Recursive Self-Improvement in Healthcare AI Agents Study-level extraction for 85 studies. The unified four-source search of 4 November 2026 re-confirmed31 previously charted studies and added52 (34 from the initial three sources,18 from OpenAlex); two comparators were retained purposively.
Publication status, mechanism evidence and clinical evaluation are separate fields. “Not reported” refers to the examined source version. Clinical-setting classification follows the reported study evaluation design. Coach Iris is retrospective; its operating patient service is documented separately. Source versions and section/table locators are given below. Author spotchecks remain pending.
1. Agent Hospital / MedAgent-Zero
Key
medhospital2024
Name
Agent Hospital / MedAgent-Zero
Title
Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents
Diagnostic skills, encounter/EMR memory, and outcome-dependent memory-importance scores.
Fixed components
Frozen GPT-3.5; constant dialogue, diagnosis and skill-reflection prompts; fixed supervisor stages and relevance/importance sampling formula.
Feedback
Professional D4 risk labels, revealed after diagnosis, guide supervisor reflection; diagnostic correctness increments or decrements retrieved-memory importance.
Timing
After simulated training consultations (Quiz); accumulated memory is reused on subsequent patients and in the single-attempt Exam setting.
Retained artifact
Conversation records, summarized EMRs, reusable diagnostic skills, and memory-importance scores.
Update generator
LLM supervisor compares predicted and reference depression/suicide risk and produces corrective skills; a fixed rule updates importance.
Selection and release
Supervisor-generated skills are written to memory; later retrieval samples ten nodes using fixed relevance/importance weighting. The described workflow does not impose an independent clinical release gate.
Evaluation separate from updating
Separate D4 test cases are evaluated in a single-attempt Exam, using professional reference labels; original versus simulated dialogues and memory/supervisor ablations are reported.
Fixed-procedure control
Memory-off, memory-layer and supervisor ablations test component value; they do not compare an evolved skill-generation procedure with its fixed predecessor.
Evidence concerning improver change
not established: skills and retrieval priorities change under constant reflection prompts and fixed importance-update/sampling rules.
Clinical setting
Chinese depression/suicide-risk classification using D4 dialogue-derived patient simulations and original dialogues; 100 selected training cases and 132 original test cases are described.
Evidence boundary
Retrospective dialogue/simulation evidence, with limited four-class accuracy and poorer simulated dialogue performance. The study explicitly lacks real-time expert feedback and prospective patient assessment.
Medical-record and reflection-experience buffers used for later diagnostic retrieval.
Fixed components
Doctor/patient/measurement roles; KNN retrieval, chain-of-thought prompting, majority-vote ensemble, and correct-diagnosis storage rule; no model-weight update is part of the described method.
Feedback
Reference diagnosis determines whether the initial or reflected diagnosis is correct; benchmark outputs are matched to labels by an LLM evaluator with manual review of accuracy logs.
Timing
After each simulated consultation; stored records are retrieved during later consultations before ensemble diagnosis.
Retained artifact
CLIP-embedded successful case records and reflection insights from initially incorrect cases corrected on a second attempt.
Update generator
The doctor reflects on an incorrect diagnosis and retries; the record-store module embeds and retains eligible results.
Selection and release
Correct initial diagnoses enter the medical buffer; only successful corrected diagnoses contribute reflection insights to the experience buffer; unsuccessful retries are discarded.
Evaluation separate from updating
NEJM, MedQA and MIMIC-IV cases are converted to interactive simulations; LLM diagnosis scoring and data conversions are manually reviewed. A separate protected memory-development/test split is not specified in the full published Methods/Experiments.
Fixed-procedure control
Incremental measurement, memory, chain-of-thought and ensembling ablations compare task pipelines, not evolved versus unchanged experience-generation procedures.
Evidence concerning improver change
not established: experience buffers grow under an unchanged retrieval, reflection, voting and admission pipeline.
Clinical setting
Interactive doctor/patient/measurement simulations derived from public clinical cases and exam data; limited human-role identification assesses perceived realism, not clinical efficacy.
Evidence boundary
Benchmark/simulation results support memory-assisted diagnosis. The human-role study is not prospective care, and cumulative additions of memory/reasoning/ensembling cannot isolate recursive improver gains.
LLM extraction of successful strategies and failure divergences; fixed credit-assignment equations update graph weights from feedback.
Selection and release
Experience reliability validation initializes quality; fixed thresholds construct edges; online feedback reweights activated memories for subsequent traversal. No clinician release gate is part of the reported benchmark pipeline.
Evaluation separate from updating
Held-out MedR-Bench diagnosis/treatment cases and five MedAgentsBench subsets; DeepSeek-V3.2 judges reasoning and is also one generator. The relation of online-evolution samples to snapshot test cases is not fully specified in §5.3/Appendices.
Fixed-procedure control
Static GSEM is compared with evolving-memory snapshots, plus retrieval ablations; this isolates graph-state calibration, not a self-revised calibration procedure.
Evidence concerning improver change
not established: feedback changes retrieval-affecting state under fixed update equations; the extraction, traversal and calibration procedures are not shown to revise themselves.
Clinical setting
Public retrospective clinical-reasoning and medical-QA benchmarks; automatic evaluation rather than live clinical interaction.
Evidence boundary
Supports outcome-informed memory selection and graph calibration. It does not establish prospective benefit, clinician-validated updates, or improvement of the memory-update algorithm itself.
Episodic cases, diagnostic procedural rules/utility statistics and tool-trust state.
Fixed components
Frozen VLM and specialist tools; fixed base protocol, memory factorization, retrieval/utility selection and tool-trust thresholds.
Feedback
The benchmark ground-truth answer is revealed after each episode; multimodal reflection uses images, predictions, traces and that answer.
Timing
Test-time, between sequential chest-X-ray cases; later cases inherit earlier-case memory.
Retained artifact
Compressed episodes, priority-tagged procedural rules and usage statistics, plus per-tool helpful/harmful/misuse counts and trust labels.
Update generator
One post-case multimodal reflection call proposes procedural patches and governance updates; fixed accounting rules maintain utility and trust.
Selection and release
Relevant episodes and utility-ranked rules enter later prompts; tool labels use preset count/rate thresholds. The published experiment does not independently validate each proposed memory patch before reuse.
Evaluation separate from updating
ChestAgentBench answers score sequential predictions; three case-order permutations are reported for GPT-5-mini. This is online benchmark evaluation with prior-answer feedback, not a separately frozen-memory external clinical cohort.
Fixed-procedure control
No-memory and episodic/procedural ablations; fixed versus evolved reflection or memory-update procedures are not compared. Tool-governance quantitative ablation is explicitly deferred.
Evidence concerning improver change
not established: diagnostic procedures and tool-use preferences change; the mechanism generating and admitting future memory changes remains specified by the fixed reflection pipeline.
Clinical setting
Sequential retrospective chest-X-ray questions on ChestAgentBench, including MCQ and open-ended tasks; no model retraining.
Evidence boundary
Initial single-modality benchmark evidence with immediate ground truth. Tool-governance is described but not quantitatively established; delayed/noisy clinical feedback and prospective care remain untested.
A growing repository of diagnostic cognition primitives (DCPs) that guides later test ordering and diagnostic decisions.
Fixed components
Frozen base LLM; shared action/prompt contract, DCP-construction template, embedding retriever, evaluator and bounded diagnostic action space.
Feedback
Accrual-case reference diagnoses, correctness flags, rule-based process feedback and charted clinician test orders guide DCP construction.
Timing
Post-encounter consolidation in a dedicated accrual pool; nested exposure-indexed repositories are tested on fixed, non-overlapping evaluation cohorts.
Retained artifact
DCPs with an experience pattern, test-ordering guidance, diagnostic-decision guidance and provenance including exposure index, pathology and source outcome.
Update generator
The same backbone uses a fixed reflection template to distill completed encounters into reusable DCPs.
Selection and release
DCPs are stored and retrieved by dense pattern similarity; the agent is instructed to use them only when compatible with acquired evidence. Clinician ratings assess sampled DCPs after construction, rather than serving as per-update release approval.
Evaluation separate from updating
Non-overlapping 400-case MIMIC-CDM evaluation; a separate 80-case reader-study split; a 293-case external hospital cohort excluded from accrual; two blinded physicians rate 40 sampled DCPs.
Fixed-procedure control
Backbone-matched DCP-free and retrieval ablations plus exposure curves isolate experiential reuse; no changed-versus-fixed DCP-construction procedure is evaluated.
Evidence concerning improver change
not established: later DCPs receive stronger clinician ratings and improve reuse, but maturation of stored guidance is not evidence that the fixed consolidation/retrieval procedure modifies itself.
Clinical setting
Retrospective EHR-based interactive workups for acute abdominal conditions, plus cross-institution/cross-language external validation; imaging is represented by narrative reports.
Evidence boundary
Stronger external retrospective evidence than simulation-only studies; no prospective patient-outcome trial. The published clinician reference uses full information, so it is not a matched-information superiority comparison.
Actor prompt rules and action-conditioned external memories.
Fixed components
Base-model weights; Process Grader and Evolver instructions; preset memory-deletion coefficients and budgets.
Feedback
Diagnosis Judge grade, encounter costs, and action-level yield/efficiency critiques.
Timing
Between diagnostic episodes; memories retrieved during subsequent encounters.
Retained artifact
Bounded operational rules and context-action-outcome memories, including error examples.
Update generator
Fixed-instruction LLM Evolver edits Actor state from graded trajectories.
Selection and release
Budgeted add/delete/merge operations; generalize identifying details; no independent candidate-release test specified.
Evaluation separate from updating
Separate diagnosis Judge, but its score feeds adaptation; the reported fixed-order case stream is not an independent frozen holdout.
Fixed-procedure control
Static, memory-only, prompt-only and transcript-feedback controls; these isolate components, not evolving versus fixed Evolver.
Evidence concerning improver change
Not established: Actor strategy changes; the described Evolver/Grader procedure is not revised.
Clinical setting
Med-Inquire simulation over 915 published clinical cases; generated missing details are not clinical ground truth.
Evidence boundary
Supports sequential prompt/memory adaptation and resource-aware diagnosis in simulation, not patient-care safety or improved update-generation ability.
Highest-reward session continues the trajectory; selected trajectories train the model. No independent skill-release test described.
Evaluation separate from updating
Separate evaluation profiles plus human ratings of 522 dialogues by two psychology-trained graduate raters; simulated clients.
Fixed-procedure control
Memory/planning, skill-evolution and internalization ablations; no dedicated control isolating shared-model improvements as an updater.
Evidence concerning improver change
Reported shared-model learning can affect skill extraction/management; improved update-generation ability is not independently established.
Clinical setting
PsychEval simulated multi-session counseling; 120 training and 20 evaluation profiles per therapeutic school.
Evidence boundary
Response-quality and memory/skill benefits; not measured therapeutic outcomes. Public release lacks the complete post-session skill-evolution pipeline.
Patient demonstration/requirement library and doctor questioning-trajectory library.
Fixed components
Model weights, initial standardized-patient requirements, quality criteria and designed recruitment topology.
Feedback
Validated question/answer quality and subsequent usefulness of retrieved library entries.
Timing
Dialogue-level storage and reuse during simulated encounters; persist across cases.
Retained artifact
Attention library of question-record-answer-requirement tuples and doctor transition trajectories.
Update generator
Evolving doctor questioning generates patient-learning interactions; attention agent extracts requirements.
Selection and release
Quality validation before insertion; keep stronger duplicates; remove entries after two harmful uses.
