold-out and patient-level fatty-liver split described.not_identifiedRanking preferences adapt; later-learning benefit of changing the improvement procedure is not isolated.Human-guided tool adaptation is demonstrated; autonomous experience-to-skill synthesis is not clearly specified.v1https://arxiv.org/abs/2509.20279[{'url': 'https://arxiv.org/html/2509.20279v1#S4.SSx3', 'location': 'Methods: Community-driven co-evolution'}, {'url': 'https://arxiv.org/html/2509.20279v1#S4.SSx6', 'location': 'Methods: Editable memory layer'}, {'url': 'https://arxiv.org/html/2509.20279v1#S2.SS5', 'location': 'Section 2.5: user-created/adapted tools'}, {'url': 'https://arxiv.org/html/2509.20279v1#S4.SSx9', 'location': 'Methods: benchmarking and separation'}, {'url': 'https://arxiv.org/html/2509.20279v1#S3', 'location': 'Discussion: human supervision and clinical integration limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
fan2026macroMACROgeneratorsParameter updates and changing training generators2-implicitretrospectiveExperience memory, composite-tool registry and agent policy parameters.Frozen visual encoder; mining criterion and SFT/GRPO update rules.Successful trajectory fragments, reusable tool sequences and LoRA policy.Training-label correctness gates memory; composite-tool invocation yields RL reward.Across training episodes, during supervised warm-start and GRPO.Trajectory extraction, frequent-subsequence mining and policy optimization.Correct episodes update memory; frequency threshold admits composites; no separate clinical release gate specified.Component ablations share hyperparameters; no isolated improver-change control.Imaging benchmarks and component ablations; RAM-W600 test-set results.not_identifiedComposite registry changes actions/reward eligibility; mining/update rules stay specified; improver gain not isolated.Training-time co-evolution does not establish ongoing clinical deployment adaptation.v1https://arxiv.org/abs/2603.05860[{'url': 'https://arxiv.org/html/2603.05860v1#S3.SS2', 'location': 'Section 3.2: correct training episodes enter memory'}, {'url': 'https://arxiv.org/html/2603.05860v1#S3.SS3', 'location': 'Section 3.3: composite registration'}, {'url': 'https://arxiv.org/html/2603.05860v1#S3.SS4', 'location': 'Section 3.4: SFT/GRPO and reward'}, {'url': 'https://arxiv.org/html/2603.05860v1#S4.SS3', 'location': 'Section 4.3: ablations'}, {'url': 'https://arxiv.org/html/2603.05860v1#S5', 'location': 'Section 5: prospective studies remain future work'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
qiu2025diaggymDiagAgent / DiagGymcomparatorsDiagnostic training and informative adaptation comparators1simulatedDiagnostic agent parameters and examination-selection policy.Previously trained DiagGym environment, reward specification and GRPO protocol.Trained diagnostic-agent weights.Simulated examination results; diagnosis match, examination alignment and turn-budget rewards.Offline cold-start training followed by interactive simulated RL.SFT then GRPO using generated diagnostic trajectories.Reward-based training updates; separate deployment release gate not described.SFT/reward ablations; no changing-versus-fixed improver comparison.Held-out and out-of-domain DiagBench cases; physician-written rubrics and judge validation.not_identifiedDiagnostic policy learns; reward/update-generator changes are not evaluated.Physician validation and clinical source records do not establish prospective clinical utility.v2https://arxiv.org/abs/2510.24654v2[{'url': 'https://arxiv.org/html/2510.24654v2#S5.SS2', 'location': 'Section 5.2: training and fixed rewards'}, {'url': 'https://arxiv.org/html/2510.24654v2#S5.SS3', 'location': 'Section 5.3: benchmark construction'}, {'url': 'https://arxiv.org/html/2510.24654v2#A1.SS1.SSS3', 'location': 'Appendix A.1.3: physician validation of metrics'}, {'url': 'https://arxiv.org/html/2510.24654v2#S3.SS3', 'location': 'Section 3.3: ablations'}, {'url': 'https://arxiv.org/html/2510.24654v2#S4', 'location': 'Section 4: virtual environment limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
feng2025doctoragentrlDoctorAgent-RLcomparatorsDiagnostic training and informative adaptation comparators1simulatedDoctor-agent questioning/diagnosis policy parameters.Patient simulator, consultation reward components and prescribed training algorithm.Fine-tuned doctor-agent weights.Simulated patient responses; diagnosis/treatment word-F1, information gain and protocol rewards.Offline SFT followed by simulated multi-turn GRPO training.Behavior cloning and GRPO under sampled dialogue-turn budgets.Training objectives update policy; separate clinical release gate not reported.SFT/RL/turn-budget ablations; no evolving-improver control.Held-out MTMedDialog; model judge and blinded human assessment of sampled dialogues.not_identifiedQuestioning strategy learns; training/evaluation procedure not reported as self-modified.Do not attribute extended v4 clinical experiments to the conference paper.ICASSP 2026 metadata; mechanism extraction from matching-title/five-author arXiv v3 (2025-10-14), not extended v4https://arxiv.org/html/2505.19630v3[{'url': 'https://www.cmsworkshops.com/ICASSP2026/view_paper.php?PaperNum=6520', 'location': 'Official conference title/authors'}, {'url': 'https://arxiv.org/html/2505.19630v3#Sx3.SSx3', 'location': 'Training Procedure'}, {'url': 'https://arxiv.org/html/2505.19630v3#Sx3.SSx2.SSSx3', 'location': 'Consultation Evaluator'}, {'url': 'https://arxiv.org/html/2505.19630v3#Sx4.SSx4', 'location': 'Blinded human evaluation'}, {'url': 'https://arxiv.org/html/2505.19630v3#Sx4.SSx5', 'location': 'Training-component ablations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
xu2026oaimsOAIMScomparatorsDiagnostic training and informative adaptation comparators1retrospectiveInteractive segmentation model parameters.Architecture, adaptation losses and prescribed post-/mid-interaction update rules.Adapted segmentation weights.Correction clicks and click-refined pseudo-label masks.After each click and/or image; updated weights persist across image sequence.Gradient updates with Dice-Focal and click-centered Gaussian losses.Direct prescribed online updates; no separate release-validation cohort described.Frozen and other fixed adaptation baselines; no changing-improver comparison.Ground-truth segmentation metrics on shifted target databases; baseline/ablation/robustness comparisons.not_identifiedModel adapts; update rule is not reported as self-modified.Non-LLM segmentation comparator; simulated clicks are not prospective clinician-use evidence.v4https://arxiv.org/abs/2503.06717v4[{'url': 'https://arxiv.org/html/2503.06717v4#S2.SS4', 'location': 'Section 2.4: update timing'}, {'url': 'https://arxiv.org/html/2503.06717v4#S3.SS2', 'location': 'Section 3.2: online evaluation'}, {'url': 'https://arxiv.org/html/2503.06717v4#S3.SS3', 'location': 'Section 3.3: robustness'}, {'url': 'https://arxiv.org/html/2503.06717v4#A1.SS2', 'location': 'Appendix A.2: clicks simulated using ground truth'}, {'url': 'https://proceedings.iclr.cc/paper_files/paper/2026/hash/8e5e1e120afdbd8700ee4ce8c42f2bf5-Abstract-Conference.html', 'location': 'Official ICLR publication'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wu2026medrsiMedRSIskillsSkills, tools and workflow adaptation2-implicitexternal_retrospectiveTool registry, generated code, trained tool models and derived training resources.Reasoning model, execution rules, judge/rubric and registration protocol.Registered tool packages, dependencies, checkpoints and evaluation histories.Diagnostic errors weighted by clinical consequence; paired candidate evaluations.Successive retrospective patient batches and candidate-trial cohorts.Failure clustering/reflection; code generation, composition and model training.Discovery screening then frozen-candidate trials; sustained benefit without increased clinical cost.Matched clinical-cost/registration ablations; no isolated evolving-improver control.Disjoint patient partitions; locked test evaluation after trajectory freezing.not_identifiedNew capabilities enable later development; improvement-procedure change/gain not independently isolated.Broad recursive capability growth; prospective workflow/patient-outcome validation remains future work.v1 (2026-09-21)https://arxiv.org/html/2609.24838v1[{'url': 'https://arxiv.org/html/2609.24838v1#S2.SS1', 'location': 'Section 2.1: fixed M/P and changing registry'}, {'url': 'https://arxiv.org/html/2609.24838v1#S2.SS2', 'location': 'Section 2.2: frozen clinical evaluator'}, {'url': 'https://arxiv.org/html/2609.24838v1#S2.SS3', 'location': 'Section 2.3: discovery and registration'}, {'url': 'https://arxiv.org/html/2609.24838v1#S3.SS1', 'location': 'Section 3.1: separation and retrospective checkpoint testing'}, {'url': 'https://arxiv.org/html/2609.24838v1#S3.T6', 'location': 'Table 6: matched mechanism ablations'}, {'url': 'https://arxiv.org/html/2609.24838v1#S4', 'location': 'Section 4: prospective validation needed'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
lan2024amcAgent Mental Clinic (AMC)memoryMemory and experience across encounters1simulatedDiagnostic skills, encounter/EMR memory, and outcome-dependent memory-importance scores.Frozen GPT-3.5; constant dialogue, diagnosis and skill-reflection prompts; fixed supervisor stages and relevance/importance sampling formula.Conversation records, summarized EMRs, reusable diagnostic skills, and memory-importance scores.Professional D4 risk labels, revealed after diagnosis, guide supervisor reflection; diagnostic correctness increments or decrements retrieved-memory importance.After simulated training consultations (Quiz); accumulated memory is reused on subsequent patients and in the single-attempt Exam setting.LLM supervisor compares predicted and reference depression/suicide risk and produces corrective skills; a fixed rule updates importance.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.Memory-off, memory-layer and supervisor ablations test component value; they do not compare an evolved skill-generation procedure with its fixed predecessor.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.not_identifiednot established: skills and retrieval priorities change under constant reflection prompts and fixed importance-update/sampling rules.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.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)https://arxiv.org/html/2409.15084v2[{'url': 'https://arxiv.org/html/2409.15084v2#Sx2', 'location': 'patient cohort, professional labels and supervisor role'}, {'url': 'https://arxiv.org/html/2409.15084v2#Sx3', 'location': 'six-stage loop; constant-prompt Eqs. 1–6; tertiary memory; retrieval Eq. 7'}, {'url': 'https://arxiv.org/html/2409.15084v2#Sx4', 'location': 'Quiz/Exam distinction, Table 1, memory and supervisor ablations'}, {'url': 'https://arxiv.org/html/2409.15084v2#Sx6', 'location': 'role-play, cultural coverage and absent real-time expert feedback'}, {'url': 'https://arxiv.org/html/2409.15084v2#A3', 'location': 'intended-use and privacy boundaries'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
almansoori2025medagentsimMedAgentSimmemoryMemory and experience across encounters1simulatedMedical-record and reflection-experience buffers used for later diagnostic retrieval.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.CLIP-embedded successful case records and reflection insights from initially incorrect cases corrected on a second attempt.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.After each simulated consultation; stored records are retrieved during later consultations before ensemble diagnosis.The doctor reflects on an incorrect diagnosis and retries; the record-store module embeds and retains eligible results.Correct initial diagnoses enter the medical buffer; only successful corrected diagnoses contribute reflection insights to the experience buffer; unsuccessful retries are discarded.Incremental measurement, memory, chain-of-thought and ensembling ablations compare task pipelines, not evolved versus unchanged experience-generation procedures.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.not_identifiednot established: experience buffers grow under an unchanged retrieval, reflection, voting and admission pipeline.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.MICCAI 2025 accepted/published version (official OA PDF); latest arXiv counterpart v2, 2025-10-01.https://papers.miccai.org/miccai-2025/paper/2575_paper.pdf[{'url': 'https://papers.miccai.org/miccai-2025/0537-Paper2575.html', 'location': 'official publication status, author list, DOI, pages, and OA version identity'}, {'url': 'https://papers.miccai.org/miccai-2025/paper/2575_paper.pdf', 'location': 'pp. 3–5, §2 and Fig. 2: two buffers, reflection and correctness admission'}, {'url': 'https://papers.miccai.org/miccai-2025/paper/2575_paper.pdf', 'location': 'pp. 5–7, §3: benchmark conversion, diagnosis scoring, Table 1 and incremental Table 2'}, {'url': 'https://arxiv.org/html/2503.22678v2#S2', 'location': 'matching methodology in current arXiv counterpart'}, {'url': 'https://api.crossref.org/works/10.1007/978-3-032-05114-1_35', 'location': 'publisher metadata and 2025-09-21 publication date'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
han2026gsemGSEMmemoryMemory and experience across encounters1offline_benchmarkExperience-node quality and inter-experience edge weights in a dual-layer graph; optional insertion of newly extracted experiences.Backbone models; typed experience schema; graph traversal actions; fixed rank-credit, node/edge calibration equations, learning rates and construction-time connection rule.Indication/contraindication experiences, entity/experience graph, reliability scores, relation weights and accumulated edge-credit state.Scalar task-outcome feedback in [-1,1] recalibrates activated nodes and edges; held-out reliability trials initialize experience quality.After each case, with reported snapshots after 50, 150 and 250 online updates.LLM extraction of successful strategies and failure divergences; fixed credit-assignment equations update graph weights from feedback.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.Static GSEM is compared with evolving-memory snapshots, plus retrieval ablations; this isolates graph-state calibration, not a self-revised calibration procedure.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.not_identifiednot established: feedback changes retrieval-affecting state under fixed update equations; the extraction, traversal and calibration procedures are not shown to revise themselves.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.Submission history From: Xiao Han [ view email ] [v1] Mon, 23 Mar 2026 15:24:15 UTC (5,083 KB)https://arxiv.org/html/2603.22096v1[{'url': 'https://arxiv.org/html/2603.22096v1#S3', 'location': 'task/experience definitions and dual-layer graph'}, {'url': 'https://arxiv.org/html/2603.22096v1#S4.SS1', 'location': 'extraction, held-out Experience Reliability Validation and graph construction'}, {'url': 'https://arxiv.org/html/2603.22096v1#S4.SS3', 'location': 'update targets and fixed Eqs. 7–10'}, {'url': 'https://arxiv.org/html/2603.22096v1#S5', 'location': 'benchmark setup, shared backbone/judging and evolution Table 2'}, {'url': 'https://arxiv.org/html/2603.22096v1#S6', 'location': 'retrieval ablations and reasoning-quality analysis'}, {'url': 'https://arxiv.org/html/2603.22096v1#Sx1', 'location': 'explicit lack of real workflow, longitudinal and expert evaluation'}, {'url': 'https://arxiv.org/html/2603.22096v1#A3', 'location': 'fixed traversal and graph hyperparameters'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
shen2026evomedagentEvo-MedAgentmemoryMemory and experience across encounters1retrospectiveEpisodic cases, diagnostic procedural rules/utility statistics and tool-trust state.Frozen VLM and specialist tools; fixed base protocol, memory factorization, retrieval/utility selection and tool-trust thresholds.Compressed episodes, priority-tagged procedural rules and usage statistics, plus per-tool helpful/harmful/misuse counts and trust labels.The benchmark ground-truth answer is revealed after each episode; multimodal reflection uses images, predictions, traces and that answer.Test-time, between sequential chest-X-ray cases; later cases inherit earlier-case memory.One post-case multimodal reflection call proposes procedural patches and governance updates; fixed accounting rules maintain utility and trust.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.No-memory and episodic/procedural ablations; fixed versus evolved reflection or memory-update procedures are not compared. Tool-governance quantitative ablation is explicitly deferred.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.not_identifiednot established: diagnostic procedures and tool-use preferences change; the mechanism generating and admitting future memory changes remains specified by the fixed reflection pipeline.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.Submission history From: Bailiang Jian [ view email ] [v1] Wed, 15 Apr 2026 23:12:02 UTC (508 KB)https://arxiv.org/html/2604.14475v1[{'url': 'https://arxiv.org/html/2604.14475v1#S3', 'location': 'sequential ground-truth reveal and fixed-parameter objective'}, {'url': 'https://arxiv.org/html/2604.14475v1#S3.SS1', 'location': 'three memory stores, utility selection, fixed tool-trust thresholds and deferred governance ablation'}, {'url': 'https://arxiv.org/html/2604.14475v1#S3.SS2', 'location': 'fixed base protocol and reflection-based patch updates'}, {'url': 'https://arxiv.org/html/2604.14475v1#S4', 'location': 'Table 1, sequential trajectories, three permutations and memory ablations'}, {'url': 'https://arxiv.org/html/2604.14475v1#S5', 'location': 'immediate-feedback, CXR-only and governance limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