Evaluation separate from updating
Separate disease-transfer assessment and human/GPT-4 preferences; no prospective human-trainee clinical-outcome test.
Fixed-procedure control
Tables 4–5 test doctor evolution and recruitment; patient-impact evolved-doctor condition bundles trajectory, question-pool and profile components.
Evidence concerning improver change
Demonstrated at interaction-generator level: evolved doctors alter patient-training experience. Recursive improvement of the whole coevolution procedure is not established.
Clinical setting
Simulated standardized patients grounded in hospital records and public data; medical-education task.
Evidence boundary
Supports co-adaptive training interactions. Bundled ablations and fixed evaluation/coevolution rules preclude an unqualified strict-RSI claim.
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.
Downstream medical-image classifier weights, through full/partial/head-only fine-tuning or gradual unfreezing; the language-model agent itself is not the trained target.
Fixed components
GPT-4.1 reasoning engine; permanent system prompts, agent roles and predefined Python tools; fixed global decline threshold and parameter-anchoring objective. User prompts may specify the fine-tuning recipe.
Feedback
Ground-truth labels for monitoring images/scans produce accuracy, precision, recall and F1 metrics; relative decline against development-test metrics triggers intervention.
Timing
User-prompted monitoring cycles compare a separate inference cohort with development-test performance, then conditionally fine-tune on a distinct cohort and re-evaluate the same inference cohort.
A master LLM agent routes comparison and fine-tuning tasks to specialist agents, which invoke predefined training tools using user constraints or configured defaults.
Selection and release
A global 5% relative-decline heuristic flags any deteriorating metric; the fine-tuning tool selects checkpoints using validation performance and the workflow rechecks inference metrics. Clinical approval is recommended in discussion, not evaluated as a release gate.
Evaluation separate from updating
Patient-separated development, monitoring and adaptation cohorts are reported. Monitoring labels select adaptation and the same monitoring cohort measures its outcome, so post-update results are not from an untouched final holdout.
Fixed-procedure control
Before/after classifier performance is compared; the paper explicitly lacks scripted or single-agent baselines and direct catastrophic-forgetting assessment. It does not compare changed versus fixed improver procedures.
Evidence concerning improver change
not established: classifier parameters adapt while the LLM, permanent prompts, comparison rule and fine-tuning toolkit remain fixed; selecting a configured training strategy is not evidence of revising the improver.
Clinical setting
Retrospective public MRI, CT and chest-radiography classification datasets; one MIMIC-CXR temporal experiment, with the remaining datasets randomly partitioned from the same distribution.
Evidence boundary
A labeled monitoring-and-retraining prototype, not label-free shift detection or prospective deployment. Reused monitoring data, a single temporal dataset and inconsistent temporal/general split descriptions limit stronger generalization claims.
PrivMeSA: Privacy-Aware Self-Evolving Multi-Agent System for Medicine via Local-Remote LLM Collaboration
Authors
Dannong Wang; Yuran Zhang; Bian Sun; Alex Stinard; Yuzhang Shang; Song Wang; Yu Tian
Year
2026
Publication status
P: arXiv preprint (v1, 29 Sep 2026) with no arXiv comment or journal_ref; listed on OpenReview as an ICLR 2027 Conference Submission (forum Ch62pc2CZs), no acceptance found; Crossref title search returned no record (checked 2026-11-03).
Two components change: a local lesson memory of distilled specialist answers, and the local policy's LoRA parameters trained by GRPO to decide when to consult, what to disclose, and whether to accept a retrieved lesson.
Fixed components
Remote GPT-5.6-Luna specialists, the frozen Gemma 4 12B distiller, judge and training attacker, the bge-small-en-v1.5 retriever with fixed thresholds (0.775 retrieval, 0.90 replacement), the GRPO recipe and reward weights (lambda_d=0.5, lambda_a=0.4) stay fixed.
Feedback
RL reward combines task accuracy against recorded MIMIC-IV-ED labels (diagnosis graded by CCS category) with discretion over injected synthetic personal details and registry-based log k-anonymity from an LLM attacker; lesson distillation uses no labels or reward.
Timing
Policy updates occur offline over 200 (Gemma) or 120 (Granite) GRPO steps; memory is updated after each completed episode (one sampled rollout per GRPO group during training, every case during test-time learning), taking effect only for later episodes.
Retained artifact
A LoRA adapter for the local policy and a lesson store of task, generalized question key, specialist-derived rule and provenance (2,026 lessons for Gemma and 1,470 for Granite at end of training).
Update generator
GRPO with group-mean-centred advantages updates the policy; a frozen Gemma 4 12B distiller converts each completed consultation exchange into a lesson or discards it, and near-duplicate same-task lessons are replaced.
Selection and release
Distilled lessons enter memory without outcome checking; at retrieval an LLM judge selects at most one of ten nearest lessons and verifies applicability in a second call, and the trained policy decides whether to use it; no clinical review or release gate is described.
Evaluation separate from updating
Patient-disjoint held-out MIMIC-IV-ED patients (100 for main results, 200 patients/897 encounters for test-time learning) with an evaluation attacker (GPT-5.6-Luna) different from the training attacker; no seeds, confidence intervals or significance tests are reported.
Fixed-procedure control
Ablations (untrained policy, no-probing training, local-only with and without RL, reward-term variants) and a paired memory-on versus memory-off test-time run; no comparison of a revised versus fixed distiller, judge, reward or training procedure.
Evidence concerning improver change
Not established: memory and policy parameters change, but the distiller, judge, retriever, GRPO recipe and reward are fixed throughout; the trained policy only indirectly shapes which exchanges are distilled, and this effect on later learning is not tested.
Clinical setting
Retrospective emergency-department replay of MIMIC-IV-ED encounters (disposition, primary ICD diagnosis by CCS category, coded procedure for admitted patients), with synthetic personal identifiers injected and re-identification measured against the de-identified 205,504-patient MIMIC-IV-ED registry.
Evidence boundary
Retrospective replay on de-identified records with synthetic personal details and a registry-conditional k measure; no clinician-facing, prospective or patient-outcome evaluation, and the test-time memory gain is mainly fewer remote messages at near-equal accuracy.
Learning from Expert-Corrected Misdiagnoses: A Self-Evolving Consultation Agent for Traditional Chinese Medicine after Failed Treatment
Authors
Ye Wang; Zhikang Lin; Jia Liu; Tianle Zhang; Nanyue Wang
Year
2026
Publication status
P: arXiv preprint (q-bio.OT), v1 only, submitted 2026-10-09; no comments or journal_ref on arXiv, and no Crossref, OpenReview or OpenAlex venue record found (OpenAlex lists only the arXiv preprint; checked 2026-11-03).
For each failed held-in consultation, per-metric outcomes plus the reference syndrome, treatment principle, error rubric and the monograph's original expert analysis with full prescription; the ablation sees only per-metric right/wrong.
Timing
Offline batch rounds (9 rounds, one proposal per round, seed 42) on fixed development sets before sealed test evaluation; no updating during test or deployment.
Retained artifact
The accepted skill-document text; two edits were accepted (rounds 1 and 6), both short pre-submission checklists.
Update generator
The same LLM that consults diagnoses failed held-in cases into signatures, clusters and ranks them into a brief with successful cases as regression checks, and proposes one minimal edit or declines.
Selection and release
An edit is accepted only if the summed composite score does not drop on either the 80-episode held-in set (drawn from training) or the 153-episode validation (held-out) set and rises on at least one; no clinician review before release is reported.
Evaluation separate from updating
Sealed 141-episode test split from the same three monographs, not used in selection, each final skill document run three times and averaged, scored by a fixed DeepSeek-Flash judge (temperature 0) plus exact-match and set-based metrics.
Fixed-procedure control
Base harness (empty document) and a learning-signal ablation (outcome-only feedback) under the same loop; no comparison of an evolved versus fixed edit-generating or gating procedure.
Evidence concerning improver change
Not established: the diagnose-propose-gate procedure, prompts and acceptance rule stay fixed across rounds, and only the skill document changes.
Clinical setting
TCM consultation after failed treatment (syndrome differentiation, recognition of the prior physician's error, corrective treatment principle), using 820 episodes digitized from three published misdiagnosis monographs with a deterministic simulator rather than live patients.
Evidence boundary
Retrospective, simulated benchmark from published case books with one backbone LLM (also used as judge), a single evolution seed and no reported confidence intervals, significance tests or physician assessment of the learned guidance.
C: published in Findings of the Association for Computational Linguistics: EMNLP 2026, pp. 44611-44629 (ACL Anthology 2026.findings-emnlp.2094, DOI 10.18653/2026.findings-emnlp.2094); arXiv v1 comment also says EMNLP 2026 Findings.
The natural-language operating manual (main-agent prompt with triage, treatment-completeness and disposition rules and a presentation-to-management table).
Fixed components
Claude Opus 4.6 backbone, MCP clinical tool interface, hand-curated skills library, subagent prompts extracted from the initial manual, proposer prompt, GPT-4.1 judge and CES simulator.
Feedback
Trace-level feedback from completed rollouts: aggregate, per-batch and per-patient LLM-judge scores on four dimensions (graded against gold answer keys), per-turn action records and LLM call audit logs.
Timing
Offline, between iterations; each of 10 proposer steps is followed by a full rollout over the same six search batches, and the selected manual is frozen at inference.
Retained artifact
The selected evolved manual M* (v7), plus a filesystem of all prior candidate manuals with their scores and traces that the proposer reads.
Update generator
A separate Claude Code proposer (Opus 4.6) with a fixed system prompt requiring edits grounded in failure patterns seen in at least 3 patients across batches, no patient- or batch-specific knowledge, and substantive edits to the current best manual.
Selection and release
After 10 iterations the manual with the best overall search-batch score (v7) is selected and reused unchanged on held-out batches.
Evaluation separate from updating
Four held-out batches (48 patients) from the same distribution, never shown to the proposer, with one rollout per configuration; baseline and full system re-graded by four further LLM judges.
Fixed-procedure control
Add-one component ablations and an evolved-versus-initial manual contrast holding skills and subagents fixed; no evolving-versus-fixed proposer test.
Evidence concerning improver change
Not reported; the proposer model, prompt and best-score selection rule are fixed across all iterations.
Clinical setting
Simulated six-hour emergency-department shift in CES (12 patients per batch, 4 beds, 6 nurses, one agent physician) with patient cards from de-identified ED records and GPT-4.1-generated physiology.
Evidence boundary
Simulator-only, LLM-graded, single rollout per batch and single backbone, no expert agreement on Asclepius traces; the evolved manual's marginal held-out gain over skills plus subagents is +0.02 (p=0.883).
ReVision3D: Attribution-Guided Recursive Self-Improvement for 3D Medical Perception
Authors
Ho Hin Lee; Yuyin Zhou; Yannan Yu; Shi Gu; Yifan Wu
Year
2026
Publication status
P: arXiv preprint (v1 26 Sep 2026, v2 2 Oct 2026; comment '28 pages', no journal_ref or DOI); OpenReview lists it only as an ICLR 2027 Conference Submission and Crossref has no record (checked 2026-11-03).
Reader perception configuration (vision-encoder LoRA blocks, rank, input scale, epochs), the RL training recipe that produces reader weights, the 3D candidate-scoring inference program, and a view-selection program.
Fixed components
Frozen LLM designer, frozen pretrained Qwen3-VL-4B weights (only adapters train), mask-grounded verifier with 5 mm tolerance, system objective J, acceptance rule and four predefined intervention levels.