ren2026dxevolveDxEvolvememoryMemory and experience across encounters1external_retrospectiveA growing repository of diagnostic cognition primitives (DCPs) that guides later test ordering and diagnostic decisions.Frozen base LLM; shared action/prompt contract, DCP-construction template, embedding retriever, evaluator and bounded diagnostic action space.DCPs with an experience pattern, test-ordering guidance, diagnostic-decision guidance and provenance including exposure index, pathology and source outcome.Accrual-case reference diagnoses, correctness flags, rule-based process feedback and charted clinician test orders guide DCP construction.Post-encounter consolidation in a dedicated accrual pool; nested exposure-indexed repositories are tested on fixed, non-overlapping evaluation cohorts.The same backbone uses a fixed reflection template to distill completed encounters into reusable DCPs.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.Backbone-matched DCP-free and retrieval ablations plus exposure curves isolate experiential reuse; no changed-versus-fixed DCP-construction procedure is evaluated.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.not_identifiednot 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.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.Submission history From: Ruiyang Ren [ view email ] [v1] Wed, 11 Mar 2026 11:41:51 UTC (1,166 KB)https://arxiv.org/html/2603.10677v1[{'url': 'https://arxiv.org/html/2603.10677v1#S4.SS1', 'location': 'workflow, three-part DCP, provenance and evidence-compatible reuse'}, {'url': 'https://arxiv.org/html/2603.10677v1#S4.SS3', 'location': 'accrual/evaluation separation, reader-study cohort and external hospital cohort'}, {'url': 'https://arxiv.org/html/2603.10677v1#S4.SS5', 'location': 'frozen backbones, common prompt contract, fixed dense retrieval and baselines'}, {'url': 'https://arxiv.org/html/2603.10677v1#S4.SS6', 'location': 'nested accrual pools, blinded clinician ratings and process metrics'}, {'url': 'https://arxiv.org/html/2603.10677v1#Sx4', 'location': 'fixed experience-construction prompt, ground truth and rule-based feedback'}, {'url': 'https://arxiv.org/html/2603.10677v1#S2.SS2', 'location': 'unmatched information access for published clinician reference'}, {'url': 'https://arxiv.org/html/2603.10677v1#S2.SS3', 'location': 'independent external validation'}, {'url': 'https://arxiv.org/html/2603.10677v1#S3', 'location': 'retrospective scope and prospective-study priorities'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
he2026evoclinicianEvoClinicianskillsSkills, tools and workflow adaptation1simulatedActor prompt rules and action-conditioned external memories.Base-model weights; Process Grader and Evolver instructions; preset memory-deletion coefficients and budgets.Bounded operational rules and context-action-outcome memories, including error examples.Diagnosis Judge grade, encounter costs, and action-level yield/efficiency critiques.Between diagnostic episodes; memories retrieved during subsequent encounters.Fixed-instruction LLM Evolver edits Actor state from graded trajectories.Budgeted add/delete/merge operations; generalize identifying details; no independent candidate-release test specified.Static, memory-only, prompt-only and transcript-feedback controls; these isolate components, not evolving versus fixed Evolver.Separate diagnosis Judge, but its score feeds adaptation; the reported fixed-order case stream is not an independent frozen holdout.not_identifiedNot established: Actor strategy changes; the described Evolver/Grader procedure is not revised.Supports sequential prompt/memory adaptation and resource-aware diagnosis in simulation, not patient-care safety or improved update-generation ability.arXiv v1, 2026-01-30https://arxiv.org/html/2601.22964v1[{'url': 'https://arxiv.org/html/2601.22964v1#S4', 'location': 'Sections 4.1–4.5; update access and mechanism'}, {'url': 'https://arxiv.org/html/2601.22964v1#S5', 'location': 'Tables 1–3; fixed-order evaluation and component ablations'}, {'url': 'https://arxiv.org/html/2601.22964v1#A6', 'location': 'fixed Evolver prompt and memory deletion'}, {'url': 'https://arxiv.org/html/2601.22964v1#A10', 'location': 'simulation and safety limitations'}, {'url': 'https://github.com/yf-he/EvoClinician', 'location': 'reference implementation README; no acceptance claim'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
liu2026evomdtEvoMDTskillsSkills, tools and workflow adaptation1external_retrospectivePrompts, consensus weights and retrieval scope (Abstract); post-task agent/workflow structure (Methods).Clinical role/schema design, guideline governance and assessment rubrics; no weight-training step specified for self-evolution.Revised prompts and coordination/retrieval configurations; versioned guideline knowledge base.Expert development-case feedback, performance metrics and reported outcome signals.Prompt development plus reported post-task instruction/workflow revision.Reported automatic prompt/workflow optimizer with expert-validated initial role prompts.Development prompts receive oncologist review; automatic candidate acceptance and rollback criteria are not detailed in Methods or supplement.No evolution-disabled or frozen-optimizer comparison identified in the main results and 16-page supplement.Six QA datasets, retrospective hospital records and single-blind physician scoring; these assess recommendations, not improvement cycles.not_identifiedReported configuration adaptation; improved ability of the updating procedure itself is not established.Supports evaluated final-system recommendation quality; does not isolate self-evolution gains or establish prospective patient benefit.Published 2026-01-07; version of record 2026-02-04https://www.nature.com/articles/s41746-025-02304-8[{'url': 'https://www.nature.com/articles/s41746-025-02304-8#Sec11', 'location': 'Multi-agent EvoMDT; prompt optimization and self-evolution'}, {'url': 'https://www.nature.com/articles/s41746-025-02304-8#Sec13', 'location': 'four-center, 405-case retrospective design'}, {'url': 'https://www.nature.com/articles/s41746-025-02304-8#Sec14', 'location': 'single-blind physician protocol'}, {'url': 'https://www.nature.com/articles/s41746-025-02304-8#Sec8', 'location': 'Discussion and limitations'}, {'url': 'https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41746-025-02304-8/MediaObjects/41746_2025_2304_MOESM1_ESM.pdf', 'location': '16-page supplement; guideline catalog, cohort figures and evaluation rubric'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
yang2026psychagentPsychAgentskillsSkills, tools and workflow adaptation2-implicitsimulatedClient memory, therapeutic skill tree and shared counseling-model parameters.Therapeutic hierarchy, reward-based session selection and rejection-fine-tuning objective.Client profiles, summaries, atomic skills and fine-tuned model weights.LLM-rated counseling trajectories; expert-annotated tasks supervise initial memory, planning and skill curation.Memory/skill revision across sessions; parameter internalization after selected multi-session rollouts.Learned skill extractor/manager appends or merges skills; best-of-eight rollouts supply RFT data.Highest-reward session continues the trajectory; selected trajectories train the model. No independent skill-release test described.Memory/planning, skill-evolution and internalization ablations; no dedicated control isolating shared-model improvements as an updater.Separate evaluation profiles plus human ratings of 522 dialogues by two psychology-trained graduate raters; simulated clients.not_identifiedReported shared-model learning can affect skill extraction/management; improved update-generation ability is not independently established.Response-quality and memory/skill benefits; not measured therapeutic outcomes. Public release lacks the complete post-session skill-evolution pipeline.arXiv v3, 2026-04-28; v1 2026-04-01https://arxiv.org/html/2604.00931v3[{'url': 'https://arxiv.org/html/2604.00931v3#S3', 'location': 'three engines; equations 4–6'}, {'url': 'https://arxiv.org/html/2604.00931v3#S4.SS1', 'location': 'profile split and implementation'}, {'url': 'https://arxiv.org/html/2604.00931v3#S4.SS4', 'location': 'Table 3 ablations'}, {'url': 'https://arxiv.org/html/2604.00931v3#A2', 'location': 'human-rater qualifications and matched dialogue sample'}, {'url': 'https://github.com/ECNU-ICALK/PsychAgent', 'location': 'README: incomplete public skill-evolution pipeline'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
sun2026skemexSkeMexskillsSkills, tools and workflow adaptation1offline_benchmarkSkill contents, context-conditioned utility values and repository membership/status.Backbone weights; retrieval formula, buffer gates, utility update schedule and governance rules.General, task-level and action-level skills with utility and lifecycle metadata.Benchmark returns, skill-adoption judgments and clinical-category reward baselines.Retrieve once per episode; write/value after 30-trajectory windows; governance every two windows.Fixed-backbone two-pass skill analysis/mutation plus prescribed valuation/governance operators.Gated trajectory buffer; novelty/quality checks; utility-based promotion, merging, deprecation and capacity limits.Matched-tool/prompt/inference-budget ablations freeze writing, valuation, governance or retrieval components; no evolving-operator comparator.Offline held-out in-domain and unseen-benchmark-family tests; cross-backbone skill transfer; separate online repeated-stream experiments.not_identifiedNot established: learned utility/state changes under specified update equations; operator rules remain prescribed.Supports utility-guided skill memory and transfer; online repeated-epoch gains are distinct from independent clinical validation.Audited arXiv v4, 2026-10-09; v1 2026-06-08https://arxiv.org/html/2606.09365v4[{'url': 'https://arxiv.org/html/2606.09365v4#S3', 'location': 'equations 5–9; writing, valuation and governance'}, {'url': 'https://arxiv.org/html/2606.09365v4#S4.SS3', 'location': 'Table 2 lifecycle ablations; Table 3 online epochs'}, {'url': 'https://arxiv.org/html/2606.09365v4#A6', 'location': 'Table 10 split and fixed hyperparameters'}, {'url': 'https://arxiv.org/html/2606.09365v4#A9', 'location': 'OOD and cross-backbone transfer'}, {'url': 'https://arxiv.org/html/2606.09365v4#A10', 'location': 'limitations'}, {'url': 'https://neurips.cc/virtual/2026/poster/150444', 'location': 'official NeurIPS 2026 Poster listing'}, {'url': 'https://openreview.net/forum?id=i09nlt9B6r', 'location': 'linked from official conference listing'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zhu2026healthflowHealthFlowskillsSkills, tools and workflow adaptation1offline_benchmarkPersistent safeguards, workflows, dataset anchors, code snippets and memory confidence/status.Four-agent orchestration, API-model weights, outcome-dependent writeback rules and type-specific retrieval gates.Typed, scoped experience records with provenance, applicability and confidence.Execution artifacts/traces, methodological-validity verdicts and recovered/failed/successful task outcomes.Evaluator-guided within-task retries; reflector-driven memory updates after each task.Fixed-model reflector synthesizes guidance and validates, weakens or retires previous memory.Outcome-aware writeback restricts failed runs to safeguards; scope/compatibility gates bound later retrieval.Progressive addition of planner/evaluator, reflector-memory and tools; no separately evolving-reflector versus fixed-reflector experiment.External Gemini judge with expert-built references; 12 blinded experts assess 20 reports. Online memory updates continue within each benchmark.not_identifiedNot established: task procedures and memories evolve within prescribed reflection/retrieval/writeback rules.Supports governed analytical reuse and artifact quality. The pilot does not establish prospective clinical outcome benefit.Published 2026-08-17; version of record 2026-08-27; supersedes arXiv:2508.02621 titlehttps://www.nature.com/articles/s41746-026-03097-0[{'url': 'https://www.nature.com/articles/s41746-026-03097-0#Sec12', 'location': 'Algorithm 1; within-task versus cross-task indices'}, {'url': 'https://www.nature.com/articles/s41746-026-03097-0#Sec13', 'location': 'equations 2–4; retrieval and outcome-specific writeback'}, {'url': 'https://www.nature.com/articles/s41746-026-03097-0#Sec15', 'location': 'independent judging, sequential online evaluation and human review'}, {'url': 'https://www.nature.com/articles/s41746-026-03097-0#Sec5', 'location': 'progressive ablation'}, {'url': 'https://www.nature.com/articles/s41746-026-03097-0#Sec9', 'location': 'limitations'}, {'url': 'https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41746-026-03097-0/MediaObjects/41746_2026_3097_MOESM1_ESM.pdf', 'location': 'supplementary judge agreement and repair dynamics'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
du2025evopatientEvoPatientgeneratorsParameter updates and changing training generators2-explicitsimulatedPatient demonstration/requirement library and doctor questioning-trajectory library.Model weights, initial standardized-patient requirements, quality criteria and designed recruitment topology.Attention library of question-record-answer-requirement tuples and doctor transition trajectories.Validated question/answer quality and subsequent usefulness of retrieved library entries.Dialogue-level storage and reuse during simulated encounters; persist across cases.Evolving doctor questioning generates patient-learning interactions; attention agent extracts requirements.Quality validation before insertion; keep stronger duplicates; remove entries after two harmful uses.Tables 4–5 test doctor evolution and recruitment; patient-impact evolved-doctor condition bundles trajectory, question-pool and profile components.Separate disease-transfer assessment and human/GPT-4 preferences; no prospective human-trainee clinical-outcome test.not_identifiedDemonstrated at interaction-generator level: evolved doctors alter patient-training experience. Recursive improvement of the whole coevolution procedure is not established.Supports co-adaptive training interactions. Bundled ablations and fixed evaluation/coevolution rules preclude an unqualified strict-RSI claim.ACL Anthology proceedings version, 2025-07; pages 17278–17306https://aclanthology.org/2025.acl-long.846.pdf[{'url': 'https://aclanthology.org/2025.acl-long.846.pdf', 'location': 'Section 3.3, pp. 17281–17282; two evolving libraries'}, {'url': 'https://aclanthology.org/2025.acl-long.846.pdf', 'location': 'Section 4.3, p. 17285; transfer from one disease to five'}, {'url': 'https://aclanthology.org/2025.acl-long.846.pdf', 'location': 'Section 4.4, Tables 4–5, pp. 17285–17286; generator and patient-impact ablations'}, {'url': 'https://aclanthology.org/2025.acl-long.846.pdf', 'location': 'Appendix E, p. 17293; correction/deletion rules'}, {'url': 'https://aclanthology.org/2025.acl-long.846.pdf', 'location': 'Algorithm 1, p. 17294; quality validation'}, {'url': 'https://aclanthology.org/2025.acl-long.846/', 'location': 'complete authors, venue, pages and DOI'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026healthclawHealthClawmemoryMemory and experience across encounters1simulatedUser profile, reusable procedures and episodic records.Shared safety rules and domain knowledge; memory-induction and routing design.User-specific profile, standard operating procedures and episodic memory.Completed episode content, evolving user state and memory-curation rules.Writeback after episodes; writeback disabled during evaluation probes.LLM induction routes information to profile, procedure, episode or exclusion.Memory routing and privacy controls; no separate validated improvement-procedure release reported.Memory/context baselines; biomedical comparisons also differ in backbone and pipeline.Hidden-ledger probes on a synthetic year; nine biomedical task datasets.not_identifiedMemory contents evolve under a fixed induction design.Longitudinal simulation is not prospective patient follow-up; task gains do not isolate memory causally.v1https://arxiv.org/abs/2607.13940v1[{'url': 'https://arxiv.org/html/2607.13940v1#Sx1', 'location': 'Introduction'}, {'url': 'https://arxiv.org/html/2607.13940v1#Sx2', 'location': 'Results'}, {'url': 'https://arxiv.org/html/2607.13940v1#Sx3', 'location': 'Discussion'}, {'url': 'https://arxiv.org/html/2607.13940v1#Sx4', 'location': 'Methods'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
shi2026emrEMRmemoryMemory and experience across encounters1offline_benchmarkClinical principles, diagnostic patterns and representative cases.LLM weights, role prompts, embedding retrieval and similarity-threshold update rule.Hierarchical experience library containing positive insights and warnings.Ground-truth answers applied to generated multi-agent trajectories.Offline iterative experience mining before benchmark evaluation.LLM extracts experience, then adds, merges or discards entries.Similarity threshold and LLM consolidation govern library updates.Experience/hierarchy ablations and cross-backbone transfer; no evolving-versus-fixed updater test.Mining and testing subsets separated; most datasets evenly resplit, PubMedQA uses official test set.not_identifiedExperience contents change; update algorithm remains fixed.Custom halves of several datasets do not match all published official test protocols; no clinical outcomes.v1https://arxiv.org/abs/2609.15161v1[{'url': 'https://arxiv.org/html/2609.15161v1#S3.SS2', 'location': 'Section 3.2: mining'}, {'url': 'https://arxiv.org/html/2609.15161v1#S3.SS3', 'location': 'Section 3.3: update and retrieval'}, {'url': 'https://arxiv.org/html/2609.15161v1#A1', 'location': 'Appendix A: mining/test split'}, {'url': 'https://arxiv.org/html/2609.15161v1#Sx1', 'location': 'Limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026medforgeMedForge-RSIskillsSkills, tools and workflow adaptation1retrospectiveError-derived experience, image-analysis tools and optional directive rules.Detector weights, official system prompt, scorer, evaluation sets and learning-stream schedule.Accepted memory entries, executable tools and reversible directive edits.Verified labels on each learning batch; replay and validation performance.Twenty simulated deployment rounds.Detector or external/ensemble reflector proposes structured edits.Replay plus disjoint validation A/B checks; rollback variants; verifier-failure fallback also reported.Fixed configurations and mechanism ablations; the improvement mechanism itself is protected.Final 4,000-image held-out evaluation, separate from validation and learning.not_identifiedAuthors explicitly claim bounded self-refinement, not modification of their own improvement mechanism.One resample set is contaminated; final held-out test is distinct. Some configurations/seeds were incomplete.v1https://arxiv.org/abs/2609.36549v1[{'url': 'https://arxiv.org/html/2609.36549v1#S3.SS4', 'location': 'Section 3.4: protected components and gates'}, {'url': 'https://arxiv.org/html/2609.36549v1#S4', 'location': 'Section 4: data splits and contaminated resample'}, {'url': 'https://arxiv.org/html/2609.36549v1#S7.SS0.SSS0.Px3', 'location': 'Section 7: limitations and bounded claim'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zhang2026vibemedVIBEMedgeneratorsParameter updates and changing training generators1offline_benchmarkMemory, model parameters and generated code/routing described at different timescales.Three-role framework and prescribed evaluation/deployment workflow.Memory, adapted model and validated