Feedback
Reference-mask verification of localized findings, a mask-grounded training reward, J over false-positive budgets on development patients, and replay-based loss attribution; the designer sees only aggregate trial statistics, attribution summaries and trial histories.
Timing
Offline revision rounds on development patients, each accepted change triggering renewed attribution; held-out test patients are read once after all choices are frozen.
Retained artifact
The accepted incumbent system (adapter weights, recipe, inference program) plus an evidence-backed playbook of successful and failed hypotheses and trial registries.
Update generator
A frozen language-model designer (Claude Opus 5 by default) proposes revisions within the level indicated by attribution, screened by replay where possible; perception changes require actual training.
Selection and release
Training revisions are accepted only if mean dev gain exceeds the incumbent's seed range and every matched seed improves; inference revisions need P(dJ>0) >= 0.9 under paired bootstrap.
Evaluation separate from updating
One-time held-out test on 31 liver and 70 kidney patients with patient-level bootstrap, though Appendix B reports that the 31 liver test patients were also used in earlier liver-study comparisons and only 40 of 70 kidney test patients were unused.
Fixed-procedure control
Equal-compute longer-training control for recipe revisions, frozen-encoder and hand-chosen perception comparisons, and a four-designer swap at the inference level; no evolving-versus-fixed improver test.
Evidence concerning improver change
The training recipe that generates reader weights is revised (one accepted learnability recipe, later recipe proposals rejected), while the designer, attribution, verifier, objective and acceptance rule stay fixed and the playbook's effect on later rounds is not tested.
Clinical setting
Retrospective lesion localization on public de-identified abdominal CT (LiTS liver, KiTS23 kidney); authors state it is not a clinical device and makes no diagnostic claim.
Evidence boundary
Benchmark-only, mask-verifiable lesion localization; routing rule formulated retrospectively and did not schedule the experiments; acquisition never passed the promotion gate; no clinician or patient-outcome evaluation.
P: arXiv preprint (cs.MA, cs.CL) submitted 2026-10-04, single version v1; no arXiv comments or journal_ref, and Crossref/OpenAlex searches (2026-11-03) returned only the arXiv record.
A persistent hierarchical reporting template (slot keys, value types, candidate values, slot-level evidence notes, slot-specific extraction patterns) plus anatomical and diagnostic knowledge graphs built from the training corpus.
Fixed components
Backbone LLMs (Qwen3.6-Flash for induction/evolution, Gemini-3.5-Flash for structuring, reconstruction and diagnosis), text-embedding-3-small, prompts, clustering hyperparameters, the deterministic reducer and the evaluation judges are fixed; no parameters are trained.
Feedback
Training-corpus statistics, structural/unit consistency checks, retrieved PubMed/Europe PMC/Google evidence, and one audit of the Clinical Agent's structuring of sampled training findings for collapse to generic values.
Timing
Offline, before evaluation: a single construction pass per training split (one round of feedback-guided granularity calibration); the frozen template and graphs are then applied to held-out reports.
Retained artifact
Final template T* with reference notes and extraction patterns, the anatomical graph, the frozen diagnostic graph, and an indexed training-sentence corpus, all with source provenance and audit logs of retrieval queries and rationales.
Update generator
LLM Evolution Agent under fixed prompts proposes assignments, merges, fine revisions, evidence notes and slot-specific patterns P_s.
Selection and release
A deterministic reducer rejects revisions failing structural or unit checks, slots need more than 3 supporting atoms, vocabularies are supplemented only with training-sentence support, and calibration adds guidance without changing allowed values; no human or clinical release gate is described.
Evaluation separate from updating
Held-out test splits of four datasets scored for LLM-matched template coverage, reference-based reconstruction metrics, LLM-judged diagnostic fidelity (PDA, KFP, CA), and expert plus LLM template-quality ratings; no statistical tests or intervals are reported.
Fixed-procedure control
Cumulative component ablation from ASTAR (M1) to full system (M6), LLM-judged pairwise Revision/Note ablation, and cross-dataset transfer without re-induction; the feedback calibration step is not ablated separately and there is no evolving-versus-fixed updater test.
Evidence concerning improver change
not established: the template and graphs change once under constant prompts, thresholds and reducer rules, and the construction procedure itself is not revised.
Clinical setting
Retrospective English radiology report corpora in whole-body 18F-FDG PET/CT (PETWB-REP, ViMed from a Vietnamese tertiary hospital, machine-translated) and abdominal CT (MERLIN, RadGPT-generated BDMAP reports); no prospective or workflow deployment.
Evidence boundary
Offline corpus-to-template construction evaluated with largely LLM-based metrics on small test splits (n=49 to 408); expert raters are not described, and the authors state that metrics and expert assessments do not establish clinical safety or deployment readiness.
AgentClinic: a multimodal benchmark for tool-using clinical AI agents
Authors
Samuel Schmidgall; Rojin Ziaei; Carl Harris; Ji Woong Kim; Eduardo Pontes Reis; Jeffrey Jopling; Michael Moor
Year
2026
Publication status
J/C: npj Digital Medicine 9, Article 499 (2026), open access; received 19 April 2025, accepted 14 April 2026, published 27 April 2026, version of record 02 July 2026.
Version
Published version of record (nature.com HTML, fetched 2026-11-03); arXiv 2405.07960v5 (2025-05-25) used only for the tool table and notebook prompt, which the published HTML does not contain.
A doctor-agent memory notebook of general diagnostic tips that persists across patient cases (in-context text, no weight updates); it is one of six Agent Toolbox tools in a benchmark paper.
Fixed components
Doctor LLM backbone weights, patient, measurement and moderator agents (GPT-4 patient/measurement), prompts, the N = 20 interaction budget and the case set are fixed.
Feedback
Per arXiv v5 Appendix M.2, after each case the notebook-update prompt shows the conversation, the correct diagnosis and the agent's own diagnosis; the published Methods describe ground-truth verification of the final answer.
Timing
Online and sequential: the notebook is rewritten after every case and used on the following AgentClinic-MedQA cases.
Retained artifact
A single notebook capped at 1000 characters that the agent overwrites each time and must re-integrate prior notes into (v5 Appendix M.2); the published Methods additionally describe a success-only vector store of reflective takeaways.
Update generator
The doctor LLM itself under a fixed notebook-writing meta-prompt, instructed to write general, non-patient-specific tips from successes and failures.
Selection and release
No acceptance test, validation gate or human review of notes is described; whatever the agent writes is truncated to the character limit and used on the next case.
Evaluation separate from updating
Not reported as a held-out split: the notebook gain is measured as AgentClinic-MedQA diagnostic accuracy over the same sequential stream in which notes are written (each case is seen before its own note update), with no seeds or intervals for the tool table.
Fixed-procedure control
Compared against a no-tool baseline per backbone and against five other tools (zero-shot, one-shot and reflection CoT, adaptive RAG book and web) on AgentClinic-MedQA, for six backbones.
Evidence concerning improver change
Not reported: the notebook-writing prompt and update procedure stay constant; only the notebook contents change.
Clinical setting
Simulated diagnostic dialogues built from USMLE-style MedQA vignettes (tool experiment), with LLM patient, measurement and moderator agents; no real patients or deployment.
Evidence boundary
Benchmark sub-experiment in simulation with an LLM moderator as judge; the notebook effect is backbone dependent (positive for four models, negative for GPT-3.5 and GPT-4o-mini) and is not tested on held-out cases or in clinical use.
J/C: MICCAI 2026 book chapter in LNCS, pp. 430-440, Springer, published 2026-09-25 (Crossref confirmed); extraction uses arXiv 2605.22872v1, which its comment labels 'MICCAI 2026 Early Accept. Submission Version'.
Version
arXiv 2605.22872v1 (2026-05-20, only version listed by the arXiv API on 2026-11-03); the LNCS chapter text was not examined.
An external, per-model experience memory of structured pairwise differential-diagnosis notes (differentials, confusions, discriminators, decision_rule, error_analysis), indexed by 11 departments and 118 organ/region paths.
Fixed components
VLM parameters are not updated (Qwen3-VL 2B/8B/30B, InternVL3.5 8B/30B, Lingshu 7B/32B); the S-PubMedBert-MS-MARCO retrieval encoder, threshold tau=0.9, top-K=10 and the prompts are fixed.
Feedback
On construction cases the agent's diagnosis is compared with the Eurorad ground-truth diagnosis, and the expert discussion (imaging findings and diagnostic reasoning) is supplied when extracting a note from each error.
Timing
Offline construction in two passes over 6,228 pre-2025 Eurorad cases (Phase I zero-shot, Phase II re-diagnosis with memory), then inference on 227 cases from 2025; updating memory during the test set is not reported.
Retained artifact
The pairwise differential notes, merged per differential pair, persist as a separate memory library built independently for each VLM.
Update generator
The same diagnostic agent (VLM) extracts each note from its own misdiagnosis, instructed to identify confusion points, discriminators, bidirectional decision rules and its reasoning failure.
Selection and release
Notes are written only for errors (d != d*); Phase II adds notes for stochastic errors and note deficiencies, refines notes that caused correct-to-incorrect transitions and merges insights into existing pairs; no separate validation gate or human review of notes is reported, and at test time retrieved notes are filtered by threshold, top-K and an agent relevance score.
Evaluation separate from updating
Accuracy on a temporally held-out set of 227 Eurorad 2025 cases versus zero-shot baseline, with retrieval recall/precision and beneficial/harmful case rates; the answer-matching method, number of runs, confidence intervals and significance tests are not reported.
Fixed-procedure control
Zero-shot baseline for all seven models, a PubMed-RAG (top-3 articles) comparison and construction/retrieval ablations, all on Qwen3-VL-8B only for the comparison and ablations.
Evidence concerning improver change
Not reported: the extraction, Phase II refinement and retrieval procedures are fixed by design; the one-round versus two-round construction ablation compares preset procedure variants rather than a procedure that revises itself.
Clinical setting
Retrospective evaluation on Eurorad radiology teaching-file cases (CT, MRI, X-ray, ultrasound with clinical history); no prospective, reader-study or deployment evaluation is reported.
Evidence boundary
Gains are single-benchmark accuracy changes of +1.7 to +7.0 points without reported variance or tests, with harmful (correct to wrong) rates of 1.3% to 5.8% and ablations on one model.
A long-term memory of past successful question and code solutions, retrieved as the K=4 few-shot demonstrations (§3.2, Eq. 1).
Fixed components
GPT-4 (gpt-4-0613, temperature 0) backbone, tool definitions, EHR metadata, prompts and Levenshtein-similarity retrieval stay fixed (§3.2, §4.1).
Feedback
Ground-truth success of each answer decides memory admission; execution error messages drive within-case debugging only (§3.2, App. F.2).
Timing
Online across the evaluated question stream; no separate training split for memory is reported (§3.2, App. F.2).
Retained artifact
Successful past cases (question plus code solution) that replace the original hand-written demonstrations (§3.2, App. H.2 Fig. 9).
Update generator
The agent's own successful solving runs; no separate generator component.
Selection and release
A case enters memory only when it succeeds against the ground-truth answer; a variant admitting all completed cases is also tested (Table 8).
Evaluation separate from updating
Not reported: success rate is measured on the same question stream from which memory accumulates (§4.1, App. F.2).
Fixed-procedure control
Ablation without long-term memory under the same backbone and prompts on MIMIC-III and eICU (Table 2, Table 7); shuffled-order runs reported as almost no influence (App. F.2).
Evidence concerning improver change
None; the storage rule and retrieval procedure are fixed.
Clinical setting
Multi-hop question answering over structured EHR tables of MIMIC-III, eICU and TREQS (§4.1).
Evidence boundary
Offline benchmark with ground-truth labels used at test time for memory admission; no clinician-facing or prospective evaluation (§4.1, Limitations).
A long-term memory of successful reflective trajectories and a per-tool store of tool-wise experience (§3.2, Eq. 3-4).
Fixed components
Qwen2-7B/72B backbone (72B with 4-bit GPTQ), 15-type clinical toolbox, prompts, BM25 retrieval and verifier procedures stay fixed (§2.2, §3.3, §4.2).
Feedback
Ground-truth answers on the optimization subset; the agent reflects on its first trajectory against the answer and retries (§3.2).
Timing
Offline optimization stage before inference; memory and experience are then frozen and only read at inference (§3.2-3.3).
Retained artifact
Successful refined trajectories as memory items, plus merged action-wise suggestions for each tool and for Plan, Think and Finish (Algorithm 1).
Update generator
The same LLM agent writes reflections and tool-wise suggestions by comparing failed and successful trajectories (Eq. 4, App. D.2 Fig. 8-9).
Selection and release
A trajectory is stored only when its refined answer equals the ground truth; tool experience is merged without further validation (Eq. 3).
Evaluation separate from updating
Test samples of CAB distinct from the optimization subset; ablations on 200 random samples per dimension; optimization subset size not reported (§3.2, §4.3).
Fixed-procedure control
Ablations removing reflective memory or tool-wise reflection (Table 3); baselines given the same long-term memory (Table 4); performance versus optimization steps (Fig. 3).
Evidence concerning improver change
None; the reflection and storage procedure is fixed.
Clinical setting
ClinicalAgent Bench: 18 public tasks over knowledge, multimodal, numerical, data understanding and trustworthiness dimensions (§2.1, App. B).
Evidence boundary
Offline benchmark; authors note correct answers do not guarantee correct trajectories (Limitations); no clinician-facing or prospective evaluation.
Two knowledge bases: CorrectKB for correct consultations and ChainKB for reflective analyses of incorrect ones (§3.3).
Fixed components
gpt-4-turbo backbone in the main text (gpt-5 for multimodal), agent roles, Lead Physician and Residual Context protocol, and text-embedding-3-small retrieval (§3, §4.2, App. C.3).
Feedback
The final consultation answer is validated against the ground truth, which routes the record to CorrectKB or ChainKB (§3.3).
Timing
Accumulated over 900 consultation rounds on training sets before testing; test-time KBs contain only training-set experience (§4.1-4.2).
Retained artifact
Structured summaries: question, answer and final-round statements (CorrectKB); hypothesis, analysis, conclusion and error reflection (ChainKB) (§3.3).
Update generator
A Chain-of-Thought Reviewer agent summarizes each finished consultation into the KB format (§3.3).
Selection and release
Routing by ground-truth correctness only; no further validation, pruning or release gate is reported (§3.3).
Evaluation separate from updating
Official MedQA and PubMedQA test sets, five runs with SD; cross-dataset KB transfer (Table 3); VQA-RAD and SLAKE official test sets (App. C.2).
Fixed-procedure control
Ablation adding CorrectKB, ChainKB or both with the discussion protocol fixed (Table 2); accuracy versus accumulated cases for eight backbones (Fig. 4).
Evidence concerning improver change
None; the reviewer and routing rule are fixed.
Clinical setting
Simulated multi-disciplinary team consultation on exam and research-abstract QA, radiology VQA and self-collected complex cases (§4.1, App. C).
Evidence boundary
Offline benchmarks with labels used for KB admission; no deployment, authors plan shadow-system validation (§6 Limitations); human evaluation on 20 samples (App. C.8).
The agent workflow graph (nodes, edges, conditional, loop and parallel structures) and node system and human prompts (§3.1-3.3).
Fixed components
Backbone LLM (GPT-4o, GPT-4o-mini or Claude 3.5 Sonnet, temperature 1.0), search-space operations, CLIP image-search tool and reference database, evolution prompts (§3.2-3.3, §4.1).
Feedback
Diagnostic errors against ground-truth labels, which the LLM classifies as image-understanding or diagnostic errors (§3.3).
Timing
Offline batched iterations over the training split, validated after each batch until validation accuracy converges; final workflow frozen for testing (§4.1).
Retained artifact
The final evolved workflow, e.g. 11 nodes and 5 branches at test time (Table 2).
Update generator
The LLM analyses failures and emits structural and prompt suggestions, reformulated into structured modifications via templates (§3.3).
Selection and release
Infeasible suggestions are filtered, structural edits pass graph-validity checks, and each updated workflow is checked on the 50-image validation set (§3.3, §4.1).
Evaluation separate from updating
Held-out test splits: 224 Skin Concepts images (before excluding 100 images for moderation or broken links) and 120 Skin Conditions images (§4.1).
Fixed-procedure control
Ablations disabling individual operations (Table 3) and operation levels (Table 4); static IO, CoT and Round Table baselines (Table 1); single seed 42.
Evidence concerning improver change
None; the analysis, suggestion and validation procedure is fixed.
Clinical setting
Dermatology image diagnosis on Skin Concepts (Fitzpatrick17k extension) and the synthetically augmented Skin Conditions dataset (§4.1).
Evidence boundary
Small 50/50 training and validation splits, one seed, one synthetic dataset; offline benchmarks only, no clinician or prospective evaluation (§4.1).
MACD: Multi-Agent Clinical Diagnosis with Self-Learned Knowledge for LLM
Authors
Wenliang Li; Rui Yan; Xu Zhang; Li Chen; Hongji Zhu; Jing Zhao; Junjun Li; Mengru Li; Wei Cao; Zihang Jiang; Wei Wei; Kun Zhang; Shaohua Kevin Zhou
Year
2025
Publication status
P: arXiv preprint (cs.AI), v1 2025-09-24, v5 2026-08-24; no comments or journal_ref; Semantic Scholar lists only arXiv (checked 2026-11-04). v5 declares an npj Digital Medicine editor conflict, suggesting submission there; no acceptance found.
Disease-specific Self-Learned Knowledge (general and rare diagnostic concepts) injected into the diagnostician prompt.
Fixed components
Base LLMs (Llama-3.1-8B/70B, DeepSeek-R1-Distill-Llama-70B), prompt templates, summarizer and refiner procedures (BioBERT MMR de-duplication to top-7, concept ablation), decoding settings, matching rules.
Feedback
Ground-truth diagnosis correctness on learning cases; refiner uses change in learning-set diagnostic accuracy when each concept is ablated.
Timing
Offline accumulation over the learning set (K_t to K_{t+1} iterations; number of iterations not reported), then frozen knowledge applied to the test set.
Retained artifact
Per-disease, per-model Self-Learned Knowledge text (Supplementary Tables 10-12), also transferred to other LLMs.
Update generator
Diagnostician diagnoses learning cases with current knowledge; Summarizer agent abstracts insights from correct diagnoses; Refiner agent merges, de-duplicates and removes concepts whose ablation does not lower accuracy.
Selection and release
Learning cases sampled from cases correctly diagnosed by at least one base agent (90 per disease per agent; all 23 for pericarditis); concepts kept by the refiner's redundancy and causal-ablation filters; no clinician release gate.
Evaluation separate from updating
Held-out MIMIC-MACD test cases never used for learning; comparisons with Mayo Clinic and guideline-derived knowledge, CoT, few-shot and LoRA fine-tuning; cross-model transfer; clinician concept-relevance ratings and a 140-case physician reader study.
Fixed-procedure control
Summarizer-only versus summarizer plus refiner ablation and knowledge-composition ablations under the same pipeline; no evolving-versus-fixed improver comparison.
Evidence concerning improver change
Not established: summarizer, refiner and selection rules stay fixed; only the knowledge text changes.
Clinical setting
Primary diagnosis of seven acute abdominal and chest diseases from MIMIC-IV emergency/inpatient text (history, examination, laboratory results, radiology reports).
Evidence boundary
Single-institution retrospective text vignettes; test set is the non-random residue after success-based learning-case selection; physician comparison on text-only vignettes; number of repeated runs not reported.
The system's own Reflector critique and pass@k draft scoring; no ground-truth feedback described for memory writes.
Timing
Memory updated after each processed case during inference.
Retained artifact
Self-generated final reports with cues and tags, reused as few-shot exemplars.
Update generator
Cue and tag functions applied to the final repaired output.
Selection and release
Outputs judged not to be a system error are stored; the criterion is not further specified.
Evaluation separate from updating
IU Chest X-ray report generation and VinBigData localization; GPT-4o-mini reference-based judge, overlap metrics and mAP50; test sizes and case order not reported.
Fixed-procedure control
Not reported; only a pass@k ablation (Fig. 2).
Evidence concerning improver change
None.
Clinical setting
Public retrospective chest X-ray datasets.
Evidence boundary
Reported gains bundle routing, reflection, repair and memory; stored exemplars are unverified self-generated outputs.
C: EMNLP 2026 main conference (ACL Anthology 2026.emnlp-main.1529); arXiv v1 examined, key numbers (0.2880, 0.2117, 100 training examples, top-1) also present in the published PDF.
Version
arXiv v1 (2026-01-20), cross-checked against EMNLP 2026 PDF
External Experience Memory Bank of actor heuristics and summarizer salience guidelines keyed by EHR embeddings.
Fixed components
LLM backbone weights (no parameter updates), RetroSum actor and summarizer prompts, MCP toolbox of 19+ tools, bge-m3 retriever with top-1 retrieval.
Feedback
Ground-truth answer sets of training instances compared with the agent's trajectory, summary and predictions by a reflection prompt.
Timing
Offline before evaluation on 100 MIMIC-IV-Common training examples per task; the same accumulated memory is then reused on all three test subsets.
Retained artifact
Triplets (EHR embedding, actor experience, summarizer experience), one per training instance.
Update generator
Reflection module that shares the agent's frozen LLM, prompted to extract procedural tool-use heuristics and information-salience guidelines.
Selection and release
No filtering, validation or pruning of generated experiences is reported.
Evaluation separate from updating
Held-out MIMIC-IV-Common cohorts plus MIMIC-IV-Rare (label shift) and MIMIC-III (schema shift) with the same memory; evolved versus non-evolved ablations; ReflecTool and ReasoningBank comparators across five backbones.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
None; only memory contents change.
Clinical setting
Retrospective hospital and ICU EHR databases (MIMIC-IV, MIMIC-III) queried through tools for six prediction tasks such as diagnoses and prescriptions.
Evidence boundary
Single-center MIMIC data; memory built once offline; disjointness of the 100 experience examples from test cohorts not stated explicitly; significance tests compare against ReSum, not against non-evolved RetroSum.
Ground-truth attributes, boxes, malignancy and BI-RADS labels: attribute accuracy plus format reward (Stage 1), diagnosis correctness with oracle attributes (Stage 2), refined trajectories for SFT (Stage 3).
Timing
Three sequential offline training stages before testing; agents fixed at inference.
Retained artifact
Trained sub-agent weights, distilled main-agent weights, and the refined trajectory set used for SFT.
Update generator
Stage-2 policy rollouts; the Stage-2 model rewrites rationales of misdiagnosed samples conditioned on ground truth; predicted boxes replaced by ground-truth boxes.
Selection and release
All sampled trajectories are refined and used; no further selection or validation gate is reported.