code; quantitative evolution test concerns parameter tuning.Completed cases, reference treatment samples and model judges.Post-encounter memory; scheduled parameter adaptation and on-demand code extension described.Clinical Evolution Manager and backbone LLM; SFT/DPO and code repair described.Model-judge screening and rollback; sandboxed code tests followed by user confirmation.No component-isolating memory/model/code evolution ablation; explicitly deferred.MedBench base-versus-tuned comparison and clinical case demonstration.not_identifiedCode evolution is described; improved effectiveness of the update procedure is not isolated.MedAnalysis and MedDiag decline after tuning; a case demonstration does not establish prospective utility.v2https://arxiv.org/abs/2606.15504v2[{'url': 'https://arxiv.org/html/2606.15504v2#S2.SS3', 'location': 'Section 2.3: three-level mechanism'}, {'url': 'https://arxiv.org/html/2606.15504v2#S3.SS2', 'location': 'Section 3.2 and Table 2: evolution evaluation'}, {'url': 'https://arxiv.org/html/2606.15504v2#S3.SS3', 'location': 'Section 3.3: case demonstration'}, {'url': 'https://arxiv.org/html/2606.15504v2#S4', 'location': 'Conclusions: untested component effects'}, {'url': 'https://doi.org/10.1016/j.metrad.2026.100223', 'location': 'Published metadata; extraction uses arXiv v2'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wang2026ecgrsiECG-RSIgeneratorsParameter updates and changing training generators2-explicitretrospectiveLearner weights; synthesis operators, prompts, reusable skills and data recipes.Acceptance gates; vision encoder/aligner; selection scoring and SFT protocol.Versioned generation rules, accepted examples, selected checkpoints and lineage.Waveform remeasurement/construction failures; learner errors on separate feedback data.Offline synthesis-repair cycles and three model-training rounds.Prompted local repairs and feedback-conditioned same-parent recipe branches.Independent operator/question regression checks; selection-set score and regression gates.Yes: frozen-operator/skill ablations; equal accepted-data volume and matched training, unequal generation counts.Patient-disjoint sources/development/tests; two physicians review synthetic samples, not clinical outcomes.identifiedGenerator-rule changes supported by synthesis-yield and downstream-performance comparisons.V1 reports code/data unreleased. Total-search compute parity not established; physician explanation support remains imperfect.v1 (2026-10-13)https://arxiv.org/html/2610.17718v1[{'url': 'https://arxiv.org/html/2610.17718v1#S3.SS1', 'location': 'Section 3.1: separate learner, training pool and construction policy; data boundaries'}, {'url': 'https://arxiv.org/html/2610.17718v1#S3.SS2.SSS2', 'location': 'Section 3.2.2: old/new operator and prompt repair comparisons'}, {'url': 'https://arxiv.org/html/2610.17718v1#A1.SS2', 'location': 'Appendix A.2: exact rule changes, proposer and independent regression checks'}, {'url': 'https://arxiv.org/html/2610.17718v1#A4.SS5', 'location': 'Prompt P5: repair generator component while acceptance criteria remain fixed'}, {'url': 'https://arxiv.org/html/2610.17718v1#S5.T10', 'location': 'Table 10: 10,000 accepted samples; candidates 16,193 full, 36,723 frozen operators, 22,398 frozen skills'}, {'url': 'https://arxiv.org/html/2610.17718v1#S5.T15', 'location': 'Main Table 15: R1 S7 full 43.22±1.40; frozen operators 39.29±1.30; frozen skills 40.15±1.30'}, {'url': 'https://arxiv.org/html/2610.17718v1#S5.SS6', 'location': 'Section 5.6/Table 11: same-parent fixed-mixture control'}, {'url': 'https://arxiv.org/html/2610.17718v1#S5.SS7', 'location': 'Section 5.7/Table 12: per-branch 10,000 additions and matched replay/training; not proof of total tree/chain compute equality'}, {'url': 'https://arxiv.org/html/2610.17718v1#S5.T7', 'location': 'Table 7: 240 generated samples, two physicians; answer correctness 0.88, explanation support 0.76'}, {'url': 'https://arxiv.org/html/2610.17718v1#S6', 'location': 'Conclusion: unreleased implementation/data and diminishing external gains'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
dang2026medzeroMedZEROgeneratorsParameter updates and changing training generators2-explicitoffline_benchmarkExaminer generator parameters, Reasoner parameters and persistent generated knowledge graph.Reward definitions, frontier-admission thresholds, causal visibility and curation protocol.Updated Examiner/Reasoner checkpoints, retained curriculum and filtered knowledge triplets.Current Reasoner agreement/frontier scores, diversity and retrieved-evidence grounding.Offline alternating GRPO rounds; counterpart frozen during each stage.Trainable Examiner generates curriculum; Reasoner performance guides Examiner optimization.Discard exploratory triplets; regenerate, filter valid/frontier-range samples, deduplicate and retain.Curriculum-source/curation/retrieval ablations; matched frozen-Examiner comparison not reported.Public benchmark gold labels; separate option-mapping judge. Curation is a separate stage, not independent truth adjudication.not_identifiedYes: the curriculum generator is optimized as the learner changes; isolated generator-change benefit remains unresolved.Incomplete grounding/consistency verification; benchmark gains do not demonstrate clinical utility.v1 (2026-10-06)https://arxiv.org/html/2610.08327v1[{'url': 'https://arxiv.org/abs/2610.08327v1', 'location': 'ArXiv comment: accepted by NIPS 2026; date/authors'}, {'url': 'https://arxiv.org/html/2610.08327v1#S3.SS1', 'location': 'Section 3.1: alternating Examiner and Reasoner updates; counterpart frozen'}, {'url': 'https://arxiv.org/html/2610.08327v1#S3.SS2', 'location': 'Section 3.2: Examiner GRPO and fixed frontier/diversity/grounding reward'}, {'url': 'https://arxiv.org/html/2610.08327v1#S3.SS3', 'location': 'Section 3.3: causal visibility; temporary buffer discarded; separate curation pass'}, {'url': 'https://arxiv.org/html/2610.08327v1#S3.SS4', 'location': 'Section 3.4: fixed Reasoner consistency/grounding reward'}, {'url': 'https://arxiv.org/html/2610.08327v1#S3.SS5', 'location': 'Section 3.5: open-ended responses and separate mapping judge'}, {'url': 'https://arxiv.org/html/2610.08327v1#S5.F3', 'location': 'Figure 3 inspected visually: panel c varies KG-only/Web-only/generated-triplets+Web curriculum source, not a frozen-Examiner control'}, {'url': 'https://arxiv.org/html/2610.08327v1/analysis_fig1.svg', 'location': 'Original Figure 3 SVG rendered for inspection'}, {'url': 'https://arxiv.org/html/2610.08327v1#S5.SS3', 'location': 'Section 5.3: three-round evolution'}, {'url': 'https://arxiv.org/html/2610.08327v1#A2', 'location': 'Appendix B: verification, persistent-error and clinical-utility limitations'}, {'url': 'https://arxiv.org/html/2610.08327v1#A6', 'location': 'Appendix F: fixed Reasoner, variable retrieval/prompt/observation budget; within-table comparison only'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
tzanis2026reclaimReclAImcomparatorsDiagnostic training and informative adaptation comparators1retrospectiveDownstream medical-image classifier weights, through full/partial/head-only fine-tuning or gradual unfreezing; the language-model agent itself is not the trained target.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.Trained/fine-tuned classifier checkpoints, configurations, evaluation/comparison JSON/CSV files, execution traces and conversational task context.Ground-truth labels for monitoring images/scans produce accuracy, precision, recall and F1 metrics; relative decline against development-test metrics triggers intervention.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.A master LLM agent routes comparison and fine-tuning tasks to specialist agents, which invoke predefined training tools using user constraints or configured defaults.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.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.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.not_identifiednot 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.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.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.https://arxiv.org/pdf/2510.17004v2[{'url': 'https://pubs.rsna.org/doi/10.1148/ryai.250923', 'location': 'title, authors, volume 8/issue 4, online date and abstract'}, {'url': 'https://api.crossref.org/works/10.1148/ryai.250923', 'location': 'article e250923 and issue publication metadata'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=5', 'location': 'Experimental setup steps I–V: label-based monitoring, separate fine-tuning data, re-evaluation on inference data'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=8', 'location': '5% trigger and eight affected model–dataset combinations'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=11', 'location': 'label availability, single temporal dataset, random partitions, no forgetting/control-pipeline tests'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=23', 'location': 'Supplement A: agent roles and fine-tuning strategies'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=24', 'location': 'fixed GPT-4.1, permanent prompt, parameter-anchoring loss'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=45', 'location': 'relative metric-change formula and threshold documentation'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=52', 'location': 'tool validation accuracy, early stopping and best checkpoint selection'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=56', 'location': 'Supplement B: general 60/20/20 with inner 70/15/15, alongside MIMIC chronological 40/gap20/recent40'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=60', 'location': 'Supplement C DenseNet121 prompts explicitly use distinct fine-tuning data and reuse the same inference file for post-update evaluation'}, {'url': 'https://arxiv.org/pdf/2510.17004v2#page=68', 'location': 'example human-specified head-only recipe for cardiomegaly InceptionV3'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wang2026privmesaPrivMeSAmemoryMemory and experience across encounters1retrospectiveTwo 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.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.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).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.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.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.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.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.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.not_identifiedNot 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.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.v1https://arxiv.org/abs/2609.38458v1[{'url': 'https://arxiv.org/html/2609.38458v1#S3.SS4', 'location': 'Section 3.4: lesson distillation, post-episode memory update, fixed retriever/judge/distiller and memory updates during RL training'}, {'url': 'https://arxiv.org/html/2609.38458v1#S3.SS3', 'location': 'Section 3.3: episode reward (accuracy, discretion, anonymity), GRPO optimization, frozen specialists/attacker/distiller/judge/retriever'}, {'url': 'https://arxiv.org/html/2609.38458v1#S4.SS1', 'location': 'Section 4.1: patient-disjoint split, registry, tasks, evaluation attacker, paired test-time learning protocol, baselines and implementation'}, {'url': 'https://arxiv.org/html/2609.38458v1#S3.T1', 'location': 'Table 1: main accuracy, leak and k<=10 results for Gemma 4 12B and Granite 4.2 8B versus Delegation, PAPILLON and Remote only'}, {'url': 'https://arxiv.org/html/2609.38458v1#A5.SS1', 'location': 'Appendix E.1 (with Table 3): test-time memory growth, share answered from memory, messages per case, memory on versus off'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wang2026tcmseaTCM-SEA / TCM-MisDxskillsSkills, tools and workflow adaptation1simulatedA free-text clinical skill document in the system prompt, which starts empty and is edited by add, replace or delete operations.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.The accepted skill-document text; two edits were accepted (rounds 1 and 6), both short pre-submission checklists.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.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.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.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.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.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.not_identifiedNot established: the diagnose-propose-gate procedure, prompts and acceptance rule stay fixed across rounds, and only the skill document changes.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.v1https://arxiv.org/abs/2610.40663v1[{'url': 'https://arxiv.org/html/2610.40663v1#S3.SS0.SSS0.Px1', 'location': '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'}, {'url': 'https://arxiv.org/html/2610.40663v1#S4.SS0.SSS0.Px2', 'location': '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'}, {'url': 'https://arxiv.org/html/2610.40663v1#S4.SS0.SSS0.Px3', 'location': 'Section 4, Learning from expert cases: feedback contents for TCM-SEA versus the outcome-only ablation'}, {'url': 'https://arxiv.org/html/2610.40663v1#S5.T2', 'location': 'Section 5, Setup and Table 2: DeepSeek-Flash for agent and judge, 9 rounds, seed 42, sealed test set, mean of three runs'}, {'url': 'https://arxiv.org/html/2610.40663v1#S5.SS0.SSS0.Px3', 'location': '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'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
yuan2026asclepiusAsclepiusskillsSkills, tools and workflow adaptation1simulatedThe natural-language operating manual (main-agent prompt with triage, treatment-completeness and disposition rules and a presentation-to-management table).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.The selected evolved manual M* (v7), plus a filesystem of all prior candidate manuals with their scores and traces that the proposer reads.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.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.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.After 10 iterations the manual with the best overall search-batch score (v7) is selected and reused unchanged on held-out batches.Add-one component ablations and an evolved-versus-initial manual contrast holding skills and subagents fixed; no evolving-versus-fixed proposer test.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.not_identifiedNot reported; the proposer model, prompt and best-score selection rule are fixed across all iterations.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).v1https://arxiv.org/abs/2609.13543v1[{'url': 'https://arxiv.org/html/2609.13543v1#S4.SS1', 'location': '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.'}, {'url': 'https://arxiv.org/html/2609.13543v1#S5.T2', 'location': 'Table 2: held-out (batches 7-10, n=48) mixed-effects differences, critical actions +0.625 (p=0.024), timeliness and overall ns.'}, {'url': 'https://arxiv.org/html/2609.13543v1#A5.T5', 'location': '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.'}, {'url': 'https://arxiv.org/html/2609.13543v1#A5.T9', 'location': '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.'}, {'url': 'https://arxiv.org/html/2609.13543v1#S8.SS0.SSS0.Px3', 'location': 'Section 8 Limitations: CES is a simulator with LLM-generated physiology, single ED configuration and single backbone; performance is a proxy for clinical competence.'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.2094/', 'location': 'Published metadata; extraction uses arXiv v1'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
lee2026revision3dReVision3DgeneratorsParameter updates and changing training generators2-explicitretrospectiveReader 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.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.The accepted incumbent system (adapter weights, recipe, inference program) plus an evidence-backed playbook of successful and failed hypotheses and trial registries.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.Offline revision rounds on development patients, each accepted change triggering renewed attribution; held-out test patients are read once after all choices are frozen.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.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.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.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.identifiedThe 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.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.v2https://arxiv.org/abs/2609.32984v2[{'url': 'https://arxiv.org/html/2609.32984v2#S3.SS0.SSS0.Px1.p2', 'location': 'Section 3, Motivation: system components (view selection, perception, weights, inference), frozen designer, and fixed verifier, masks, objective and acceptance rule.'}, {'url': 'https://arxiv.org/html/2609.32984v2#S3.SS0.SSS0.Px8.p2', 'location': 'Section 3, Recursive improvement: seed-range and bootstrap acceptance rules, playbook carried across rounds, routing rule formulated retrospectively.'}, {'url': 'https://arxiv.org/html/2609.32984v2#A5.T9', 'location': 'Appendix E.1, Table A9: liver J and max recall after each accepted revision and who made each change.'}, {'url': 'https://arxiv.org/html/2609.32984v2#S5.T2', 'location': 'Section 5.3, Table 2: development (selection) and single-run test J for perception and inference revisions.'}, {'url': 'https://arxiv.org/html/2609.32984v2#A7.T16', 'location': 'Appendix G.1, Table A16: inference-level search repeated with four designers from three model families.'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
hu2026templarTEMPLARcomparatorsDiagnostic training and informative adaptation comparators1retrospectiveA 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.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.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.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.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.LLM Evolution Agent under fixed prompts proposes assignments, merges, fine revisions, evidence notes and slot-specific patterns P_s.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.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.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.not_identifiednot established: the template and graphs change once under constant prompts, thresholds and reducer rules, and the construction procedure itself is not revised.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.v1https://arxiv.org/abs/2610.05247v1[{'url': 'https://arxiv.org/html/2610.05247v1#S3.SS2', 'location': 'Section 3.2: template evolution as assembly, revision, external evidence grounding and one feedback-guided granularity calibration.'}, {'url': 'https://arxiv.org/html/2610.05247v1#A2.SS2.SSS3', 'location': 'Appendix B.2.3: calibration audits sampled training structurings and adds slot-specific patterns without changing allowed values.'}, {'url': 'https://arxiv.org/html/2610.05247v1#S4.SS1', 'location': 'Section 4.1: four datasets, train/test splits, ASTAR and nine LLM baselines.'}, {'url': 'https://arxiv.org/html/2610.05247v1#S4.SS4', 'location': 'Section 4.4: cumulative ablation M1 (ASTAR) to M6 on PETCT, ViMed and PETCT-to-ViMed transfer.'}, {'url': 'https://arxiv.org/html/2610.05247v1#S5.SSx2', 'location': 'Ethics Statement: research use only; metrics and expert assessments do not establish clinical safety or deployment readiness.'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
schmidgall2026agentclinicAgentClinicmemoryMemory and experience across encounters1simulatedA 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.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.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.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.Online and sequential: the notebook is rewritten after every case and used on the following AgentClinic-MedQA cases.The doctor LLM itself under a fixed notebook-writing meta-prompt, instructed to write general, non-patient-specific tips from successes and failures.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.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.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.not_identifiedNot reported: the notebook-writing prompt and update procedure stay constant; only the notebook contents change.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.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.https://www.nature.com/articles/s41746-026-02674-7[{'url': 'https://www.nature.com/articles/s41746-026-02674-7#Abs1', 'location': 'Abstract: Llama-3 shows up to 92% relative improvement with a notebook that persists across cases.'}, {'url': 'https://www.nature.com/articles/s41746-026-02674-7#Sec5', 'location': 'Results, Comparing tools from the Agent Toolbox: six tools, Claude 3.5 56.1% with Notebook, Llama3-70b 41.1%, GPT-3.5 declines.'}, {'url': 'https://www.nature.com/articles/s41746-026-02674-7#Sec21', 'location': 'Methods, Tool descriptions: notebook persists across patients; experiential learning module stores takeaways from correct cases.'}, {'url': 'https://arxiv.org/html/2405.07960v5#A15.T6', 'location': 'arXiv v5 Table 6: per-tool accuracy with change from no-tool baseline for six backbones.'}, {'url': 'https://arxiv.org/html/2405.07960v5#A13.SS2', 'location': 'arXiv v5 Appendix M.2: notebook prompt reveals correct diagnosis, 1000-character overwrite limit.'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
feng2026medexpmemMedExpMemmemoryMemory and experience across encounters1retrospectiveAn 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.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.The pairwise differential notes, merged per differential pair, persist as a separate memory library built independently for each VLM.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.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.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.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.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.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.not_identifiedNot 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.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.arXiv 2605.22872v1 (2026-05-20, only version listed by the arXiv API on 2026-11-03); the LNCS chapter text was not examined.https://arxiv.org/html/2605.22872v1[{'url': 'https://arxiv.org/html/2605.22872v1#S2.T1', 'location': 'Section 2.2, Table 1: note schema, and 11 departments / 118 organ-region taxonomy.'}, {'url': 'https://arxiv.org/html/2605.22872v1#S2.SS3', 'location': 'Section 2.3: Phase I error-driven extraction with expert discussion; Phase II re-diagnosis, refinement and merging.'}, {'url': 'https://arxiv.org/html/2605.22872v1#S3.SS1', 'location': 'Section 3.1: 6,228 pre-2025 construction cases, 227 test cases from 2025, tau=0.9, K=10, per-agent memory.'}, {'url': 'https://arxiv.org/html/2605.22872v1#S3.T2', 'location': 'Table 2: baseline vs w/Exp accuracy, recall, precision, beneficial and harmful rates for seven VLMs.'}, {'url': 'https://arxiv.org/html/2605.22872v1#S3.T3', 'location': 'Table 3: Qwen3-VL-8B PubMed-RAG comparison and construction/retrieval ablations.'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
shi2024ehragentEHRAgentmemoryMemory and experience across encounters1offline_benchmarkA long-term memory of past successful question and code solutions, retrieved as the K=4 few-shot demonstrations (§3.2, Eq. 1).GPT-4 (gpt-4-0613, temperature 0) backbone, tool definitions, EHR metadata, prompts and Levenshtein-similarity retrieval stay fixed (§3.2, §4.1).Successful past cases (question plus code solution) that replace the original hand-written demonstrations (§3.2, App. H.2 Fig. 9).Ground-truth success of each answer decides memory admission; execution error messages drive within-case debugging only (§3.2, App. F.2).Online across the evaluated question stream; no separate training split for memory is reported (§3.2, App. F.2).The agent's own successful solving runs; no separate generator component.A case enters memory only when it succeeds against the ground-truth answer; a variant admitting all completed cases is also tested (Table 8).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).Not reported: success rate is measured on the same question stream from which memory accumulates (§4.1, App. F.2).not_identifiedNone; the storage rule and retrieval procedure are fixed.Offline benchmark with ground-truth labels used at test time for memory admission; no clinician-facing or prospective evaluation (§4.1, Limitations).arXiv v3 (2024-10-04) read in full; numbers checked against the EMNLP 2024 Anthology PDFhttps://arxiv.org/html/2401.07128v3[{'url': 'https://arxiv.org/html/2401.07128v3#S3.SS2', 'location': '§3.2 Demonstration Optimization through Long-Term Memory'}, {'url': 'https://arxiv.org/html/2401.07128v3#S4.T2', 'location': 'Table 2, ablation on MIMIC-III'}, {'url': 'https://arxiv.org/html/2401.07128v3#A6.T8', 'location': 'Table 8, LTM with success vs completion vs none'}, {'url': 'https://arxiv.org/html/2401.07128v3#A6.T7', 'location': 'Table 7, eICU ablation'}, {'url': 'https://aclanthology.org/2024.emnlp-main.1245.pdf#page=4', 'location': 'Published version, Algorithm 1 and §3.2'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
liao2025reflectoolReflecToolmemoryMemory and experience across encounters1offline_benchmarkA long-term memory of successful reflective trajectories and a per-tool store of tool-wise experience (§3.2, Eq. 3-4).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).Successful refined trajectories as memory items, plus merged action-wise suggestions for each tool and for Plan, Think and Finish (Algorithm 1).Ground-truth answers on the optimization subset; the agent reflects on its first trajectory against the answer and retries (§3.2).Offline optimization stage before inference; memory and experience are then frozen and only read at inference (§3.2-3.3).The same LLM agent writes reflections and tool-wise suggestions by comparing failed and successful trajectories (Eq. 4, App. D.2 Fig. 8-9).A trajectory is stored only when its refined answer equals the ground truth; tool experience is merged without further validation (Eq. 3).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).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).not_identifiedNone; the reflection and storage procedure is fixed.Offline benchmark; authors note correct answers do not guarantee correct trajectories (Limitations); no clinician-facing or prospective evaluation.arXiv v3 (2025-06-16) read in full; key table values match the ACL 2025 Anthology PDFhttps://arxiv.org/html/2410.17657v3[{'url': 'https://arxiv.org/html/2410.17657v3#S3.SS2', 'location': '§3.2 Optimization Stage'}, {'url': 'https://arxiv.org/html/2410.17657v3#S3.SS3', 'location': '§3.3 Inference Stage'}, {'url': 'https://arxiv.org/html/2410.17657v3#S3.T3', 'location': 'Table 3, ablation of memory and tool-wise reflection'}, {'url': 'https://arxiv.org/html/2410.17657v3#S5.T4', 'location': 'Table 4, long-term memory given to baselines'}, {'url': 'https://arxiv.org/html/2410.17657v3#S5.F3', 'location': 'Figure 3, effect of optimization step'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026mdteamgptMDTeamGPTmemoryMemory and experience across encounters1offline_benchmarkTwo knowledge bases: CorrectKB for correct consultations and ChainKB for reflective analyses of incorrect ones (§3.3).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).Structured summaries: question, answer and final-round statements (CorrectKB); hypothesis, analysis, conclusion and error reflection (ChainKB) (§3.3).The final consultation answer is validated against the ground truth, which routes the record to CorrectKB or ChainKB (§3.3).Accumulated over 900 consultation rounds on training sets before testing; test-time KBs contain only training-set experience (§4.1-4.2).A Chain-of-Thought Reviewer agent summarizes each finished consultation into the KB format (§3.3).Routing by ground-truth correctness only; no further validation, pruning or release gate is reported (§3.3).Ablation adding CorrectKB, ChainKB or both with the discussion protocol fixed (Table 2); accuracy versus accumulated cases for eight backbones (Fig. 4).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).not_identifiedNone; the reviewer and routing rule are fixed.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).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)https://aclanthology.org/2026.findings-acl.1427.pdf[{'url': 'https://aclanthology.org/2026.findings-acl.1427.pdf#page=5', 'location': '§3.3 Summary and Output Stage, KB construction'}, {'url': 'https://aclanthology.org/2026.findings-acl.1427.pdf#page=7', 'location': 'Table 2 ablation and §4.5 Self-Evolving Capability'}, {'url': 'https://aclanthology.org/2026.findings-acl.1427.pdf#page=8', 'location': 'Fig. 4 and Table 3 cross-dataset generalization'}, {'url': 'https://aclanthology.org/2026.findings-acl.1427.pdf#page=9', 'location': '§6 Limitations'}, {'url': 'https://aclanthology.org/2026.findings-acl.1427.pdf#page=22', 'location': 'App. C.2-C.3, Table 6 multimodal results'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zhuang2025doctorLearning to Be A Doctor (medical agent workflow search)skillsSkills, tools and workflow adaptation1offline_benchmarkThe agent workflow graph (nodes, edges, conditional, loop and parallel structures) and node system and human prompts (§3.1-3.3).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).The final evolved workflow, e.g. 11 nodes and 5 branches at test time (Table 2).Diagnostic errors against ground-truth labels, which the LLM classifies as image-understanding or diagnostic errors (§3.3).Offline batched iterations over the training split, validated after each batch until validation accuracy converges; final workflow frozen for testing (§4.1).The LLM analyses failures and emits structural and prompt suggestions, reformulated into structured modifications via templates (§3.3).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).Ablations disabling individual operations (Table 3) and operation levels (Table 4); static IO, CoT and Round Table baselines (Table 1); single seed 42.Held-out test splits: 224 Skin Concepts images (before excluding 100 images for moderation or broken links) and 120 Skin Conditions images (§4.1).not_identifiedNone; the analysis, suggestion and validation procedure is fixed.Small 50/50 training and validation splits, one seed, one synthetic dataset; offline benchmarks only, no clinician or prospective evaluation (§4.1).arXiv v2 (2025-08-15, comment: Accepted at ACM MM 2025); published version not readhttps://arxiv.org/html/2504.11301v2[{'url': 'https://arxiv.org/html/2504.11301v2#S3.SS3', 'location': '§3.3 Workflow Evolution'}, {'url': 'https://arxiv.org/html/2504.11301v2#S4.SS1', 'location': '§4.1 datasets, splits, implementation details'}, {'url': 'https://arxiv.org/html/2504.11301v2#S4.T1', 'location': 'Table 1, top-k accuracy vs baselines'}, {'url': 'https://arxiv.org/html/2504.11301v2#S4.T3', 'location': 'Table 3, operation ablation'}, {'url': 'https://arxiv.org/html/2504.11301v2#S4.T4', 'location': 'Table 4, operation-level ablation'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2025macdMACDmemoryMemory and experience across encounters1retrospectiveDisease-specific Self-Learned Knowledge (general and rare diagnostic concepts) injected into the diagnostician prompt.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.Per-disease, per-model Self-Learned Knowledge text (Supplementary Tables 10-12), also transferred to other LLMs.Ground-truth diagnosis correctness on learning cases; refiner uses change in learning-set diagnostic accuracy when each concept is ablated.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.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.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.Summarizer-only versus summarizer plus refiner ablation and knowledge-composition ablations under the same pipeline; no evolving-versus-fixed improver comparison.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.not_identifiedNot established: summarizer, refiner and selection rules stay fixed; only the knowledge text changes.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.v5https://arxiv.org/abs/2509.20067v5[{'url': 'https://arxiv.org/html/2509.20067v5#Sx5.SSx1', 'location': 'Methods, Dataset construction: 4,390 cases, success-driven learning-set selection'}, {'url': 'https://arxiv.org/html/2509.20067v5#Sx5.SSx4', 'location': 'Methods, Multi-agent clinical diagnosis framework: Eqs. 1-4, summarizer, refiner, causal concept ablation'}, {'url': 'https://arxiv.org/html/2509.20067v5#Sx3.SSx2', 'location': 'Results: SLK vs authoritative knowledge, CoT, few-shot, LoRA, SOTA models (Fig. 2)'}, {'url': 'https://arxiv.org/html/2509.20067v5#Sx3.T2', 'location': 'Table 2: ablations including Sum. Only vs Sum. + Ref.'}, {'url': 'https://arxiv.org/html/2509.20067v5#Sx3.SSx6', 'location': 'Results: MACD-human workflow and physician comparison (Fig. 4)'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zheng2026meddcrMedDCRskillsSkills, tools and workflow adaptation1offline_benchmarkExecutable medical-coding workflow programs (tool and strategy composition) and the archive of scored past designs.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.Archive tuples of plan, compiled workflow, score and reflection; the best-scoring workflow is the released output.Validation micro-F1-based score with guideline-compliance and cost terms, plus Reflector textual critique of intermediate LLM outputs.Offline search for 100 iterations before use; no parameter training.LLM Designer proposes plans conditioned on elite and recent archive entries; Coder compiles them with a self-fixing retry loop.Every candidate is appended to the archive; the final workflow is the argmax validation score over the archive (Alg. 1, line 18).Ablations removing top-k, recent-n, score feedback or text feedback (Table 4); no changing-improver comparison.Not reported; optimisation uses D_val and the paper does not describe a separate test split for Table 1.not_identifiedNone reported; Designer, Reflector and archive-selection rule stay fixed while workflows change (archived workflows do serve as Designer exemplars).Benchmark workflow search; overlap between selection and reporting data cannot be excluded from the text, and no clinical deployment.arXiv v1 (17 Nov 2025) read in full; ACL Anthology PDF checked and Tables 1–4 and Algorithm 1 match.https://arxiv.org/html/2511.13361v1[{'url': 'https://arxiv.org/html/2511.13361v1#S3.SS2', 'location': '§3.2 framework overview and Algorithm 1'}, {'url': 'https://arxiv.org/html/2511.13361v1#S3.SS4', 'location': '§3.4 memory archive and plug-and-play'}, {'url': 'https://arxiv.org/html/2511.13361v1#S4.SS1', 'location': '§4.1 datasets and 100-iteration implementation'}, {'url': 'https://arxiv.org/html/2511.13361v1#S4.SS3', 'location': '§4.3 ablations, Table 4'}, {'url': 'https://arxiv.org/html/2511.13361v1#S6', 'location': '§6 limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
lu2025keaKEA (on PG-Bench)memoryMemory and experience across encounters1offline_benchmarkExperience bank of correct case–answer pairs and reflection bank of corrected cases with self-summarised error analyses.Backbone LLM weights (GPT-4o, DeepSeek-V3, BianCang-7B-Instruct), BGE retriever, static medical knowledge base and prompt templates.Prompt–answer pairs in the experience bank; prompt, both answers and reflection stored only if the revised answer matches the label.Gold department labels on training dialogues; incorrect answers trigger LLM reflection and re-prediction.Offline evolution rounds (1–3, extended to 7 in Table IX) over the 70% training split before test inference.The same frozen task LLM writes reflections on its own errors.Label correctness gates storage; no other release criterion reported.Component ablations (Table VIII) and round-count comparisons; no changing-improver comparison.Held-out 30% test split of PG-Bench; separate high- and low-frequency test sets (Tables V–VI); human evaluation assesses dataset realism only.not_identifiedNone; reflection prompt and storage rule fixed across rounds.Synthetic-dialogue benchmark; best GPT-4o accuracy is 37.91% and gains plateau after round 5; no deployment study.Author manuscript in IEEE template from the UTS OPUS repository (open-access copy listed by Semantic Scholar).https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf[{'url': 'https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf#page=3', 'location': '§III-B PG-Bench construction and 7:3 split (pp. 3–4)'}, {'url': 'https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf#page=4', 'location': '§IV-A evolving knowledge pool'}, {'url': 'https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf#page=5', 'location': '§IV-B dynamic evolution rule'}, {'url': 'https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf#page=7', 'location': 'Tables II–IV round-wise results'}, {'url': 'https://opus.lib.uts.edu.au/bitstream/10453/190868/2/main.pdf#page=9', 'location': '§VII-E/F ablation and extended rounds, Tables VIII–IX (pp. 9–10)'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