Evaluation separate from updating
Held-out BUSBRA, BUSI and BUDIAT splits (BUS-CoT protocol) and external BrEaST OOD set; ablations without oracle training and without self-distillation; GT-box and GT-attribute upper bounds.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
The trained Stage-2 policy also rewrites its successor's SFT data; this role is not isolated from box replacement in the ablation.
Clinical setting
Breast ultrasound image diagnosis (malignancy and BI-RADS) on public datasets.
Evidence boundary
Single 3B backbone and single runs without variance; BI-RADS agreement low (overall kappa 0.224, OOD 0.037); no learning after deployment.
Described as updating when outcomes become available; the order and timing of memory population in the reported experiments are not reported.
Retained artifact
Patient feature profiles with outcomes; empirical outcome rates for feature combinations shared by three or more cases.
Update generator
Rule-based: cases appended and subgroup rates recomputed; no LLM-generated memory content.
Selection and release
Patterns formed only when at least three cases share a feature combination; no other gate reported.
Evaluation separate from updating
35 real patients and 400 synthetic cases; single-LLM, MedAgents, MDAgents and MedAgent-Pro baselines; w/o SEA-Mem ablation on real data.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
None; the update rule is fixed.
Clinical setting
Pre-therapy outcome prediction for 177Lu-PSMA radioligand therapy in mCRPC from PSMA PET/CT reports, laboratory tests and clinical notes.
Evidence boundary
Single hospital, n=35; exclusion of the query patient from memory (leave-one-out or sequential order) not reported; synthetic cases are LLM-generated; no variance or significance tests.
Accumulated clinical experience of 38 physician agents produced by post-case reflection; storage form not specified.
Fixed components
Qwen3-Next-80B-A3B weights for physician and patient agents; Agent-Kernel action space; synthetic patient records and personas.
Feedback
Ground-truth diagnosis, gold examination list and guideline-based treatment comparison used in post-case self-critique.
Timing
After each case, across 22 sequential training batches of about 500 cases.
Retained artifact
Diagnostic, investigative and treatment reflections described as accumulated clinical experience; format and retrieval not described.
Update generator
Physician agents' own multi-dimensional self-critique.
Selection and release
Not reported.
Evaluation separate from updating
A 9:1 train/test split of 12,000 synthetic records is stated, but reported trajectories are plotted across training batches; no ablation or comparator system.
Fixed-procedure control
Not reported.
Evidence concerning improver change
None reported.
Clinical setting
Simulated multi-department hospital with LLM patient agents built from synthetic records (583 diseases, 19 departments).
Evidence boundary
Fully synthetic, text-only arena; mechanism under-specified; no no-reflection control or repeated runs; authors state it is not for clinical decision support.
EvoDerma-Mem graph database of historical cases (image embedding, key findings, expert diagnosis) and versioned per-disease diagnostic guidelines.
Fixed components
Qwen3-VL vision agent, fine-tuned disease-diagnosis model, Qwen3 case-review and summarization agents and prompts, Oxford-handbook RAG store, image feature extractor.
Feedback
Expert diagnosis labels of stored historical (training) cases.
Timing
Guidelines rewritten when new samples reach threshold N_thresh; the N_thresh value and whether test cases are added are not reported.
Retained artifact
Case triplets and guideline versions G_C^t per disease.
Update generator
LLM Summarization Agent that merges the previous guideline with new key findings.
Selection and release
Threshold-triggered update with contradiction resolution inside the summarizer; no validation gate before use reported.
Evaluation separate from updating
Held-out test splits of Dermnet, HAM10000, DDI31, Fitzpatrick-17k, Dermnet498 and RSDD; with/without memory ablation; blinded review of 600 evolutionary cases by two dermatologists.
Fixed-procedure control
No comparison of a changing versus a fixed improvement procedure.
Evidence concerning improver change
None; the summarizer procedure is fixed (each guideline update takes the previous version as input).
Clinical setting
Dermatology image diagnosis on public clinical and dermoscopic datasets and a web-curated rare-disease set.
Evidence boundary
Offline public benchmarks; memory built from labelled training data; RSDD images gathered from websites; no prospective use.
Qwen3-4B/8B policy weights trained by round-wise GRPO to diagnose and issue memory operations; at deployment the short-term case memory (capacity K) and long-term rule memory change while weights stay frozen (Sec 3.1-3.3, App F.2.2).
Fixed components
Closed-set candidate pipeline (ground truth plus 199 GPT-5.2-ranked distractors, fixed seeds), memory tool schema (list/append/pop/consolidate), reward weights and schedule, decoding settings; weights frozen in the long-horizon stream (Sec 4.1.1, App D, App E.1-E.2).
Feedback
Ground-truth diagnosis returned after each case; training reward of +5/-5 for correctness plus a short-term-memory occupancy penalty (alpha = 3), linearly re-weighted across rounds (Sec 3.2-3.3, Sec 4.2.1).
Timing
Offline RL on the MedCaseReasoning training split (about 48 GPU-hours on 8 A100 for Qwen3-8B); online memory updates after every case in the ER-Reason stream (Sec 4.2.1, App E.2).
Retained artifact
Outcome-annotated short-term case records and natural-language diagnostic rules consolidated from evicted cases, carried forward across stream rounds (Sec 3.1).
Update generator
The trained policy itself decides when to append or pop cases and writes the consolidated rules through tool calls (Sec 3.1, App E.1).
Selection and release
No gate on memory writes beyond the agent's own consolidate actions; training checkpoint chosen by best validation accuracy (Sec 3.1, App E.2).
Evaluation separate from updating
MedCaseReasoning official test split (each case independent, 10 training cases as memory context, App F.2.3); ER-Reason never trained on and streamed once with frozen weights (Sec 4.2.1, Table 4).
Fixed-procedure control
Same dual-memory tool and prompt with untrained or SFT writers (Qwen-8b SFT + Dual Memory, GPT-5.2 + Dual Memory) and memoryless SFT/RL baselines (Table 2); component ablations on MedCaseReasoning (Table 3).
Evidence concerning improver change
RL with a memory reward revises the policy that writes memory, and stream gains are compared with untrained writers using the identical memory tool (Table 2); writer and diagnostician share weights, so the improver effect is not separated from reasoning gains.
Clinical setting
Closed-set diagnosis (1 of 200 candidates) from published case reports (MedCaseReasoning) and emergency-room cases (ER-Reason); no deployment (Sec 4, App B).
Evidence boundary
Closed-set labels with immediate ground-truth feedback; one stream order, no seeds or intervals for stream metrics; Table 1 and Table 3 gains come from a setting without cross-case carryover; rule quality rated by one internal medicine resident; preprint.
Full parameters of Qwen3.5-9B trained from scratch (no SFT warm-up) by vanilla GRPO or TT-OPD, i.e. GRPO with cosine length reward plus turn-level KL to an EMA teacher, for 60 steps (Sec 4.2, Sec 5.1, App F).
Fixed components
Gym with 10 domains, 3,631 tasks (2,657 train, 307 validation), 135 tools, 828K-passage knowledge base, domain policy prompts, 5D reward weights, AgentRunner; EMA rule alpha = 0.995 every 5 steps with hard copy every 30 steps (Sec 3, Sec 4.2, App A).
Feedback
Episode-level 5D reward (accuracy 0.25, process 0.20, safety 0.20, format 0.10, coherence 0.10) plus cosine length reward; outcome-conditioned hints given only to the teacher (Sec 3, Sec 4.2).
Timing
Online RL during training (60 steps on 8 A100 80GB); no updates at evaluation (Sec 5.1).
Retained artifact
Trained policy checkpoint, evaluated with the same multi-turn AgentRunner, tools and knowledge base (Sec 5.2).
Update generator
Student on-policy rollouts in the gym; an EMA copy of the student, conditioned on outcome hints, supplies per-turn KL targets (Sec 4.2).
Selection and release
Not reported; results given for TT-OPD at step 60 and GRPO peak at step 55 on the validation set (Sec 5.3).
Evaluation separate from updating
307 held-out validation tasks from the training distribution; 18 external benchmarks (MC QA, VQA, EHR, long-form QA) with test-set fingerprinting for contamination (Sec 5.1, Table 1, App G).
Fixed-procedure control
Base+AR (same tools and knowledge base, no RL), vanilla GRPO with identical hyperparameters, four OPD variants adding one component each (Table 1, Sec 6.1, Fig 4).
Evidence concerning improver change
The teacher that shapes each update is itself updated as an EMA of the student; ablations compare teacher rules (periodic reset, EMA, EMA plus hints, full) on validation curves, not the learned teacher's contribution alone (Sec 6.1).
Clinical setting
Simulated tool-based clinical tasks (EHR, triage, drug interaction, visual and clinical diagnosis, psychiatry, obstetrics, pathways) and offline medical QA/VQA/EHR benchmarks; no clinician or patient involvement (Sec 3, App A.2).
Evidence boundary
Training tasks are converted from the same benchmark families used for testing; single run without seeds or intervals; text says best on 10 and on 12 of 18 benchmarks; below text-only base on MMLU-Med, SLAKE, PMC-VQA; preprint.
Shared library of verified Python functions, one per MIMIC-IV clinical concept (63 concepts), built offline and then given to inference agents (Sec 4.1-4.3).
Fixed components
Inference LLM weights, MIMIC-IV schema and query_db tool, LLM-generated guidelines from PubMed abstracts, verification agent, threshold 0.90, up to 100 iterations; the library built with Qwen3.5-27B is reused for every backbone (Sec 4, Sec 5, App D, App F).
Feedback
Accuracy of each candidate function on the 10% verification split (about 100 labeled rows per variant), with per-question error traces fed back to the builder (Sec 4.2, App A).
Timing
Offline construction per concept in topological dependency order (about 20 GPU-hours per concept); library frozen as a static resource at inference (Sec 4.3, App I).
Retained artifact
Library of verified concept functions exposed as signatures; later concept builds can search and load earlier verified functions (Sec 4.1, App G.1).
Update generator
ReAct autoformalization agent (Qwen3.5-27B) that explores the schema, retrieves guidelines and writes code (Sec 4.1).
Selection and release
Function saved to the shared library when verification accuracy reaches 0.90; otherwise the best candidate across iterations is retained (Sec 4.2, App F Algorithm 1).
Evaluation separate from updating
Held-out 90% of compositional queries (56,700 test vs 6,300 train) and the longitudinal ICU surveillance task (2,000 held-out stays, subject-level split) not used to build the library (Sec 3.1-3.2, Sec 5).
Fixed-procedure control
Zero-shot agent with the same schema and query_db tool; GRPO fine-tuning with 50 examples per concept on the same backbone (Table 3, Fig 6-7).
Evidence concerning improver change
Not reported; the build-and-verify procedure stays fixed.
Clinical setting
Retrospective MIMIC-IV ICU EHR: compositional concept extraction (e.g. SOFA, Sepsis-3) and 4-hourly surveillance over 48 h with 13 checkpoints (Sec 3).
Evidence boundary
Single institution; library built by one model; small verification sets; LLM-written guidelines; frontier models run on a 21% subset; longitudinal gains shown only in Fig 6 without a numeric table; preprint.
Full parameters of Qwen3-4B-Instruct through Stage I SFT (about 100k instances), Stage II behavior cloning (about 1k), Stage III GRPO (about 3k trajectories, 300 steps) in the world model (Sec 4, App G.2).
Fixed components
Clinical World Model trained beforehand on the MIMIC-IV training split, GPT-4.1-synthesized rationales and propose-simulate-refine traces, reward coefficients, guideline constraints, 5x5 fluid-vasopressor action grid (Sec 3, Sec 4.2-4.4, App G.3).
Feedback
Composite reward from world-model rollouts: survival +15 or mortality -15, SOFA and lactate step terms, -10 per guideline violation, -5 for not terminating; scaled by 0.1 and clipped to [-2, 2] (Sec 4.4, App G.3).
Timing
Offline staged training; no updates during evaluation (Sec 4, Sec 5.1).
Retained artifact
Trained agent checkpoint used with the propose-simulate-refine workflow at inference; outcome prediction also tested without simulator access (Sec 2.3, Sec 5.4).
Update generator
GPT-4.1 for Stage I-II supervision; the agent's own multi-turn rollouts inside the world model for Stage III (Sec 4.2-4.4).
Selection and release
Not reported beyond terminating GRPO after 300 steps (App G.2).
Evaluation separate from updating
725-episode held-out MIMIC-IV test set with off-policy estimators (DR, WIS, WPDIS), guideline adherence and unsafe-action rates never optimized in training; 2,862 MIMIC-III episodes (Sec 5.1, App I).
Fixed-procedure control
Same world model and workflow without training (o3, Gemini-3-Flash, GPT-4.1-mini + WM), untrained base model, and cumulative stage ablation (Table 3, Table 4).
Evidence concerning improver change
Not reported.
Clinical setting
Retrospective ICU sepsis fluid and vasopressor dosing at 4-hour steps on 20,092 MIMIC-IV stays; no clinician review or deployment (Sec 3, App C).
Evidence boundary
Off-policy value estimates only; world model and policy trained on the same source; MIMIC-III check comes from the same hospital system; no seeds or intervals; authors state not for clinical use; preprint.
Clinician accept or reject per proposal with structured rejection rationale; final accepted settings and a Note-Generator-extracted preference signal drive the update (Sec 2.2-2.3).
Timing
Online, one bandit update at the close of each adjustment cycle (Algorithm 1).
Retained artifact
Clinician preference state passed to the reasoning engine as preference context that modulates candidate ranking in later cycles; summaries written to long-term memory (Sec 2.3).
Update generator
Note Generator LLM extracts the preference signal; BanditUpdate rule, whose algorithm is not specified (Sec 2.3).
Selection and release
Updates only from clinician-accepted outcomes, not from rejected proposals; no further gating reported (Sec 2.3).
Evaluation separate from updating
Retrospective replay on a multi-center ICU cohort (1,309 records, 7,447 setting entries, 13 modes) for next-setting prediction; expert ratings of 100 cycles (Sec 3, Table 1).
Fixed-procedure control
Single-model end-to-end generation with the same backbones; NoPref, NoImg and NoImgNoPref ablations within VDSS (Table 1).
Evidence concerning improver change
Not reported.
Clinical setting
ICU mechanical-ventilation titration with clinician-in-the-loop review, retrospective trajectory replay and prototype interface (Sec 3-4).
Evidence boundary
Cross-cycle learning shown only by a regret trend over 100 cycles from one clinician (Fig 2) without numbers or statistics; patient span, number of experts, replay feedback source and bandit algorithm not stated.
A non-parametric experience pool (ExPool) of rejection-sampled reasoning traces and the QLoRA weights of the worker and manager GPT-OSS-20B agents (multi-agent reward-ranked fine-tuning).
Fixed components
Traj-CoA chain-of-agents architecture and prompts, nomic-embed-text-v1.5 embedder, label-conditioned selection rule, binary reward thresholds (s>=6 for cases, s<=4 for controls), and the test sets.
Feedback
Ground-truth one-year lung cancer labels of training patients, used for best-of-4 label-conditioned selection (ExPool) and a binary acceptance reward (MARL) (Sec 3.3-3.4).
Timing
Offline before testing; a stream of verified patients is simulated by incrementally sampling the training set, MARL is one QLoRA epoch, and inference uses the full training-set ExPool (Sec 3.1, 3.5, 4).
Retained artifact
ExPool vector store of (summary embedding, manager rationale, predicted risk, label) for 13,629 training patients, plus fine-tuned worker and manager adapters.
Update generator
The base system's own rollouts (m=4, temperature 1.5) filtered by fixed rejection sampling; in the combined variant MARL rollouts are conditioned on leave-one-out ExPool retrieval (Sec 3.5).
Selection and release
No accept/reject gate on updates; hyperparameters and thresholds set on a validation split; single training round released (Appendix B).
Evaluation separate from updating
Disjoint held-out test sets from the same institution (overall n=1,000; never-smoker n=835), bootstrap 1,000 resamples (Sec 4).
Fixed-procedure control
Static zero-shot Traj-CoA and RAG with the same backbone, plus ExPool-only, MARL-only and combined variants (Tables 1-2).
Evidence concerning improver change
None; rejection rule, reward and training recipe are fixed.
Clinical setting
One-year incident lung cancer prediction from up to five years of multimodal EHR (codes, labs, notes, radiology reports) in a 1:10 matched case-control cohort from one medical center.
Evidence boundary
Single-institution retrospective case-control design with a simulated patient stream and one MARL round; authors note label delay and need for prospective, multi-institution validation (Limitations).
The doctor agent's external memory bank of success, failure-correction and action-rule cards, retrieved by TF-IDF (top K=3) before later decisions.
Fixed components
Doctor model (Qwen-Turbo), patient simulator (Qwen-Turbo), manager scorer (Kimi-K2.5), test catalog, rubrics, case order, action budget of 10.
Feedback
Manager-side rubric scores of diagnosis, evidence and management after FINALIZE, with access to the reference diagnosis (Sec 3.3-3.4).
Timing
Online sequential write-back after each episode over a 500-case stream (passes E1-E3) and continued during held-out evaluation (H3).
Retained artifact
Unbounded memory bank; 4,500 records after E3 (312 success, 1,188 corrective, 3,000 rule cards) (Appendix B.4).
Update generator
The manager LLM writes compact IF-THEN and IF-AVOID cards after scoring.
Selection and release
No acceptance gate or eviction; every card is written back.
Evaluation separate from updating
200 held-out cases sharing no content with the stream, but same source and H3 keeps writing memory during evaluation; two-physician audit of 80 scored outputs.
Fixed-procedure control
No-memory conditions E0 and H0 with identical doctor, simulator, order and scorer; frozen-bank relevant vs case-mismatched retrieval control (Table 27).
Evidence concerning improver change
None.
Clinical setting
Simulated Chinese outpatient first-visit diagnostic episodes converted from MedQA-derived cases, with action-gated inquiry and test requests.
Evidence boundary
Simulated standardized episodes; stream gains E2-E3 combine repeated exposure with memory; held-out clinical score and strict accuracy not significant after Holm (Sec 6, Table 21).
Long-term memory of structured diagnostic entries (failed descriptor, diagnosed cause, successful correction), accumulated within a dataset and reset between datasets.
Fixed components
GPT-4o backbone at temperature 0, 21 perception and descriptor tools, distilled skill set (properties, tiered rankings, parameters), Perception-Reasoning-Action-Reflection pipeline, classifiers.
Feedback
LLM Reflection verdict on feature-vector quality (sparsity, variance, kurtosis and similar statistics); no ground-truth labels enter the memory (Sec 3.2-3.3).
Timing
Online during runs; an entry is written whenever Reflection fails and triggers a retry.
Retained artifact
Per-dataset long-term memory entries consulted by later Determination and Reflection steps.
Update generator
The LLM Reflect step writes the diagnosis and correction (Algorithm 1, line 17).
Selection and release
No gate; entries are written on failure without verification.
Evaluation separate from updating
26 TopoBenchmark datasets with 5-fold CV on frozen splits, but the skill set was distilled from the same benchmark at object-type level; three external datasets for downstream integration.
Fixed-procedure control
Ablations without dual memory and without long-term memory under the same pipeline (Table 2).
Evidence concerning improver change
None.
Clinical setting
Selecting topological descriptors for 2D medical image classification across 11 modalities (histopathology, retina, X-ray, dermoscopy and others).
Evidence boundary
Memory is per dataset and reset; skill set built from evaluation datasets; endpoint is classifier accuracy, not clinical decisions; about 32 s per image (Sec 5).
Episodic memory of the agent's own stage decisions with rewards, and a semantic memory of up to 10 stage-indexed stop/acquire rules.
Fixed components
Frozen Qwen-2.5-7B-Instruct, frozen Transformer survival predictor and uncertainty head, text tools, FAISS index of training cases, MCDA burdens and reward formulas.
Feedback
Stage reward (-burden + uncertainty reduction) and terminal reward ((1-uncertainty) - total burden); top and bottom reward quartiles label episodes good or poor to score rule effectiveness (Sec 2.3).
Timing
Experience accumulation on training patients, 3 episodes each (about 408 episodes per fold), reflection every 10 patients, 10 cycles per fold (Sec 3.2, 3.5).
Retained artifact
Episode store and active rule set; four rules recurred in at least 4 of 5 folds (Table 3).
Update generator
The frozen LLM's reflection module proposes rules from an effectiveness report.
Selection and release
Proposed rules are checked for direction clarity and novelty; rules with consistently poor outcomes are deprecated; 10.8 candidates and 4.6 survivors per fold (Sec 3.5).
Evaluation separate from updating
Nested 5x5 cross-validation on 170 complete-modality patients with outer-fold test patients; whether memory keeps updating on test patients is not reported.
Fixed-procedure control
Component ablation from base LLM through tools, episodic and semantic memory with the same frozen LLM (Table 2); Reflexion baseline (Table 1).
Evidence concerning improver change
None.
Clinical setting
Glioma survival prediction with ordered acquisition of demographics, radiology, pathology and genomics (TCGA-LGG, TCGA-GBM, BraTS).
Evidence boundary
Small public cohort (170 evaluation patients); rewards use model uncertainty and burden rather than outcomes; C-index slightly below full-workup fusion.
The post-training data mixture and resulting Kimi-K2.6 (1T) weights, trained per round by SFT screen, RL and self-distillation fine-tuning.
Fixed components
Base model, benchmark harnesses (one MedAgentBench harness bug fix counted as a round), data-synthesis skill set, training stack, human review step.
Feedback
Graded benchmark trajectories and scores on MedAgentBench, HealthBench Professional/Hard, MedXpertQA and AgentClinic, categorized into failure modes (Sec 3.2).
Timing
Offline rounds per capability loop (1-5 rounds shown), then consolidation of the accumulated mixture into the released model (Sec 4).
Retained artifact
Validated data refinements accumulated into a consolidated mixture; kept checkpoints serve as the next round's start; released Cura 1T weights.
Update generator
A training agent (model identity not reported) using reasoning-correction (frontier-model explanations), knowledge-injection, behavior-calibration, retention-anchor and mixture-curation skills.
Selection and release
A human approves each plan and makes keep, revert, refine or deploy decisions from benchmark scores; reverted rounds are reported (Fig 1, Tables 2-5).
Evaluation separate from updating
None for healthcare: the selection benchmarks also give the reported scores (authors call them development measurements, Sec 4.6); AIME, GPQA-Diamond and tau2-Bench were not used for selection.
Fixed-procedure control
Base Kimi-K2.6 and frontier comparators only; authors state isolating contributions or further ablations was beyond capability (Limitations).
Evidence concerning improver change
Not tested.
Clinical setting
Patient-care answers, expert medical QA with images, simulated interactive diagnosis (AgentClinic) and FHIR EHR tool use (MedAgentBench).
Evidence boundary
Benchmarks guide keep/revert and report results; human-in-the-loop decisions; no ablations; not evidence of clinical safety (Sec 6).