ji2026spagentsSpAgentsmemoryMemory and experience across encounters1external_retrospectiveLong-term case memory of ClinicalBERT-encoded patient data with physician diagnostic labels.DeepSeek-Chat Planner/Data/Tool agents, DeepSeek-Reasoner DoctorAgent, SPARCC MRI tool, prompts and top-3 FAISS retrieval.Patient case text plus diagnostic label vectors, with create/read/update/delete management.Physician diagnostic labels submitted through a 'learning' task.Memory updated immediately on feedback; when updates occurred in the reported experiments is not reported.None; labelled cases are stored directly.Not reported beyond storing physician-labelled cases.With versus without memory (Table 3); no changing-improver comparison.Single-centre validation set (n=186) and external test set (n=51) from other hospitals; with/without-memory comparison.not_identifiedNone.Memory ablation gain shown, but memory contents during testing unreported; external hospitals counted as five (Abstract) and six (Methods).Published HTML article and Supplementary Information PDF (algorithms, Fig. S3).https://www.nature.com/articles/s41746-026-02372-4[{'url': 'https://www.nature.com/articles/s41746-026-02372-4#Sec5', 'location': 'Results: Impact of long-term memory, Table 3'}, {'url': 'https://www.nature.com/articles/s41746-026-02372-4#Sec16', 'location': 'Methods: data preparation and cohorts'}, {'url': 'https://www.nature.com/articles/s41746-026-02372-4#Sec18', 'location': 'Methods: Design of DoctorAgent (diagnosis and learning tasks)'}, {'url': 'https://www.nature.com/articles/s41746-026-02372-4#Sec20', 'location': 'Methods: Evaluation of the long-term memory'}, {'url': 'https://static-content.springer.com/esm/art%3A10.1038%2Fs41746-026-02372-4/MediaObjects/41746_2026_2372_MOESM1_ESM.pdf', 'location': 'Supplementary Note 1.4 and Fig. S3 (memory retrieval/update)'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
mokssit2026fhiragentevalFHIR-AgentEval memory configurationsmemoryMemory and experience across encounters1simulatedDual FAISS memory of micro reflections (FHIR resource–operation tips) and macro reflections (task strategy).GPT-4.1-mini ReAct evaluation agent, GPT-4.1 training agent, evaluator and reflector prompts/models, MCP tools, deterministic validators.Reflection lessons indexed for semantic and exact-key retrieval.Deterministic validator outcomes, execution logs and a GPT-4.1 evaluator critique.Separate offline training phase (up to 2 variations and 2 trials per task, early stop on success); memory then frozen.GPT-o4-mini reflector, optionally grounded in the FHIR R4 specification server.Not reported; generated reflections appear to be indexed without a gate.Five configurations varying memory and specification access; no changing-improver comparison.Unseen task variations (Experiments 2–3) and 22 held-out tasks (Experiment 4), three runs each, deterministic validation.not_identifiedNone.Sandbox administrative workflows; held-out transfer gain is smaller than within-task gains; no patient-care evidence.Research Square preprint v1 (DOI 10.21203/rs.3.rs-8746188/v1); published preview page shows abstract only.https://www.researchsquare.com/article/rs-8746188/v1[{'url': 'https://www.researchsquare.com/article/rs-8746188/v1', 'location': 'Methodology: Reflexion long-term memory MCP server'}, {'url': 'https://www.researchsquare.com/article/rs-8746188/v1', 'location': 'Methodology: Reflexion training process (micro and macro reflections)'}, {'url': 'https://www.researchsquare.com/article/rs-8746188/v1', 'location': 'Methodology: four experimental settings, Experiment 4 held-out tasks'}, {'url': 'https://www.researchsquare.com/article/rs-8746188/v1', 'location': 'Results: Figure 5a success rates'}, {'url': 'https://www.researchsquare.com/article/rs-8746188/v1', 'location': 'Discussion: memory effects and specification grounding'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
xing2026clinicalretrialClinicalReTrialmemoryMemory and experience across encounters1simulatedGlobal memory of edit-category success rates and mean rewards, plus within-trial local memory and redesign pool.GPT-4o-mini agents, LightGBM outcome predictor used as reward, RAG safety validator, beam search and prompts.Edit-category priors (success rate, mean reward) used for confidence calibration and generation-budget allocation.Change in predicted success probability, attributed to individual augmentations as marginal rewards.Within-trial over 5 iterations; global memory updated after each trial converges.LLM synthesis distils each trial's reward distribution into the global pool.Only safety-validated augmentations receive rewards; no further release gate reported.Paired ablations without memory, pool or analysis (Table 3); no changing-improver comparison.60 failed trials from the held-out TrialBench test split; HINT re-scoring never used in optimisation (App. C.2); 8 retrospective case studies.not_identifiedNone.Gains are predicted probabilities; order of global-memory accumulation and separation of global from local memory are not reported.arXiv v3 (13 Sep 2026) read in full; ACL Anthology PDF checked for Table 3 and Appendix C.2.https://arxiv.org/html/2601.00290v3[{'url': 'https://arxiv.org/html/2601.00290v3#S3.SS3.SSS3', 'location': '§3.3.3 cross-trial global memory'}, {'url': 'https://arxiv.org/html/2601.00290v3#A1', 'location': 'Appendix A Algorithm 1'}, {'url': 'https://arxiv.org/html/2601.00290v3#S4.SS3', 'location': '§4.3 Tables 2–3 and ablations'}, {'url': 'https://arxiv.org/html/2601.00290v3#A3.SS2', 'location': 'Appendix C.2 cross-predictor HINT evaluation'}, {'url': 'https://arxiv.org/html/2601.00290v3#Sx1', 'location': 'Limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
sayeedi2026r4R4AgentmemoryMemory and experience across encounters1offline_benchmarkExemplar memory of task type, specialization, cue, final report and tags from processed cases.Frozen VLM backbones (Gemini-2.5-Flash, MedGemma-4B, Qwen2.5-VL-7B), router rules, reflector/repairer prompts, lexical-overlap retrieval.Self-generated final reports with cues and tags, reused as few-shot exemplars.The system's own Reflector critique and pass@k draft scoring; no ground-truth feedback described for memory writes.Memory updated after each processed case during inference.Cue and tag functions applied to the final repaired output.Outputs judged not to be a system error are stored; the criterion is not further specified.Not reported; only a pass@k ablation (Fig. 2).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.not_identifiedNone.Reported gains bundle routing, reflection, repair and memory; stored exemplars are unverified self-generated outputs.arXiv v2 (8 Mar 2026).https://arxiv.org/html/2601.08192v2[{'url': 'https://arxiv.org/html/2601.08192v2#S3.SS2.SSSx2', 'location': '§3.2 Exemplar Memory'}, {'url': 'https://arxiv.org/html/2601.08192v2#S3.SS3', 'location': '§3.3 Self-Improvement via Exemplar Curation'}, {'url': 'https://arxiv.org/html/2601.08192v2#S4.SS1', 'location': '§4.1 Datasets'}, {'url': 'https://arxiv.org/html/2601.08192v2#S5', 'location': '§5 Results, Table 1 and Fig. 2 pass@k ablation'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
liao2026agentehrAgentEHR / RetroSummemoryMemory and experience across encounters1offline_benchmarkExternal Experience Memory Bank of actor heuristics and summarizer salience guidelines keyed by EHR embeddings.LLM backbone weights (no parameter updates), RetroSum actor and summarizer prompts, MCP toolbox of 19+ tools, bge-m3 retriever with top-1 retrieval.Triplets (EHR embedding, actor experience, summarizer experience), one per training instance.Ground-truth answer sets of training instances compared with the agent's trajectory, summary and predictions by a reflection prompt.Offline before evaluation on 100 MIMIC-IV-Common training examples per task; the same accumulated memory is then reused on all three test subsets.Reflection module that shares the agent's frozen LLM, prompted to extract procedural tool-use heuristics and information-salience guidelines.No filtering, validation or pruning of generated experiences is reported.No comparison of a changing versus a fixed improvement procedure.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.not_identifiedNone; only memory contents change.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.arXiv v1 (2026-01-20), cross-checked against EMNLP 2026 PDFhttps://arxiv.org/html/2601.13918v1[{'url': 'https://arxiv.org/html/2601.13918v1#S3.SS3', 'location': 'Sec. 3.3 Evolving Strategy: experience generation, memory construction, evolved inference'}, {'url': 'https://arxiv.org/html/2601.13918v1#A4.SS2', 'location': 'App. D.2: 100 training examples per task, same memory for all subsets, top-1 bge-m3 retrieval'}, {'url': 'https://arxiv.org/html/2601.13918v1#S3.T2', 'location': 'Table 2: MIMIC-IV-Common results, evolved vs non-evolved, five backbones'}, {'url': 'https://arxiv.org/html/2601.13918v1#S4.T3', 'location': 'Table 3: OOD subsets MIMIC-IV-Rare and MIMIC-III'}, {'url': 'https://arxiv.org/html/2601.13918v1#S4.T4', 'location': 'Table 4: ablation of retrospect and evolved components'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zhu2026ultrasoundagentsUltrasoundAgentsgeneratorsParameter updates and changing training generators2-implicitoffline_benchmarkWeights of the attribute sub-agent and the main diagnostic agent (Qwen2.5-VL-3B).Backbone architecture, crop-and-zoom tool, BUS-CoT attribute taxonomy, reward definitions, GRPO and SFT hyperparameters.Trained sub-agent weights, distilled main-agent weights, and the refined trajectory set used for SFT.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).Three sequential offline training stages before testing; agents fixed at inference.Stage-2 policy rollouts; the Stage-2 model rewrites rationales of misdiagnosed samples conditioned on ground truth; predicted boxes replaced by ground-truth boxes.All sampled trajectories are refined and used; no further selection or validation gate is reported.No comparison of a changing versus a fixed improvement procedure.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.not_identifiedThe trained Stage-2 policy also rewrites its successor's SFT data; this role is not isolated from box replacement in the ablation.Single 3B backbone and single runs without variance; BI-RADS agreement low (overall kappa 0.224, OOD 0.037); no learning after deployment.arXiv v1 (2026-03-11)https://arxiv.org/html/2603.10852v1[{'url': 'https://arxiv.org/html/2603.10852v1#S2.SS2', 'location': 'Sec. 2.2: three-stage decoupled training and corrective self-distillation'}, {'url': 'https://arxiv.org/html/2603.10852v1#S3.SS1', 'location': 'Sec. 3.1: datasets, splits, implementation and baselines'}, {'url': 'https://arxiv.org/html/2603.10852v1#S3.T1', 'location': 'Table 1: main diagnostic performance, in-domain and BrEaST OOD'}, {'url': 'https://arxiv.org/html/2603.10852v1#S3.T2', 'location': 'Table 2: ablation and upper-bound analysis'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026theraagentTheraAgentmemoryMemory and experience across encounters1retrospectiveSEA-Mem case memory (extracted feature profile and outcome per patient) and subgroup outcome-rate patterns.GPT-4o radiologist, biochemist, oncologist, integrator and reasoning agents and prompts; 23-publication trial knowledge base; feature-matching retrieval.Patient feature profiles with outcomes; empirical outcome rates for feature combinations shared by three or more cases.Documented treatment outcomes (PSA response, overall survival beyond 12 months).Described as updating when outcomes become available; the order and timing of memory population in the reported experiments are not reported.Rule-based: cases appended and subgroup rates recomputed; no LLM-generated memory content.Patterns formed only when at least three cases share a feature combination; no other gate reported.No comparison of a changing versus a fixed improvement procedure.35 real patients and 400 synthetic cases; single-LLM, MedAgents, MDAgents and MedAgent-Pro baselines; w/o SEA-Mem ablation on real data.not_identifiedNone; the update rule is fixed.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.arXiv v1 (2026-03-14)https://arxiv.org/html/2603.13676v1[{'url': 'https://arxiv.org/html/2603.13676v1#S2.SS3', 'location': 'Sec. 2.3: SEA-Mem structure, pattern learning, retrieval, update'}, {'url': 'https://arxiv.org/html/2603.13676v1#S3.SS1', 'location': 'Sec. 3.1: real and synthetic data, baselines'}, {'url': 'https://arxiv.org/html/2603.13676v1#S2.T1', 'location': 'Table 1: real-patient results'}, {'url': 'https://arxiv.org/html/2603.13676v1#S3.T2', 'location': 'Table 2: synthetic results by complexity'}, {'url': 'https://arxiv.org/html/2603.13676v1#S3.T3', 'location': 'Table 3: ablation including w/o SEA-Mem'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
liu2026openhospitalOpenHospitalmemoryMemory and experience across encounters1simulatedAccumulated clinical experience of 38 physician agents produced by post-case reflection; storage form not specified.Qwen3-Next-80B-A3B weights for physician and patient agents; Agent-Kernel action space; synthetic patient records and personas.Diagnostic, investigative and treatment reflections described as accumulated clinical experience; format and retrieval not described.Ground-truth diagnosis, gold examination list and guideline-based treatment comparison used in post-case self-critique.After each case, across 22 sequential training batches of about 500 cases.Physician agents' own multi-dimensional self-critique.Not reported.Not reported.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.not_identifiedNone reported.Fully synthetic, text-only arena; mechanism under-specified; no no-reflection control or repeated runs; authors state it is not for clinical decision support.arXiv v3 (2026-05-31)https://arxiv.org/html/2603.14771v3[{'url': 'https://arxiv.org/html/2603.14771v3#S4.SS2', 'location': 'Sec. 4.2: dataset split, 22 batches, baseline reflection mechanism'}, {'url': 'https://arxiv.org/html/2603.14771v3#S4.SS3', 'location': 'Sec. 4.3: evolution results'}, {'url': 'https://arxiv.org/html/2603.14771v3#S4.F4', 'location': 'Figure 4: performance trajectories across batches'}, {'url': 'https://arxiv.org/html/2603.14771v3#S3.SS1', 'location': 'Sec. 3.1: synthetic patient data construction'}, {'url': 'https://arxiv.org/html/2603.14771v3#Sx1', 'location': 'Limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026skingptxSkinGPT-XmemoryMemory and experience across encounters1offline_benchmarkEvoDerma-Mem graph database of historical cases (image embedding, key findings, expert diagnosis) and versioned per-disease diagnostic guidelines.Qwen3-VL vision agent, fine-tuned disease-diagnosis model, Qwen3 case-review and summarization agents and prompts, Oxford-handbook RAG store, image feature extractor.Case triplets and guideline versions G_C^t per disease.Expert diagnosis labels of stored historical (training) cases.Guidelines rewritten when new samples reach threshold N_thresh; the N_thresh value and whether test cases are added are not reported.LLM Summarization Agent that merges the previous guideline with new key findings.Threshold-triggered update with contradiction resolution inside the summarizer; no validation gate before use reported.No comparison of a changing versus a fixed improvement procedure.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.not_identifiedNone; the summarizer procedure is fixed (each guideline update takes the previous version as input).Offline public benchmarks; memory built from labelled training data; RSDD images gathered from websites; no prospective use.arXiv v1 (2026-03-27)https://arxiv.org/html/2603.26122v1[{'url': 'https://arxiv.org/html/2603.26122v1#S4.SS3', 'location': 'Sec. IV-C: EvoDerma-Mem storage and guideline evolution'}, {'url': 'https://arxiv.org/html/2603.26122v1#S4.SS4.SSS3', 'location': 'Sec. IV-D3: case review with retrieved cases and evolved guidelines'}, {'url': 'https://arxiv.org/html/2603.26122v1#S2.SS5', 'location': 'Sec. II-E: memory ablation and dermatologist review'}, {'url': 'https://arxiv.org/html/2603.26122v1#S2.F5', 'location': 'Figure 5: memory ablation, guideline evolution timeline, physician ratings'}, {'url': 'https://arxiv.org/html/2603.26122v1#S4.SS2', 'location': 'Sec. IV-B: datasets and splits'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026seaSEAmemoryMemory and experience across encounters2-implicitoffline_benchmarkQwen3-4B/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).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).Outcome-annotated short-term case records and natural-language diagnostic rules consolidated from evicted cases, carried forward across stream rounds (Sec 3.1).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).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).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).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).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).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).not_identifiedRL 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.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.arXiv v2, 2026-08-07 (v1 2026-04-08)https://arxiv.org/html/2604.07269v2[{'url': 'https://arxiv.org/html/2604.07269v2#S3.SS1', 'location': 'Sec 3.1: dual-memory design and memory operations as agent actions'}, {'url': 'https://arxiv.org/html/2604.07269v2#S3.SS2', 'location': 'Sec 3.2-3.3: diagnostic and memory rewards, round-wise schedule, GRPO'}, {'url': 'https://arxiv.org/html/2604.07269v2#S4.T2', 'location': 'Table 2: ER-Reason long-horizon final accuracy and Delta Acc@50/@100'}, {'url': 'https://arxiv.org/html/2604.07269v2#A6.T4', 'location': 'App F.2.3 and Table 4: standard vs long-horizon protocols, no test carryover in standard setting'}, {'url': 'https://arxiv.org/html/2604.07269v2#S5.SS1', 'location': 'Sec 5.1 and App G: expert evaluation by one physician'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