Online during RL (750 steps, about one epoch); memory update after rollout rounds with at least 4 unique failed cases.
Retained artifact
Accepted memory saved with each policy checkpoint; released MIRA-VL-8B weights, trajectories and tool environment.
Update generator
A reflection optimizer and editor (model not reported) propose bounded patch edits; an MCTS teacher engine (Seed 1.8, Gemini 2.5 Pro, GPT-5.2) built the SFT data.
Selection and release
Candidate memory accepted only if mean validation reward under the frozen updated policy improves; rejected edits stored to steer later proposals (Appendix E).
Evaluation separate from updating
Test splits of nine medical VQA benchmarks reported as never used for training or selection; whether the final memory is injected at test time is not stated explicitly.
Fixed-procedure control
RL ablation with format+accuracy, plus consistency, plus reflection memory (Table 4); validation curve with and without memory (Fig A7).
Evidence concerning improver change
None; the gate and editor are fixed, only the edit budget follows a preset schedule.
Clinical setting
Medical image question answering across radiology, pathology, dermatology, lesion and expert multimodal benchmarks.
Evidence boundary
Benchmark VQA rather than clinical workflow; single runs; memory's role inside the reward not isolated; authors note limited memory generalization (Sec 5).
Best retained checkpoint, plus per-round records of states, mixtures, reward version and controller scores.
Update generator
Stage Bank Data Factory teacher models synthesize leakage-checked fictional medical-imaging tasks and executable sandboxes (20,299 states); the router picks the weak stage.
Selection and release
Best-so-far checkpoint retained; fallback after three consecutive drops or KL shift above a threshold; best checkpoints auto-selected (Sec 2.2.2).
Evaluation separate from updating
AutoMedBench-Lite scores both steer training and report results (authors: not an untouched final test, Sec 5.1-5.2); ABRA, MedXpertQA-Text and eight external benchmarks were not used for selection.
Fixed-procedure control
Same SFT start trained by plain GRPO with a single final reward, and Stage Bank pool ablations (Table 2, Fig 5).
Evidence concerning improver change
None.
Clinical setting
Long-horizon medical-imaging AI development workflows (plan, setup, validate, inference, submit) and radiology viewer tasks (ABRA).
Evidence boundary
Benchmark participates in adaptation; authors say differences describe observed scores rather than a matched ablation (Sec 4.1); research workflows, not patient care.
Per-agent FAISS memories holding question context, answer, post-hoc notes and, after incorrect predictions, role-specific reflection entries; model weights are not updated (Sec III-C).
Fixed components
GPT-4o for every role, LangGraph routing (Moderator, Recruiter, specialists, Decision Maker, Ethical Overseer), text-embedding-3-large, top-5 retrieval, and a RAG corpus frozen before evaluation (Sec IV-B).
Feedback
Ground-truth multiple-choice label; reflection triggers only when the prediction differs from the label (Sec III-C, Alg. 1).
Timing
Alg. 1 writes feedback after each labelled question; Sec IV-B says test questions, predictions and reflections were never written back, so test-time memory was static; the phase that filled it is not reported.
Retained artifact
Role-specific memory entries and corrective reflection entries in FAISS, retrieved top-5 per agent (Sec III-C, IV-B).
Update generator
The same GPT-4o agents produce answers and reasoning summaries; a reflection step writes corrective feedback against the label (Sec III-C).
Selection and release
No filtering, pruning or validation of memory entries; the authors name pruning and confidence-based retention as future work (Sec V).
Evaluation separate from updating
MedQA official test split and a disjoint MedMCQA subset, evaluated in sequential batches of 50; test-set sizes, seeds and intervals are not reported (Sec IV-A, IV-B).
Fixed-procedure control
Ablation against a multi-agent baseline without memory, reflection or RAG, and feedback-only and memory-only variants (Table IV); no comparison of a revised versus a fixed update procedure.
Evidence concerning improver change
None; the reflection rule and memory format stay fixed.
Clinical setting
USMLE-style (MedQA) and AIIMS/NEET (MedMCQA) multiple-choice exam questions; no patient data.
Evidence boundary
Retrospective exam benchmarks only; Ethical Overseer not clinician-validated; source of test-time memory contents unreported; no uncertainty estimates (Sec IV-B, V).
Four typed external banks: Clinical Skills, Process Rules, Symbolic Schemas and Measurement Skills (visual procedures); backbone weights untouched (Sec 2.1, 2.2).
Fixed components
Hosted Qwen3.6-Flash or DeepSeek-V4-Flash Doctor, Process-Constrained Preference Harness (three candidates, symbolic verifier, Clinical Process Critic, risk audit), Patient and Measurement environment, evaluator, case order and turn ceiling (Sec 3.1).
Feedback
Post-encounter evaluation of each completed training trajectory against reference diagnosis and benchmark targets, then reflection (Sec 2.2, Table S1).
Timing
Offline: proposals after each of 700 training encounters (200 NEJM cases for the Measurement Bank) are merged into snapshots; the final snapshot is frozen and writes are disabled at test (Sec 2.2, 3.1).
Retained artifact
Versioned bank artifacts with content, applicability scope, trajectory provenance and lifecycle status, published as immutable hashed snapshots (Sec 2.2, S1, Table S3).
Update generator
An LLM reflector proposes typed add, merge, patch, deprecate or discard operations after each trajectory (Sec 2.2, S2.1, S2.2).
Selection and release
Typed merge operator and validator check type consistency, provenance, label leakage, controller-defined safety and replay on the proposing trajectories; failing artifacts are withheld or rejected (Sec 2.2).
Evaluation separate from updating
300 held-out MIMIC-IV FullChain encounters, 180 stress conditions from 30 disjoint cases, and 100 NEJM image cases; one rollout per case at temperature 0; LLM judge is the comparison backbone (Sec 3.1-3.3).
Fixed-procedure control
No-memory and single-bank-removal variants on a 100-case subset (Table 7), and six baselines with method-native frozen memory trained on the same 700 encounters (Table 4, Table S3); no revised versus fixed update-procedure test.
Evidence concerning improver change
None; reflector, merge and validation operators stay fixed.
Clinical setting
Simulated doctor-patient-measurement encounters from MIMIC-IV records (history to prescription) and NEJM image-challenge cases under the AgentClinic protocol.
Evidence boundary
Automatic LLM-judged labels without clinician calibration; single run per configuration; the authors say results do not establish clinical safety or causal credit for components (Sec 3.3, 4, 5).
A Counterfactual Diagnostic Evidence Graph of diagnosis, evidence and acquisition-action nodes plus diagnosis-revision edges (Graph Consolidation).
Fixed components
Doctor-agent backbones (Gemini-3-Flash-Preview, Gemini-3.1-Pro, GPT-5.6-Luna, Claude-Sonnet-5), patient environment and tools, 30-turn budget, Qwen3-Embedding and SapBERT matching thresholds (Implementation Details).
Feedback
Correctness of each trajectory's ICD-aligned final diagnosis against the reference, then verifier preference shifts under counterfactual evidence edits (Candidate Evidence Mining; Counterfactual Validation).
Timing
One offline construction pass: K=4 rollouts per construction case from two Gemini agents; the graph is frozen and shared across all evaluated backbones (Implementation Details).
Retained artifact
Frozen graph with 1,156 nodes and 1,426 relations, including 270 diagnosis-revision edges (Table 6).
Update generator
Failed-successful trajectory pairing (cosine > 0.7), UMLS evidence-atom mapping and counterfactual verification by GPT-5.5 (Implementation Details).
Selection and release
A candidate is kept only if the preference shift holds in both diagnosis orders; 692 of 2,271 candidates (30.47%) retained (Counterfactual Validation; Table 7).
Evaluation separate from updating
200 patient-disjoint MIMIC-IDx cases (ID) and 107 AgentClinic-MedQA cases (OOD); accuracy judged by GPT-5.6-Sol; repeated runs or intervals not reported (Benchmarks; Metrics; App A).
Fixed-procedure control
Vanilla agent, self-reflection, Trajectory RAG and Flat Experience baselines from the same construction split, plus a graph without counterfactual validation (Tables 2, 3); no revised versus fixed construction procedure over rounds.
Evidence concerning improver change
None; pairing, mapping and verifier stay fixed; online graph evolution is named as future work (Conclusion).
Clinical setting
Long-horizon interactive diagnosis over MIMIC-IV admissions with notes and matched MIMIC-CXR reports, plus AgentClinic MedQA scenarios.
Evidence boundary
Retrospective records in an interactive environment; LLM-judged accuracy; single construction pass; Claude-Sonnet-5 ID accuracy is below Trajectory RAG (Table 2).
Same pipeline with and without MistakeKDB on identical notes, plus raw-example and reflection-summary memory formats from the same error pool (Tables 1-3); routing policy not ablated; no revised update-procedure test.
Evidence concerning improver change
None; MistakeAnalyzer and routing rule stay fixed.
Clinical setting
ICD-9/CPT coding of MIMIC-III discharge summaries and ICD-10/HCPCS coding of MIMIC-IV notes.
Evidence boundary
Retrospective single-system study; one model family; administrative labels as proxy; taxonomy not coder-validated; CPT/HCPCS performance remains low (Sec 6).
Full weights of one shared Qwen3-VL-8B-Instruct policy acting as Radiologist, Anatomist and Consultant, including the visual pathway (Sec 5).
Fixed components
Tool harness and replayed evidence (Vista3D masks, MedGemma-27B reports), role prompts and tool budgets, Kimi K2.6 rubric judge, a frozen Consultant skill store built from teacher trajectories (Sec 5.1, 5.3, App B).
Feedback
Exact-match answer reward, format reward, Kimi K2.6 rubric score with top-2 linkage, and a Causal-KL penalty from an answer-privileged view of the same policy snapshot (Sec 5.4, 5.5, Eq. 7).
Timing
Offline: RSFT on screened teacher trajectories, then one epoch of on-policy GRPO with G=8 rollouts per prompt; the trained policy is evaluated frozen (Sec 5.2, 5.3, Fig. 3).
Retained artifact
Post-trained CASE-8B policy weights.
Update generator
The learner's own on-policy rollouts; the refreshed policy snapshot also serves as privileged self-teacher for the OPSD penalty (Sec 5.3, 5.4).
Selection and release
Last checkpoint of the RL run; RSFT data screened by tool use and a rubric score of at least 0.4 (Eq. 3); no checkpoint-gating procedure described.
Component ablation from shared RSFT initialization: RSFT only, GRPO, GRPO+RLAIF, GRPO+OPSD, full (Table 2); the evolving self-teacher is not compared with a frozen teacher.
Evidence concerning improver change
The privileged self-teacher is the evolving policy itself and is refreshed after updates (Sec 5.4); rubric and reward weights stay fixed.
Clinical setting
Longitudinal multi-sequence MRI from UK Biobank baseline and follow-up visits with ICD-10-derived multiple-choice diagnostic questions.
Evidence boundary
Single cohort; labels may be unobservable from images; generated evidence and LLM judges are fallible; utility beyond this cohort untested (Sec 7).
External doctor-agent artifacts (skills, SkillBank, utility-scored memory, experience bases, cheatsheet, playbook, tool registry) with Qwen3-32B frozen; Appendix C also updates Qwen3-8B weights by SkillRL-GRPO for 30 steps.
Fixed components
Environment roles (dual-layer patient simulator, laboratory physician with ClinLab-2B, environment controller), rule-based judge, ClinPatient cases, and the Qwen3-32B backbone for the nine Table 2 methods (Sec 2.2, App B).
Feedback
Deterministic trajectory-level scores against EHR ground truth: weighted information collection rate, test and abnormal-test coverage, ICD-10 block and three-character accuracy (Sec 2.4).
Timing
L1 offline adaptation from 5,000 trajectories; L2 online and L3 methods over 1,000 optimization trajectories (Sec 4.3); whether online updates ran on the 100 test cases is not reported.
Each borrowed method's own LLM distiller, reflector, curator or improver working on doctor-agent trajectories; Table 3 also tests MemRL with Qwen3-32B versus GPT-5.6-Luna trajectories.
Selection and release
Method-specific: MemRL utility re-ranking, ACE curator deduplication, MCE iterative evaluation selection, MedRSI fast-discovery slow-registration over later cohorts; ClinVerse adds no separate release gate (App B).
Evaluation separate from updating
100 cases randomly sampled from a held-out test set, 5 independent runs, mean and SD (Sec 4.1); ClinVerse-adapted agents tested on AIHospital without further adaptation (Table 4).
Fixed-procedure control
Unadapted Qwen3-32B (L0) under the same environment and metrics, plus fixed-procedure L1/L2 methods beside procedure-revising L3 methods (Table 2); no seed-improver versus improved-improver comparison.
Evidence concerning improver change
STOP applies its improver to its own code and MCE evolves its context-engineering skill (App B), but their effect on later learning is reported only as end performance, not isolated (Table 2).
Clinical setting
Simulated diagnostic consultation built from de-identified hospital and ICU EHRs (MIMIC-IV v3.1, MIMIC-IV-Note, SICdb), 150k cases, ICD-10 primary diagnosis as hidden label (Sec 3.1).
Evidence boundary
Simulated environment scored by the authors' own rules; methods reimplemented (SkillRL parameter-free); many gains within one SD over 5 runs; no clinician review of trajectories; preprint.
Norm trajectories and fixed AI preference feedback.
Timing
Cross-episode prompt edits and periodic identity updates.
Retained artifact
Structured norm memory and revised role prompts.
Update generator
Fixed diagnosis/reflection and identity-update routines.
Selection and release
Constraint checks described; no clinical deployment gate.
Evaluation separate from updating
Three synthetic healthcare communication tasks.
Fixed-procedure control
Component ablations; no isolated changing-improver control.
Evidence concerning improver change
Task role policies change; evaluator/update routine stays fixed.
Clinical setting
Simulated medication review, emergency triage and discharge handoff.
Evidence boundary
Judge count differs between Appendix A.1 and A.3 (5 versus 1); identity consistency is described as embedding distance and LLM scoring. No prospective patient evaluation.
Prompts, BM25 knowledge base, synthetic-data generator and GRPO routine.
Feedback
JSON-format and reference risk-label rewards after full agent trajectories.
Timing
Offline GRPO through Agent Lightning before separate consultation evaluation.
Retained artifact
Updated weights reused across different consultations.
Update generator
Fixed DeepSeek-R1 generates risk/no-risk training samples.
Selection and release
Rule-based terminal rewards; no prospective clinical release gate.
Evaluation separate from updating
Separate real/synthetic conversations and retrospective expert assessment.
Fixed-procedure control
Component ablations; no evolving-versus-fixed improver comparison.
Evidence concerning improver change
Policy changes; training-data generator and learning procedure stay fixed.
Clinical setting
Retrospective Chinese telemedicine risk checking.
Evidence boundary
Global memory is patient-specific and planner refinement is within-claim; eligibility rests on trajectory-based policy learning. Expert ratings are not prospective patient outcomes.
Persistent clinical guidance and final adapter gating parameter.
Fixed components
RetFound/BioClinicalBERT towers, adapter MLPs and identity specification.
Feedback
Original reports simulate physician adjudication of conflicting cases.
Timing
Updates every 100 iterations across repeated test-set passes.
Retained artifact
Special-case MEMORY, distilled SOUL guidance and final gate.
Update generator
Periodic fixed case summarisation and selective gate optimisation.
Selection and release
Conflict cases trigger feedback; no independent clinical release gate.
Evaluation separate from updating
Private fundus-image data; repeated adaptation on test cases.
Fixed-procedure control
Adapter-capacity comparison; no matched evolving-improver control.
Evidence concerning improver change
Task guidance and gate change; update procedure remains fixed.
Clinical setting
Retrospective fundus diagnosis/report generation.
Evidence boundary
Agent LLM backbone unspecified. Evolution evaluated by conflict counts, with report-simulated feedback and reused test data, not independent diagnostic improvement or prospective care.
Backbone during rule evolution, clustering and selection routine.
Feedback
Automatic impression scores and difficult training reports.
Timing
Offline iterative evolution before different test reports.
Retained artifact
Selected experience rules and historical reports.
Update generator
LLM induces/refines rules using retained rules and difficult examples.
Selection and release
Discard lower-scoring half; stop at predefined count/performance thresholds.
Evaluation separate from updating
MIMIC-CXR and Open-I; separate training/test comparisons.
Fixed-procedure control
Experience/component ablations; no isolated evolving rule-generator control.
Evidence concerning improver change
Experience rules provide context for their own refinement; the induction/selection procedure stays fixed, as in the existing Level 1 notebook rule.
Clinical setting
Retrospective radiology report summarisation.
Evidence boundary
Text findings-to-impression task; no image interpretation or prospective care. Contextual refinement of task rules does not establish a changed improver.
MedQA/PubMedQA test sets; experience banks contain training-set cases only.
Fixed-procedure control
Memory/discussion variants; no changing-improver control.
Evidence concerning improver change
Memory grows; abstraction and retrieval procedures stay fixed.
Clinical setting
Offline clinical and biomedical QA.
Evidence boundary
No prospective care. Comparisons mix MedQA option counts; role names do not establish image use. Distinct earlier ICASSP paper, not the retitled ACL MDTeamGPT record.
Internal partition of the public benchmark, not its official hidden test; search-mini feedback differs across strategies (256 versus128). No prospective clinical evaluation.
Experience vector database containing clinician corrections and prompts.
Fixed components
Qwen-7B, five prompted clinical modules, retrieval/tool interfaces and update routine.
Feedback
Doctors add prompts or edit LLM inputs/outputs; all edits and prompts are re-encoded into the vector database.
Timing
Feedback retained after human interventions for later similar clinical situations.
Retained artifact
Indexed clinician corrections, prompts and case-derived experience.
Update generator
Fixed recording, encoding and retrieval procedure.
Selection and release
Clinician intervention; no independent update-release threshold reported.
Evaluation separate from updating
461 retrospectively collected elderly outpatient prescriptions/clinical records from one hospital; two pharmacist raters and a third for disagreement. No separate adaptation/test split reported.
Fixed-procedure control
With/without HITL comparison; no isolated persistent-memory or changing-improver control.
Evidence concerning improver change
Task experience changes; update procedure and model are not shown to change.
Clinical setting
Polypharmacy review, interaction checks, medication schedules and patient education.
Evidence boundary
Abstract calls the mechanism RLHF, but methods document clinician-edit vector-memory reuse rather than parameter RL. HITL also includes current-case corrections, so its comparison does not isolate cross-case learning. Clinical data are retrospective, and deployment effectiveness is unvalidated.
Development and validation of a human-supervised AI-augmented living oncology evidence platform: a breast cancer pilot study
Authors
H. Rugo; Anna Forsythe; D. Flora; S. Glück; S. Grieve; R. Campden; R. J. Liu; M. Musat; J. M. Rege; P. A. Kaufman; R. L. Coleman; P. Tarantino; J. O’Shaughnessy; L. Schwartzberg
Year
2026
Publication status
J/C (ESMO Real World Data and Digital Oncology 13:100770)
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.
AgentEHR-Bench/MIMIC-IV and private CN-EHR with four backbones; all tasks are processed sequentially with memory from earlier tasks, without separate held-out memory learning.
Fixed-procedure control
Ablations remove UCB, step scoring and negative memory; no changing-improver control.
Evidence concerning improver change
Task memory and retrieval estimates change under fixed scoring/extraction/UCB rules.
Clinical setting
Six EHR prediction tasks and Chinese diagnosis coding.
Evidence boundary
Same-stream sequential adaptation, not an independent post-learning holdout. CN-EHR is externally sourced but also evaluated sequentially. No prospective patient outcomes; memory-on/off and component ablations do not isolate an evolving updater.
Shared agent prompts, response parameters and retrieval data; model refinement is described.
Fixed components
Backbone and optimizer Not reported.
Feedback
Human FAST ratings and SME reviews.
Timing
Continuous monitoring; revision cadence Not reported.
Retained artifact
Adjusted prompts/RAG/settings; fine-tuned model described.
Update generator
Human reviewers, SMEs and ML team.
Selection and release
Unacceptable ratings trigger refinement; release gate Not reported.
Evaluation separate from updating
Five reviewers initially assessed 12 recorded dialogues; the article reports framework development/face validity and ongoing service monitoring, without an independent post-update learning evaluation.
Fixed-procedure control
Not reported.
Evidence concerning improver change
Human rating rubric was revised; no effect of those revisions on subsequent agent learning was tested.
Clinical setting
Retrospective FAST assessment of recorded health-coaching dialogues from an operating patient service; deployment status is documented separately from the evaluation design.
Evidence boundary
The service has active patient users, but this article evaluates a human review framework on recorded dialogues. It does not report a prospective trial of agent adaptation or quantify post-update patient/learning benefits. Human-led model/rubric refinements are distinguished from autonomous self-improvement.
Case filtering described; checkpoint acceptance/release criterion absent.
Evaluation separate from updating
2000 heldout cases separate from RL, per author response; patient-level disjointness and external validation unspecified.
Fixed-procedure control
Table3 matchedSFT, outcome-onlyRL and full-reward ablations; same initialization/data/steps/hyperparameters stated.
Evidence concerning improver change
The process reward uses policy-model diagnostic likelihood, assigning the trained component a scoring role in later policy updates. Whether scoring parameters are synchronized after every update is not specified; the reward formula and GRPO procedure stay fixed.
Clinical setting
MIMIC-grounded virtual consultation.
Evidence boundary
Simulated clinical interactions; no prospective deployment. Author response states 2000 held-out cases differ from RL cases, but patient-level separation is unspecified. Policy likelihood is specified as process reward, not ordinary action-logprob loss; scoring-parameter synchronization and a frozen-scoring-model control are absent.
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.
Textual and multimodal multidisciplinary diagnostic QA.
Evidence boundary
Benchmark evidence only. Context-length stability and Playbook ablations do not establish an improved updater. SeM-Agents and MDTeamGPT are separately cited predecessors, not publication aliases. Imprint2027 follows first online2026 within cutoff.
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.
Selection and release
Change record requires frozen trigger evidence, root-cause account, proposed rule, adjudication, successor version, rollback target, regression result and independent-evaluation boundary.
Evaluation separate from updating
300 virtual-EHR tasks in11 first-pass stages; controller versions change across stages. Only named same-stage retests support version contrasts; no separate held-out learning-transfer test.
Fixed-procedure control
Named predecessor/successor controller runs, without a matched frozen-improver control; aggregate across versions is explicitly descriptive.
Evidence concerning improver change
The external controller changes task action admission. No revision and later-learning test of the change-proposal/adjudication procedure itself is demonstrated.
Clinical setting
Virtual-EHR task execution with FHIR-oriented tools, including medication and service writes.
Evidence boundary
Single-creator preprint; same-stage retests, missing scorer outcomes and candidate governance labels are disclosed. Version changes across different stages do not isolate causal gains; no clinical care was delivered.