jeong2026healthcaregymHealthcare AI GYM (TT-OPD)generatorsParameter updates and changing training generators2-implicitoffline_benchmarkFull 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).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).Trained policy checkpoint, evaluated with the same multi-turn AgentRunner, tools and knowledge base (Sec 5.2).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).Online RL during training (60 steps on 8 A100 80GB); no updates at evaluation (Sec 5.1).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).Not reported; results given for TT-OPD at step 60 and GRPO peak at step 55 on the validation set (Sec 5.3).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).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).not_identifiedThe 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).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.arXiv v1, 2026-05-01https://arxiv.org/html/2605.02943v1[{'url': 'https://arxiv.org/html/2605.02943v1#S3', 'location': 'Sec 3: environment design, tools, 5D reward'}, {'url': 'https://arxiv.org/html/2605.02943v1#S4.SS2', 'location': 'Sec 4.2: TT-OPD, EMA teacher, outcome-conditioned hints'}, {'url': 'https://arxiv.org/html/2605.02943v1#S5.T1', 'location': 'Table 1: 18 benchmarks, Base (text), Base+AR, GRPO, TT-OPD'}, {'url': 'https://arxiv.org/html/2605.02943v1#S6.SS1', 'location': 'Sec 6.1 and Fig 4: OPD failure progression ablation'}, {'url': 'https://arxiv.org/html/2605.02943v1#A1.SS2', 'location': 'App A.2: task sources, AutoTaskGenerator conversion of MedQA/MedMCQA/MMLU/VQA/MIMIC'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
ossowski2026codeclinicCodeClinic (clinical autoformalization)skillsSkills, tools and workflow adaptation1retrospectiveShared 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).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).Library of verified concept functions exposed as signatures; later concept builds can search and load earlier verified functions (Sec 4.1, App G.1).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).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).ReAct autoformalization agent (Qwen3.5-27B) that explores the schema, retrieves guidelines and writes code (Sec 4.1).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).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).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).not_identifiedNot reported; the build-and-verify procedure stays fixed.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.arXiv v1, 2026-05-10https://arxiv.org/html/2605.09675v1[{'url': 'https://arxiv.org/html/2605.09675v1#S4', 'location': 'Sec 4: autoformalization pipeline (tool generation, verification, inference)'}, {'url': 'https://arxiv.org/html/2605.09675v1#S5.T3', 'location': 'Table 3: zero-shot vs GRPO vs autoformalization on held-out compositional test set'}, {'url': 'https://arxiv.org/html/2605.09675v1#S5.F6', 'location': 'Fig 6: longitudinal step accuracy, zero-shot vs autoformalized library'}, {'url': 'https://arxiv.org/html/2605.09675v1#S3.SS2', 'location': 'Sec 3.2: concepts, dependency levels, 10%/90% split'}, {'url': 'https://arxiv.org/html/2605.09675v1#A1', 'location': 'App A: limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wu2026sepsisagentSepsisAgentgeneratorsParameter updates and changing training generators1retrospectiveFull 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).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).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).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).Offline staged training; no updates during evaluation (Sec 4, Sec 5.1).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).Not reported beyond terminating GRPO after 300 steps (App G.2).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).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).not_identifiedNot reported.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.arXiv v1, 2026-05-14https://arxiv.org/html/2605.14723v1[{'url': 'https://arxiv.org/html/2605.14723v1#S4.SS4', 'location': 'Sec 4.4: world-model-based agentic RL, rollouts and composite reward'}, {'url': 'https://arxiv.org/html/2605.14723v1#S5.T3', 'location': 'Table 3: OPE, guideline adherence and unsafe actions vs baselines'}, {'url': 'https://arxiv.org/html/2605.14723v1#S5.T4', 'location': 'Table 4: cumulative stage ablation and IHM/VR prediction without simulator'}, {'url': 'https://arxiv.org/html/2605.14723v1#A9.T11', 'location': 'App I, Table 11: MIMIC-III cross-dataset OPE'}, {'url': 'https://arxiv.org/html/2605.14723v1#A7', 'location': 'App G: training data, settings and reward coefficients'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026vdssVDSScomparatorsDiagnostic training and informative adaptation comparators1external_retrospectivePer-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).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).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).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).Online, one bandit update at the close of each adjustment cycle (Algorithm 1).Note Generator LLM extracts the preference signal; BanditUpdate rule, whose algorithm is not specified (Sec 2.3).Updates only from clinician-accepted outcomes, not from rejected proposals; no further gating reported (Sec 2.3).Single-model end-to-end generation with the same backbones; NoPref, NoImg and NoImgNoPref ablations within VDSS (Table 1).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).not_identifiedNot reported.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.arXiv v1, 2026-05-22 (examined); LNCS version not examinedhttps://arxiv.org/html/2605.23320v1[{'url': 'https://arxiv.org/html/2605.23320v1#S2.SS2', 'location': 'Sec 2.2 and Algorithm 1: cycle-end preference update and long-term memory writes'}, {'url': 'https://arxiv.org/html/2605.23320v1#S2.SS3', 'location': 'Sec 2.3: contextual bandit over 12 preference categories'}, {'url': 'https://arxiv.org/html/2605.23320v1#S3.T1', 'location': 'Table 1: replay metrics, clinician ratings, NoPref/NoImg ablations'}, {'url': 'https://arxiv.org/html/2605.23320v1#S3.F2', 'location': 'Fig 2: regret over 100 cycles from a single clinician'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zeng2026trajevolveTraj-EvolvememoryMemory and experience across encounters2-implicitretrospectiveA 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).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.ExPool vector store of (summary embedding, manager rationale, predicted risk, label) for 13,629 training patients, plus fine-tuned worker and manager adapters.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).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).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).No accept/reject gate on updates; hyperparameters and thresholds set on a validation split; single training round released (Appendix B).Static zero-shot Traj-CoA and RAG with the same backbone, plus ExPool-only, MARL-only and combined variants (Tables 1-2).Disjoint held-out test sets from the same institution (overall n=1,000; never-smoker n=835), bootstrap 1,000 resamples (Sec 4).not_identifiedNone; rejection rule, reward and training recipe are fixed.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).arXiv v1 (2026-06-01); no journal_ref or acceptance commenthttps://arxiv.org/html/2606.02812v1[{'url': 'https://arxiv.org/html/2606.02812v1#S3.SS3', 'location': 'Sec 3.3 Evolving Experience Pool'}, {'url': 'https://arxiv.org/html/2606.02812v1#S3.SS4', 'location': 'Sec 3.4 Multi-Agent Reinforcement Learning'}, {'url': 'https://arxiv.org/html/2606.02812v1#S5.T1', 'location': 'Table 1 overall population results'}, {'url': 'https://arxiv.org/html/2606.02812v1#S5.F4', 'location': 'Fig 4 and Sec 5.2 ExPool size dynamics'}, {'url': 'https://arxiv.org/html/2606.02812v1#S6', 'location': 'Sec 6 Conclusion and Limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
zhang2026medevoevalMedEvoEvalmemoryMemory and experience across encounters1simulatedThe doctor agent's external memory bank of success, failure-correction and action-rule cards, retrieved by TF-IDF (top K=3) before later decisions.Doctor model (Qwen-Turbo), patient simulator (Qwen-Turbo), manager scorer (Kimi-K2.5), test catalog, rubrics, case order, action budget of 10.Unbounded memory bank; 4,500 records after E3 (312 success, 1,188 corrective, 3,000 rule cards) (Appendix B.4).Manager-side rubric scores of diagnosis, evidence and management after FINALIZE, with access to the reference diagnosis (Sec 3.3-3.4).Online sequential write-back after each episode over a 500-case stream (passes E1-E3) and continued during held-out evaluation (H3).The manager LLM writes compact IF-THEN and IF-AVOID cards after scoring.No acceptance gate or eviction; every card is written back.No-memory conditions E0 and H0 with identical doctor, simulator, order and scorer; frozen-bank relevant vs case-mismatched retrieval control (Table 27).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.not_identifiedNone.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).arXiv v3 (2026-10-18), author comment: camera-ready accepted at NeurIPS 2026 Evaluations and Datasets Trackhttps://arxiv.org/html/2606.28900v3[{'url': 'https://arxiv.org/html/2606.28900v3#S3.SS4', 'location': 'Sec 3.4 Experience write-back'}, {'url': 'https://arxiv.org/html/2606.28900v3#S5.T4', 'location': 'Table 4 longitudinal, held-out and BWT results'}, {'url': 'https://arxiv.org/html/2606.28900v3#A4.T21', 'location': 'Table 21 paired H3-H0 held-out endpoints'}, {'url': 'https://arxiv.org/html/2606.28900v3#A4.T27', 'location': 'Table 27 frozen-bank retrieval control'}, {'url': 'https://arxiv.org/html/2606.28900v3#S6', 'location': 'Sec 6 Discussion, limitations and ethics'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
meng2026topoagentTopoAgentmemoryMemory and experience across encounters1offline_benchmarkLong-term memory of structured diagnostic entries (failed descriptor, diagnosed cause, successful correction), accumulated within a dataset and reset between datasets.GPT-4o backbone at temperature 0, 21 perception and descriptor tools, distilled skill set (properties, tiered rankings, parameters), Perception-Reasoning-Action-Reflection pipeline, classifiers.Per-dataset long-term memory entries consulted by later Determination and Reflection steps.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).Online during runs; an entry is written whenever Reflection fails and triggers a retry.The LLM Reflect step writes the diagnosis and correction (Algorithm 1, line 17).No gate; entries are written on failure without verification.Ablations without dual memory and without long-term memory under the same pipeline (Table 2).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.not_identifiedNone.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).arXiv v1 (2026-06-29)https://arxiv.org/html/2606.29763v1[{'url': 'https://arxiv.org/html/2606.29763v1#S3.SS3', 'location': 'Sec 3.3 Dual Memory and Algorithm 1'}, {'url': 'https://arxiv.org/html/2606.29763v1#S4.T1', 'location': 'Table 1 main results'}, {'url': 'https://arxiv.org/html/2606.29763v1#S4.T2', 'location': 'Table 2 ablation'}, {'url': 'https://arxiv.org/html/2606.29763v1#S5', 'location': 'Sec 5 Conclusions and limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
qu2026sageagentSAGEAgentmemoryMemory and experience across encounters1offline_benchmarkEpisodic memory of the agent's own stage decisions with rewards, and a semantic memory of up to 10 stage-indexed stop/acquire rules.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.Episode store and active rule set; four rules recurred in at least 4 of 5 folds (Table 3).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).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).The frozen LLM's reflection module proposes rules from an effectiveness report.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).Component ablation from base LLM through tools, episodic and semantic memory with the same frozen LLM (Table 2); Reflexion baseline (Table 1).Nested 5x5 cross-validation on 170 complete-modality patients with outer-fold test patients; whether memory keeps updating on test patients is not reported.not_identifiedNone.Small public cohort (170 evaluation patients); rewards use model uncertainty and burden rather than outcomes; C-index slightly below full-workup fusion.arXiv v1 (2026-07-10) examined; published in MICCAI 2026, LNCS pp. 536-546 (Crossref)https://arxiv.org/html/2607.09521v1[{'url': 'https://arxiv.org/html/2607.09521v1#S2.SS3', 'location': 'Sec 2.3 episodic and semantic memory'}, {'url': 'https://arxiv.org/html/2607.09521v1#S3.T1', 'location': 'Table 1 core results'}, {'url': 'https://arxiv.org/html/2607.09521v1#S3.T2', 'location': 'Table 2 component ablation'}, {'url': 'https://arxiv.org/html/2607.09521v1#S3.T3', 'location': 'Table 3 discovered rules and Sec 3.5'}, {'url': 'https://doi.org/10.1007/978-3-032-38059-3_51', 'location': 'MICCAI 2026 LNCS version, pp. 536-546'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026curaCura 1TgeneratorsParameter updates and changing training generators2-explicitoffline_benchmarkThe post-training data mixture and resulting Kimi-K2.6 (1T) weights, trained per round by SFT screen, RL and self-distillation fine-tuning.Base model, benchmark harnesses (one MedAgentBench harness bug fix counted as a round), data-synthesis skill set, training stack, human review step.Validated data refinements accumulated into a consolidated mixture; kept checkpoints serve as the next round's start; released Cura 1T weights.Graded benchmark trajectories and scores on MedAgentBench, HealthBench Professional/Hard, MedXpertQA and AgentClinic, categorized into failure modes (Sec 3.2).Offline rounds per capability loop (1-5 rounds shown), then consolidation of the accumulated mixture into the released model (Sec 4).A training agent (model identity not reported) using reasoning-correction (frontier-model explanations), knowledge-injection, behavior-calibration, retention-anchor and mixture-curation skills.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).Base Kimi-K2.6 and frontier comparators only; authors state isolating contributions or further ablations was beyond capability (Limitations).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.not_identifiedNot tested.Benchmarks guide keep/revert and report results; human-in-the-loop decisions; no ablations; not evidence of clinical safety (Sec 6).arXiv v3 (2026-10-05); no journal_refhttps://arxiv.org/html/2607.15314v3[{'url': 'https://arxiv.org/html/2607.15314v3#S3.SS2', 'location': 'Sec 3.2 Recursive self-improvement loop'}, {'url': 'https://arxiv.org/html/2607.15314v3#S4.T2', 'location': 'Table 2 MedAgentBench rounds'}, {'url': 'https://arxiv.org/html/2607.15314v3#S4.T4', 'location': 'Table 4 MedXpertQA rounds'}, {'url': 'https://arxiv.org/html/2607.15314v3#S4.SS6', 'location': 'Sec 4.6 Improvement across benchmarks'}, {'url': 'https://arxiv.org/html/2607.15314v3#S6', 'location': 'Sec 6 Conclusion, Limitations, Safety'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wang2026miraMIRAmemoryMemory and experience across encounters2-implicitoffline_benchmarkQwen3-VL-8B policy weights (SFT then GRPO) and a global textual reflection memory of diagnostic principles injected into the system prompt.Six-tool suite, fixed Seed-series reward judges and prompts, reward coefficients, validation set, static quality gate, edit-budget schedule.Accepted memory saved with each policy checkpoint; released MIRA-VL-8B weights, trajectories and tool environment.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).Online during RL (750 steps, about one epoch); memory update after rollout rounds with at least 4 unique failed cases.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.Candidate memory accepted only if mean validation reward under the frozen updated policy improves; rejected edits stored to steer later proposals (Appendix E).RL ablation with format+accuracy, plus consistency, plus reflection memory (Table 4); validation curve with and without memory (Fig A7).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.not_identifiedNone; the gate and editor are fixed, only the edit budget follows a preset schedule.Benchmark VQA rather than clinical workflow; single runs; memory's role inside the reward not isolated; authors note limited memory generalization (Sec 5).arXiv v1 (2026-08-11), PDF onlyhttps://arxiv.org/pdf/2608.10827v1[{'url': 'https://arxiv.org/pdf/2608.10827v1#page=9', 'location': 'Sec 3.2-3.3 composite reward and on-policy reflection memory'}, {'url': 'https://arxiv.org/pdf/2608.10827v1#page=10', 'location': 'Table 1 main results (p. 10)'}, {'url': 'https://arxiv.org/pdf/2608.10827v1#page=13', 'location': 'Table 4 RL ablation (p. 13)'}, {'url': 'https://arxiv.org/pdf/2608.10827v1#page=30', 'location': 'Appendix E online reflection memory details (pp. A13-A16)'}, {'url': 'https://arxiv.org/pdf/2608.10827v1#page=14', 'location': 'Sec 5 limitations (p. 14)'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
liu2026batBaT (Benchmark-as-Teacher)generatorsParameter updates and changing training generators2-explicitoffline_benchmarkQwen3.5-4B/9B policy weights (SFT cold start then multi-round GRPO) and the per-round stage curriculum.OpenHands execution environment with public stage skills, pool proportions, LLM rubric verifier, controller evaluation, fallback rule.Best retained checkpoint, plus per-round records of states, mixtures, reward version and controller scores.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).Offline rounds (about ten) alternating GRPO blocks with re-evaluation (Fig 1c, Sec 4.2).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.Best-so-far checkpoint retained; fallback after three consecutive drops or KL shift above a threshold; best checkpoints auto-selected (Sec 2.2.2).Same SFT start trained by plain GRPO with a single final reward, and Stage Bank pool ablations (Table 2, Fig 5).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.not_identifiedNone.Benchmark participates in adaptation; authors say differences describe observed scores rather than a matched ablation (Sec 4.1); research workflows, not patient care.arXiv v1 (2026-08-17)https://arxiv.org/html/2608.16211v1[{'url': 'https://arxiv.org/html/2608.16211v1#S2.SS1', 'location': 'Sec 2.1 Stage Bank'}, {'url': 'https://arxiv.org/html/2608.16211v1#S2.SS2', 'location': 'Sec 2.2 BiCuRL outer and inner loops'}, {'url': 'https://arxiv.org/html/2608.16211v1#S4.T2', 'location': 'Table 2 policy-level results'}, {'url': 'https://arxiv.org/html/2608.16211v1#S4.F5', 'location': 'Fig 5 pool ablation'}, {'url': 'https://arxiv.org/html/2608.16211v1#S5.SS2', 'location': 'Sec 5.2 Limits'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
murugesan2026amrAMRmemoryMemory and experience across encounters1offline_benchmarkPer-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).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).Role-specific memory entries and corrective reflection entries in FAISS, retrieved top-5 per agent (Sec III-C, IV-B).Ground-truth multiple-choice label; reflection triggers only when the prediction differs from the label (Sec III-C, Alg. 1).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.The same GPT-4o agents produce answers and reasoning summaries; a reflection step writes corrective feedback against the label (Sec III-C).No filtering, pruning or validation of memory entries; the authors name pruning and confidence-based retention as future work (Sec V).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.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).not_identifiedNone; the reflection rule and memory format stay fixed.Retrospective exam benchmarks only; Ethical Overseer not clinician-validated; source of test-time memory contents unreported; no uncertainty estimates (Sec IV-B, V).arXiv v1, 2026-08-19https://arxiv.org/html/2608.19029v1[{'url': 'https://arxiv.org/html/2608.19029v1#S3.SS3', 'location': 'Sec III-C: agent-specific memory and reflective update'}, {'url': 'https://arxiv.org/html/2608.19029v1#alg1', 'location': 'Algorithm 1: reflective feedback learning loop'}, {'url': 'https://arxiv.org/html/2608.19029v1#S4.SS2', 'location': 'Sec IV-B: frozen retrieval corpus, no test write-back'}, {'url': 'https://arxiv.org/html/2608.19029v1#S4.T4', 'location': 'Table IV: memory, feedback and RAG ablation'}, {'url': 'https://arxiv.org/html/2608.19029v1#S4.T5', 'location': 'Table V: comparison with GPT-4, MDAgents, MedAgents'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wu2026mediskillevoMediSkill-EvoskillsSkills, tools and workflow adaptation1simulatedFour typed external banks: Clinical Skills, Process Rules, Symbolic Schemas and Measurement Skills (visual procedures); backbone weights untouched (Sec 2.1, 2.2).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).Versioned bank artifacts with content, applicability scope, trajectory provenance and lifecycle status, published as immutable hashed snapshots (Sec 2.2, S1, Table S3).Post-encounter evaluation of each completed training trajectory against reference diagnosis and benchmark targets, then reflection (Sec 2.2, Table S1).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).An LLM reflector proposes typed add, merge, patch, deprecate or discard operations after each trajectory (Sec 2.2, S2.1, S2.2).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).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.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).not_identifiedNone; reflector, merge and validation operators stay fixed.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).arXiv v2, 2026-08-25https://arxiv.org/html/2608.23397v2[{'url': 'https://arxiv.org/html/2608.23397v2#S2.SS2', 'location': 'Sec 2.2: four-bank self-evolution, merge and validation'}, {'url': 'https://arxiv.org/html/2608.23397v2#S3.T4', 'location': 'Table 4: FullChain results on 300 encounters, two backbones'}, {'url': 'https://arxiv.org/html/2608.23397v2#S3.T5', 'location': 'Table 5: controlled-stress aggregates on 180 conditions'}, {'url': 'https://arxiv.org/html/2608.23397v2#S3.SS7', 'location': 'Table 7: no-memory and bank-removal ablation'}, {'url': 'https://arxiv.org/html/2608.23397v2#S3.T3a', 'location': 'Table S3: comparator ledger and frozen object counts'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
dai2026cdegCDEGmemoryMemory and experience across encounters1simulatedA Counterfactual Diagnostic Evidence Graph of diagnosis, evidence and acquisition-action nodes plus diagnosis-revision edges (Graph Consolidation).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).Frozen graph with 1,156 nodes and 1,426 relations, including 270 diagnosis-revision edges (Table 6).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).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).Failed-successful trajectory pairing (cosine > 0.7), UMLS evidence-atom mapping and counterfactual verification by GPT-5.5 (Implementation Details).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).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.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).not_identifiedNone; pairing, mapping and verifier stay fixed; online graph evolution is named as future work (Conclusion).Retrospective records in an interactive environment; LLM-judged accuracy; single construction pass; Claude-Sonnet-5 ID accuracy is below Trajectory RAG (Table 2).arXiv v3, 2026-10-14https://arxiv.org/html/2608.22899v3[{'url': 'https://arxiv.org/html/2608.22899v3#Sx3.SSx2', 'location': 'Methods: candidate mining, counterfactual validation, graph consolidation'}, {'url': 'https://arxiv.org/html/2608.22899v3#Sx4.SSx1.SSS0.Px4', 'location': 'Implementation Details: construction agents, frozen shared graph'}, {'url': 'https://arxiv.org/html/2608.22899v3#Sx4.T2', 'location': 'Table 2: accuracy across four backbones, ID and OOD'}, {'url': 'https://arxiv.org/html/2608.22899v3#Sx4.T3', 'location': 'Table 3: ablations incl. without counterfactual validation'}, {'url': 'https://arxiv.org/html/2608.22899v3#A2.T6', 'location': 'Tables 6-7: construction and validation statistics'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
ghaffari2026learnactcoderLearnActCodermemoryMemory and experience across encounters1retrospectiveMistakeKDB, a structured JSON error memory; false-negative lessons go to the Coder prompt, false-positive lessons to the Judge prompt (Sec 3.2).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).Lessons with error category, implicated codes, explanation and transferable lesson, consolidated from 1,028 MIMIC-III LEARN mismatches (Sec 3.2, 5.4).Mismatches between predictions and reference administrative codes from MIMIC tables (false positives and false negatives) (Sec 3.2, 4).Offline LEARN phase on 25 labelled notes per dataset, then memory reused unchanged in the ACT phase (Sec 3.2, 4, 5.3).A post-hoc LLM MistakeAnalyzer that sees note, predictions and reference codes and assigns one of six error categories (Sec 3.2).Similar lessons are consolidated and frequent patterns prioritized before prompt injection; no separate validation gate (Sec 3.2).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.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).not_identifiedNone; MistakeAnalyzer and routing rule stay fixed.Retrospective single-system study; one model family; administrative labels as proxy; taxonomy not coder-validated; CPT/HCPCS performance remains low (Sec 6).arXiv v1, 2026-09-17https://arxiv.org/html/2609.19721v1[{'url': 'https://arxiv.org/html/2609.19721v1#S3.SS2', 'location': 'Sec 3.2: Learn-Then-Act error memory and role-aware routing'}, {'url': 'https://arxiv.org/html/2609.19721v1#S5.T1', 'location': 'Table 1: matched 150-note MIMIC-III ablation with CIs'}, {'url': 'https://arxiv.org/html/2609.19721v1#S5.T2', 'location': 'Table 2: matched 250-note MIMIC-IV ablation'}, {'url': 'https://arxiv.org/html/2609.19721v1#S5.T3', 'location': 'Table 3: memory-format comparison'}, {'url': 'https://arxiv.org/html/2609.19721v1#S6.SS0.SSS0.Px4', 'location': 'Sec 6: limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
shao2026caseCASEgeneratorsParameter updates and changing training generators2-implicitoffline_benchmarkFull weights of one shared Qwen3-VL-8B-Instruct policy acting as Radiologist, Anatomist and Consultant, including the visual pathway (Sec 5).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).Post-trained CASE-8B policy weights.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).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).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).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.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.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).not_identifiedThe privileged self-teacher is the evolving policy itself and is refreshed after updates (Sec 5.4); rubric and reward weights stay fixed.Single cohort; labels may be unobservable from images; generated evidence and LLM judges are fallible; utility beyond this cohort untested (Sec 7).arXiv v3, 2026-10-14https://arxiv.org/html/2609.39566v3[{'url': 'https://arxiv.org/html/2609.39566v3#S5.SS3', 'location': 'Sec 5.3: on-policy interaction with replayable environment'}, {'url': 'https://arxiv.org/html/2609.39566v3#S5.SS5', 'location': 'Sec 5.5: rubric feedback, composite reward, GRPO'}, {'url': 'https://arxiv.org/html/2609.39566v3#S6.T1', 'location': 'Table 1: accuracy versus medical and frontier models'}, {'url': 'https://arxiv.org/html/2609.39566v3#S6.SS3', 'location': 'Sec 6.3 / Table 2: post-training ablations'}, {'url': 'https://arxiv.org/html/2609.39566v3#A2.SS0.SSS0.Px4', 'location': 'App B: Consultant skill memory, frozen across runs'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026clinverseClinVersecomparatorsDiagnostic training and informative adaptation comparators2-implicitsimulatedExternal 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.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).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).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).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.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.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).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.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).not_identifiedSTOP 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).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.arXiv v1, 2026-10-26https://arxiv.org/html/2610.34878v1[{'url': 'https://arxiv.org/html/2610.34878v1#S3.T2', 'location': 'Table 2: L0-L3 results for Qwen3-32B with nine methods'}, {'url': 'https://arxiv.org/html/2610.34878v1#S4.SS3', 'location': 'Sec 4.3: evolution environment results and trajectory budgets'}, {'url': 'https://arxiv.org/html/2610.34878v1#A2', 'location': 'Appendix B: descriptions of the nine self-evolution methods'}, {'url': 'https://arxiv.org/html/2610.34878v1#S5.T4', 'location': 'Table 4: transfer to AIHospital'}, {'url': 'https://arxiv.org/html/2610.34878v1#A3', 'location': 'Appendix C / Table 7: SkillRL-GRPO weight updates on Qwen3-8B'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
cao2026guideskillGuideSkillskillsSkills, tools and workflow adaptation1offline_benchmarkGuideline-grounded executable ICD-10 scoring skills.Backbone weights, skill compiler and generator remain fixed.Compiled Python scoring programs and diagnostic rules.Labelled training cases and recurring diagnostic rationale patterns.Offline evolution before disjoint benchmark evaluation.Claude Sonnet 4.6 distils grouped case rationales.Syntax and tier-contract checks; no clinical release gate reported.Skill-component ablations; no changing-improver versus frozen-improver control.Four benchmarks; MedThinkBench excluded from evolution.not_identifiedTask skills change; update procedure stays fixed.No prospective care evaluation; external executable skill learning rather than parameter posttraining.arXiv v1, 2026-07-28https://arxiv.org/html/2607.26160v1[{'url': 'https://arxiv.org/html/2607.26160v1', 'location': '§3.3 skill evolution'}, {'url': 'https://arxiv.org/html/2607.26160v1', 'location': '§4 and Tables 1–3'}, {'url': 'https://arxiv.org/html/2607.26160v1', 'location': 'Appendix H'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
ruan2026normcomplianceRole-based norm agentskillsSkills, tools and workflow adaptation1simulatedNorm memory, system prompts and role identity.Qwen-plus weights and feedback/update design.Structured norm memory and revised role prompts.Norm trajectories and fixed AI preference feedback.Cross-episode prompt edits and periodic identity updates.Fixed diagnosis/reflection and identity-update routines.Constraint checks described; no clinical deployment gate.Component ablations; no isolated changing-improver control.Three synthetic healthcare communication tasks.not_identifiedTask role policies change; evaluator/update routine stays fixed.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.Official ACL Anthology conference PDFhttps://aclanthology.org/2026.findings-acl.1133.pdf[{'url': 'https://aclanthology.org/2026.findings-acl.1133.pdf#page=4', 'location': '§3.4–3.7; Algorithm 1'}, {'url': 'https://aclanthology.org/2026.findings-acl.1133.pdf#page=6', 'location': '§4; Table 1'}, {'url': 'https://aclanthology.org/2026.findings-acl.1133.pdf', 'location': 'Appendix A.1 and A.3 reproducibility'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
shi2026medguardMedGuardskillsSkills, tools and workflow adaptation1retrospectiveShared Qwen2.5-7B agent policy weights.Prompts, BM25 knowledge base, synthetic-data generator and GRPO routine.Updated weights reused across different consultations.JSON-format and reference risk-label rewards after full agent trajectories.Offline GRPO through Agent Lightning before separate consultation evaluation.Fixed DeepSeek-R1 generates risk/no-risk training samples.Rule-based terminal rewards; no prospective clinical release gate.Component ablations; no evolving-versus-fixed improver comparison.Separate real/synthetic conversations and retrospective expert assessment.not_identifiedPolicy changes; training-data generator and learning procedure stay fixed.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.Publisher Article in Press PDFhttps://www.nature.com/articles/s41746-026-03116-0_reference.pdf[{'url': 'https://www.nature.com/articles/s41746-026-03116-0_reference.pdf#page=21', 'location': 'Methods; MedGuard Agent'}, {'url': 'https://www.nature.com/articles/s41746-026-03116-0_reference.pdf#page=24', 'location': 'Domain-Transfer Reinforcement learning, pp.24–26'}, {'url': 'https://www.nature.com/articles/s41746-026-03116-0_reference.pdf#page=12', 'location': 'Doctor Evaluation and retrospective validation, pp.12–15'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
jiang2026sempaSEMPAskillsSkills, tools and workflow adaptation2-implicitretrospectiveEpisodic bank and Proposer/Critic natural-language policies.GLM-5 weights, retrieval, pool partition and rewrite routine.Past episodes and appended role-policy principles.Reference-drug overlap, DDI, molecular grounding and evolving Critic scores.Periodic policy rewrites across encounters; policies freeze before held-out evaluation.Fixed LLM meta-reflection contrasts success/failure pools.Utility-based pools and semantic-anchor checks.Evolution-module ablation; changing Critic is not isolated.Disjoint evolution stream and 100 held-out MIMIC-IV encounters.not_identifiedCritic affects later update selection; no tested revision of meta-reflection procedure.Small single-centre test; no prospective intervention. Drug-class F1 and exact-name Jaccard use different matching definitions.Official ACL Anthology conference PDFhttps://aclanthology.org/2026.findings-emnlp.994.pdf[{'url': 'https://aclanthology.org/2026.findings-emnlp.994.pdf#page=4', 'location': '§3.4–3.5; Eqs.5–7'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.994.pdf#page=5', 'location': '§4.1 evolve–freeze–test'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.994.pdf#page=9', 'location': 'Table 5'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.994.pdf#page=15', 'location': 'Appendix A.7 Algorithm 1'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026skillcdpoSkill-CDPOskillsSkills, tools and workflow adaptation1offline_benchmarkRuntime-derived tool-use skills and DR-Tulu-8B weights.Tools, expert model, critical-step selector and DPO procedure.Markdown skill cards, generated guards and trained weights.Invocation failures and expert/local rollout outcome differences.Offline skill construction, then single-round CDPO and held-out evaluation.LLM summarises historical failure clusters; fixed expert supplies alternatives.Rollout divergence selects preference steps; no clinical release gate.Skill and preference-learning ablations; no evolving-improver control.CURE-Bench, MedBrowseComp and PubMed Search with contamination filters.not_identifiedTask skills/weights change; update and selection procedure stays fixed.Only single-round CDPO tested; retained experience-based skills establish eligibility independently. No patient-care evaluation.Official ACL Anthology conference PDFhttps://aclanthology.org/2026.findings-emnlp.1985.pdf[{'url': 'https://aclanthology.org/2026.findings-emnlp.1985.pdf#page=3', 'location': '§3.1'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.1985.pdf#page=7', 'location': 'Table 1'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.1985.pdf#page=11', 'location': 'Appendix A.1'}, {'url': 'https://aclanthology.org/2026.findings-emnlp.1985.pdf#page=9', 'location': 'Limitations; single-round training'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
feng2026gatedopenclawGated adapter / OpenClaw agentskillsSkills, tools and workflow adaptation1retrospectivePersistent clinical guidance and final adapter gating parameter.RetFound/BioClinicalBERT towers, adapter MLPs and identity specification.Special-case MEMORY, distilled SOUL guidance and final gate.Original reports simulate physician adjudication of conflicting cases.Updates every 100 iterations across repeated test-set passes.Periodic fixed case summarisation and selective gate optimisation.Conflict cases trigger feedback; no independent clinical release gate.Adapter-capacity comparison; no matched evolving-improver control.Private fundus-image data; repeated adaptation on test cases.not_identifiedTask guidance and gate change; update procedure remains fixed.Agent LLM backbone unspecified. Evolution evaluated by conflict counts, with report-simulated feedback and reused test data, not independent diagnostic improvement or prospective care.Publisher PDF, 2026https://ijocsit.org/pdf/2026/10.62051-ijcsit.v8n4.01.pdf[{'url': 'https://ijocsit.org/pdf/2026/10.62051-ijcsit.v8n4.01.pdf#page=7', 'location': '§2.3.1–2.3.3'}, {'url': 'https://ijocsit.org/pdf/2026/10.62051-ijcsit.v8n4.01.pdf#page=9', 'location': '§3.1; Table 1'}, {'url': 'https://ijocsit.org/pdf/2026/10.62051-ijcsit.v8n4.01.pdf#page=10', 'location': '§3.2; Figure 4'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026emaiEMAImemoryMemory and experience across encounters1external_retrospectiveCluster-specific natural-language summarisation rules.Backbone during rule evolution, clustering and selection routine.Selected experience rules and historical reports.Automatic impression scores and difficult training reports.Offline iterative evolution before different test reports.LLM induces/refines rules using retained rules and difficult examples.Discard lower-scoring half; stop at predefined count/performance thresholds.Experience/component ablations; no isolated evolving rule-generator control.MIMIC-CXR and Open-I; separate training/test comparisons.not_identifiedExperience rules provide context for their own refinement; the induction/selection procedure stays fixed, as in the existing Level 1 notebook rule.Text findings-to-impression task; no image interpretation or prospective care. Contextual refinement of task rules does not establish a changed improver.Publisher HTML methods/results; first online 2025-09-29, issue January 2026https://www.sciencedirect.com/science/article/abs/pii/S016926072500495X[{'url': 'https://www.sciencedirect.com/science/article/abs/pii/S016926072500495X', 'location': '§2.2 Self-evolutive nearest-neighbor explicit experience induction'}, {'url': 'https://www.sciencedirect.com/science/article/abs/pii/S016926072500495X', 'location': '§2.4 Agentic framework'}, {'url': 'https://www.sciencedirect.com/science/article/abs/pii/S016926072500495X', 'location': '§4.2–4.3; Table 3; Figure 4'}, {'url': 'https://www.sciencedirect.com/science/article/abs/pii/S016926072500495X', 'location': '§4.5 manual evaluation'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026bridgegapBridgeGapskillsSkills, tools and workflow adaptation1simulatedQwen3-8B diagnostic dialogue policy weights.Patient simulator, profile verifier and reward constants.Trained policy weights; within-case belief states are temporary.Environment-verified belief progress and terminal diagnostic correctness.SFT followed by online simulated dialogue PPO across profiles.Frozen Qwen2.5-72B patient; fixed rule-based reward computation.Profile-grounded NLI verifies symptom evidence; no clinical release gate.Training/reward ablations; no changing-improver comparison.HealthBench, AgentClinic and static medical QA.not_identifiedTask policy changes; simulator, reward and learning procedure stay fixed.Belief-state persistence alone is within-case; eligibility rests on cross-episode policy learning. No prospective patient-care evaluation.Official ACL Anthology conference PDFhttps://aclanthology.org/2026.emnlp-main.1974.pdf[{'url': 'https://aclanthology.org/2026.emnlp-main.1974.pdf#page=3', 'location': '§3.1–3.2'}, {'url': 'https://aclanthology.org/2026.emnlp-main.1974.pdf#page=5', 'location': '§3.4.2 online belief-progress RL'}, {'url': 'https://aclanthology.org/2026.emnlp-main.1974.pdf', 'location': '§4.1 and Appendix C PPO details'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2025semagentsSeM-AgentsmemoryMemory and experience across encounters1offline_benchmarkCorrect-case and error-reflection experience databases.GPT-4-turbo weights, residual consultation protocol and reviewer prompts.CorrectKB cases and ChainKB structured error analyses.Reference correctness routes successful cases or error abstractions.Offline training consultations before test-set evaluation.Fixed Chain-of-Thought Reviewer abstracts incorrect consultations.Reference correctness gates storage; safety/ethics reviewer screens answers.Memory/discussion variants; no changing-improver control.MedQA/PubMedQA test sets; experience banks contain training-set cases only.not_identifiedMemory grows; abstraction and retrieval procedures stay fixed.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.IEEE Xplore full-text HTML, accessed through existing Tohoku University entitlementhttps://ieeexplore.ieee.org/document/10889517[{'url': 'https://ieeexplore.ieee.org/document/10889517', 'location': '§II.C Summary and output; Algorithm 2'}, {'url': 'https://ieeexplore.ieee.org/document/10889517', 'location': '§III.A Datasets'}, {'url': 'https://ieeexplore.ieee.org/document/10889517', 'location': '§III.B–D; Tables I–II; Figure 2'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
li2026decisionrulesEvolved multi-image decision rulesskillsSkills, tools and workflow adaptation1offline_benchmarkStrategy-specific prompt-program configuration for multi-image VQA.Qwen3-Coder-Next proposer, Qwen3.5-35B-A3B task VLM, ShinkaEvolve archive/selection and evaluation.Selected inference policy/configuration reused on different medical VQA cases.Clinical VQA accuracy on evolution cases, with larger-pool periodic reevaluation.50 offline evolution generations before held-out generation selection and final testing.Fixed Qwen3-Coder-Next patch proposals through ShinkaEvolve.Archive/island selection, periodic evolution-pool rescoring and a separate 665-case holdout.Common search engine and generation budget; no comparison that revises the improvement procedure.855 final-test examples, grouped by video ID apart from 1331 evolution and 665 selection cases; five independent main-method repeats.not_identifiedTask inference configurations change; proposal, archive and selection pipeline remain fixed.Internal partition of the public benchmark, not its official hidden test; search-mini feedback differs across strategies (256 versus128). No prospective clinical evaluation.arXiv v1, 2026-07-30https://arxiv.org/abs/2607.27564v1[{'url': 'https://arxiv.org/abs/2607.27564v1', 'location': '§3 and Table1, frozen split'}, {'url': 'https://arxiv.org/abs/2607.27564v1', 'location': '§4 and Figure4, mutable configuration'}, {'url': 'https://arxiv.org/abs/2607.27564v1', 'location': '§5 and Table3, models/budget/repeated results'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
dou2024shennongagentShennong-AgentmemoryMemory and experience across encounters1retrospectiveExperience vector database containing clinician corrections and prompts.Qwen-7B, five prompted clinical modules, retrieval/tool interfaces and update routine.Indexed clinician corrections, prompts and case-derived experience.Doctors add prompts or edit LLM inputs/outputs; all edits and prompts are re-encoded into the vector database.Feedback retained after human interventions for later similar clinical situations.Fixed recording, encoding and retrieval procedure.Clinician intervention; no independent update-release threshold reported.With/without HITL comparison; no isolated persistent-memory or changing-improver control.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.not_identifiedTask experience changes; update procedure and model are not shown to change.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.Published IEEE HTML, accessed through existing Tohoku University entitlementhttps://ieeexplore.ieee.org/document/10822302[{'url': 'https://ieeexplore.ieee.org/document/10822302', 'location': '§III opening paragraphs, correction/prompt recording and similar-situation reuse'}, {'url': 'https://ieeexplore.ieee.org/document/10822302', 'location': '§III.B vector database; §IV.C Qwen-7B'}, {'url': 'https://ieeexplore.ieee.org/document/10822302', 'location': '§IV.A–B retrospective cohort and raters'}, {'url': 'https://ieeexplore.ieee.org/document/10822302', 'location': '§V.B HITL comparison; §VI limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
rugo2026livingoepLiving-OEPskillsSkills, tools and workflow adaptation1offline_benchmarkAnnotation manual, prompt templates and few-shot examples.GPT-4.1/o3/Claude Sonnet-4 and human-supervised update procedure.Revised instructions and validated annotation examples.Human corrections of earlier abstract annotations.Repeated development iterations before locked, newly arriving validation publications.Human scientists revise instructions and agent components.Variable-specific accuracy targets; unsuccessful variables remain manually annotated.Commercial chatbot comparison; no isolated persistent-update or changing-improver control.1997 later breast-cancer publications excluded from training; eight clinical evidence scenarios.not_identifiedTask annotation guidance changes; no revised updater tested.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.Published 2026-09-01; full PMC articlehttps://pmc.ncbi.nlm.nih.gov/articles/PMC13563553/[{'url': 'https://pmc.ncbi.nlm.nih.gov/articles/PMC13563553/', 'location': 'Materials and methods: agentic AI construction and annotation process'}, {'url': 'https://pmc.ncbi.nlm.nih.gov/articles/PMC13563553/', 'location': 'Validation of the agentic AI system: independent locked publications'}, {'url': 'https://pmc.ncbi.nlm.nih.gov/articles/PMC13563553/', 'location': 'Results: iterative extraction refinement; chatbot-comparison limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
wu2026uehrU-EHRmemoryMemory and experience across encounters1external_retrospectivePositive/negative experience bank, memory quality estimates and retrieval statistics.Backbone weights, step-scoring/extraction prompts, three-score scheme and UCB update rules.Task summaries, effective action patterns, errors/reflections and historical retrieval/success counts.Relevance, progress and correctness scores for prior actions; completed-task success updates memory retrieval counts.After each clinical task, memories are extracted and statistics updated before the next task.Fixed LLM scoring, extraction and memory consolidation procedures.Positive/negative step thresholds 0.7/0.3; UCB retrieval uses fixed quality weights and exploration coefficient.Ablations remove UCB, step scoring and negative memory; no changing-improver control.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.not_identifiedTask memory and retrieval estimates change under fixed scoring/extraction/UCB rules.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.Published 2026-08-08; online publisher PDF read through existing browserhttps://dl.acm.org/doi/pdf/10.1145/3770855.3819038[{'url': 'https://dl.acm.org/doi/pdf/10.1145/3770855.3819038#page=4', 'location': '§3 and Figure2, retained bank and next-task reuse'}, {'url': 'https://dl.acm.org/doi/pdf/10.1145/3770855.3819038#page=5', 'location': '§4.1–4.3, scoring, extraction and UCB'}, {'url': 'https://dl.acm.org/doi/pdf/10.1145/3770855.3819038#page=6', 'location': '§4.4, Tables1–3'}, {'url': 'https://dl.acm.org/doi/pdf/10.1145/3770855.3819038#page=7', 'location': '§5.2–5.4 and Table4'}, {'url': 'https://dl.acm.org/doi/pdf/10.1145/3770855.3819038#page=11', 'location': 'AppendixA, sequential task order and fixed prompts'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
neary2025fastCoach Iris / FASTskillsSkills, tools and workflow adaptation1retrospectiveShared agent prompts, response parameters and retrieval data; model refinement is described.Backbone and optimizer Not reported.Adjusted prompts/RAG/settings; fine-tuned model described.Human FAST ratings and SME reviews.Continuous monitoring; revision cadence Not reported.Human reviewers, SMEs and ML team.Unacceptable ratings trigger refinement; release gate Not reported.Not reported.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.not_identifiedHuman rating rubric was revised; no effect of those revisions on subsequent agent learning was tested.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.published18Jun2025https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/full[{'url': 'https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/pdf#page=4', 'location': '§2.4 SME-directed refinements; PDFp4'}, {'location': '§2.5 updates from monitoring', 'page': 4, 'url': 'https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/pdf'}, {'location': '§3.1.1 unacceptable feedback to ML fine-tuning', 'page': 5, 'url': 'https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/pdf'}, {'location': '§3.4 rubric changes', 'page': 8, 'url': 'https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/pdf'}, {'location': '§4 face validity and absent measure validation', 'page': '8–9', 'url': 'https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1460236/pdf'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
fan2026medenvMedEnv / Diagnosis-R1generatorsParameter updates and changing training generators2-implicitsimulatedDoctor-policy parameters.Simulator/tool interfaces, disease KB and GRPO/reward formula.Trained Diagnosis-R1 checkpoint.Terminal diagnosis correctness plus confidence gain toward correct diagnosis; unsupported-action penalty.SFT then stochastic interactive RL rollouts and parameter updates.GRPO under rllm; SFT policy filters cases using8 rollouts, discarding uniformly correct/incorrect cases.Case filtering described; checkpoint acceptance/release criterion absent.Table3 matchedSFT, outcome-onlyRL and full-reward ablations; same initialization/data/steps/hyperparameters stated.2000 heldout cases separate from RL, per author response; patient-level disjointness and external validation unspecified.not_identifiedThe 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.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.Public accepted10-page paper; society HTML abstract still calls environment DiagWorld and500K+, final PDF/ Springer useMedEnv496K+.https://papers.miccai.org/miccai-2026/paper/5084_paper.pdf[{'url': 'https://papers.miccai.org/miccai-2026/paper/5084_paper.pdf#page=4', 'location': '§2.1, clinical simulators/tools'}, {'url': 'https://papers.miccai.org/miccai-2026/paper/5084_paper.pdf#page=5', 'location': '§2.2, interactive GRPO and policy-likelihood process reward'}, {'url': 'https://papers.miccai.org/miccai-2026/paper/5084_paper.pdf#page=7', 'location': '§3.1/3.3, training and Table3'}, {'url': 'https://papers.miccai.org/miccai-2026/0631-Paper5084.html', 'location': 'Author Feedback, held-out-case separation'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
mao2026srmaprSR-MAPRmemoryMemory and experience across encounters1offline_benchmarkHeuristic rule memory for disease-specific physician assessment.Specialist/Summary/Reflection prompts, backbone options, clustering, retrieval and expert reference labels.Rule-based memory alongside labelled case memory.Summary predictions that disagree with expert gold ratings trigger error analysis.Error-derived rules are added for future physician-profile evaluations; update/test scheduling not detailed.Fixed Reflection Agent produces generalized heuristics from rating errors.New rules stored in memory; acceptance/release criteria not specified.Rules/case/RAG removal ablations; no fixed-versus-changing improver comparison.7362 physician profiles with stated80/20 split, four Summary backbones, ranking/classification metrics and component ablations.not_identifiedTask assessment heuristics change; criteria-generation and reflection procedures stay fixed.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.Published IEEE HTML; conference 2026-01-30–02-01; added2026-03-31https://ieeexplore.ieee.org/document/11452303[{'url': 'https://ieeexplore.ieee.org/document/11452303', 'location': '§III.A dataset and split'}, {'url': 'https://ieeexplore.ieee.org/document/11452303', 'location': '§III.C.3–4 dual memory and self-reflective rule storage'}, {'url': 'https://ieeexplore.ieee.org/document/11452303', 'location': '§IV.A fixed backbone/retrieval implementation'}, {'url': 'https://ieeexplore.ieee.org/document/11452303', 'location': '§V.C and TableIII ablations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2027mdtaceMDTacememoryMemory and experience across encounters1offline_benchmarkGlobal Bullet Playbook of structured consultation experience.GPT-4o/backbone variants, specialist selection, lead/reflector prompts and update/retrieval procedure.Up to600 recent training-case bullets; cosine-similar entries can be replaced.Completed training-case deliberations, disagreements and tool-use feedback are distilled and filtered by the Reflector.Bullets accumulated from training cases, then retrieved on separate test cases.Fixed Lead Physician aggregation and Reflector filtering.High-value reusable bullets retained; capacity/similarity replacement rule and top5 retrieval fixed.Primary-doctor, context-layer and Playbook removal ablations; no changing-improver control.MedQA, PubMedQA, VQA-RAD and SLAKE test sets; all stored bullets explicitly training-derived.not_identifiedTask memory changes and conditions subsequent consultation summaries; memory-construction procedure remains fixed.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.Publisher full HTML; first online2026-07-19; citation/imprint2027https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6[{'url': 'https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6#Sec7', 'location': '§3.2, Reflector-to-Playbook memory write'}, {'url': 'https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6#Sec8', 'location': '§3.3, cross-case bullet aggregation/filtering'}, {'url': 'https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6#Sec10', 'location': '§4.1, training-only memory/test evaluation'}, {'url': 'https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6#Sec11', 'location': '§4.2, capacity/retrieval, results and ablations'}, {'url': 'https://link.springer.com/chapter/10.1007/978-981-92-3403-5_6#Sec3', 'location': '§2.1, distinct SeM-Agents/MDTeamGPT predecessors'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496
chen2026rhisRHISskillsSkills, tools and workflow adaptation1simulatedVersioned external action-governance controller rules, payload adapters and terminal-obligation checks.Kimi kimi-k3 with high reasoning effort; frozen MedAgentBench source and isolated official scorer; declared change-admission record.Controller, self-evolver, configuration templates and versioned rules; v0.3.25 package reports128 passing unit/regression tests.Visible execution failures, internal runtime-calibration signals or authorized external instructions trigger proposed controller revisions.Successor controller versions are released between frozen runs and reused across later benchmark stages; same-stage recoveries and full retests are reported separately.Self-evolver generates versioned rule changes under a fixed protocol; detailed implementation of its proposal procedure is not reported.Change record requires frozen trigger evidence, root-cause account, proposed rule, adjudication, successor version, rollback target, regression result and independent-evaluation boundary.Named predecessor/successor controller runs, without a matched frozen-improver control; aggregate across versions is explicitly descriptive.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.not_identifiedThe external controller changes task action admission. No revision and later-learning test of the change-proposal/adjudication procedure itself is demonstrated.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.SSRN Working Paper Version0.1, September2026, posted2026-09-21;13-page online PDFhttps://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519[{'url': 'https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519', 'location': 'PDFp3, §3.1 benchmark source and300-task stages'}, {'url': 'https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519', 'location': 'PDFp4, §3.2–3.4 controller and controlled self-evolution'}, {'url': 'https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519', 'location': 'PDFpp5–6, §5 first-pass versus named post-evolution runs'}, {'url': 'https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519', 'location': 'PDFp7, §6 costs and state-closure boundary'}, {'url': 'https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7482519', 'location': 'PDFpp10–11, §10 limitations'}]Author-approved manuscript evidence-profile table, source commit f974dfbb5da9d4b51725f62e788a99e409721496