Medical agents increasingly combine general reasoning models with specialized clinical tools, yet their capabilities remain largely fixed by what clinicians and engineers design before deployment. Recursive self-improvement (RSI) offers a different paradigm in which agents learn from their own failures and autonomously expand their capabilities, but directly applying RSI to medicine introduces fundamental safety challenges. We introduce MedRSI, the first recursive self-improvement framework for medicine, which continuously transforms diagnostic failures into new clinical capabilities through tool composition and task-specific model training. Inspired by clinical practice, MedRSI introduces two mechanisms for clinically aligned self-evolution. Clinical-cost-aware failure prioritization directs improvement toward errors according to their potential clinical consequences rather than frequency alone. Fast discovery with slow registration separates rapid capability invention from conservative adoption, allowing new tools to enter the persistent agent only after demonstrating sustained benefit across subsequent patient cohorts. Across public glaucoma and heart disease benchmarks and two private clinical tasks, MedRSI progressively develops segmentation, measurement, prediction, multimodal reasoning, and generative capabilities, surpasses manually engineered medical agents, and autonomously discovers solutions to clinical problems not anticipated by its original designers. Our results show that medical agents need not remain constrained by capabilities specified before deployment: with clinically grounded mechanisms governing what to improve and what to retain, they can continuously construct, validate, and accumulate new capabilities from diagnostic experience. Code is available at https://github.com/ImprintLab/MedRSI.
Figures & tables
Figure 1: Recursive self-improvement of a medical agent in MedRSI. a , The medical agent Hr combines a reasoning model, a diagnostic workflow and a stable tool registry that grows across rounds. In each round the agent diagnoses a patient batch, reviews its failures with clinical consequence weighting, invents candidate tools, places them in an experimental pool and evaluates them on later cohorts before registration. b , Candidate tools arise from code generation, tool composition and model development. Each candidate is frozen with a version hash and tracked across a discovery cohort and K trial cohorts. A candidate with sustained gain and stable clinical cost is registered, and a candidate whose early gain fades is discarded.
Glaucoma (REFUGE2)
Heart disease (MITEA)
Family
Method
bAcc
F1
AUC
bAcc
F1
AUC
General multimodal
GPT-4o OpenAI (2024)
56.4
21.1
—
56.8
28.1
—
Janus-Pro-7B Chen et al. (2025)
53.4
13.3
—
52.3
10.7
—
LLaVA-Med Li et al. (2023)
50.0
0.0
—
50.0
0.0
—
BioMedCLIP Zhang et al. (2023)
58.1
21.3
—
47.0
37.8
—
Qwen2.5-VL-7B Bai et al. (2025)
54.3
16.3
—
50.0
0.0
—
Table 1: Diagnostic performance on the public tasks. Balanced accuracy (bAcc), F1 and area under the receiver operating characteristic curve (AUC) are percentages. Comparator values are reported by MedAgent-Pro Wang et al. (2025b) , including its evaluation of the ophthalmology models and the REFUGE2 challenge leaderboard. MedRSI values are means over five independent improvement trajectories, and their confidence intervals are given in Table 2 . A dash indicates that the source provides no corresponding value.
Task
Round
bAcc (95% CI)
Sens.
Spec.
PPV
NPV
MCC
ECE
Cost
Severe
Glaucoma
0
56.2 (53.6–58.8)
28.4
84.0
18.7
90.0
0.10
0.29
12.4
9.8
5
80.9 (78.2–83.6)
71.6
90.2
48.7
96.1
0.53
0.17
5.9
4.4
10
91.2 (89.6–92.8)
87.3
95.1
69.8
98.3
0.75
0.09
3.0
2.1
20
94.1 (92.8–95.4)
92.0
96.2
75.9
98.9
0.81
0.05
2.5
1.5
Heart disease
0
56.5 (53.1–59.9)
22.8
90.2
56.7
67.5
0.18
0.31
14.1
10.6
5
64.1 (60.5–67.7)
49.2
79.0
56.9
73.4
0.29
0.24
10.2
7.5
Table 2: Complete diagnostic metric panel across improvement rounds. All values are percentages except the Matthews correlation coefficient (MCC) and the expected calibration error (ECE), which are reported on their native scales. Clinical cost is the mean consequence-weighted error per 100 test cases. Severe errors are errors graded serious or critical as a percentage of all test cases. Values are means over five independent trajectories, with 95% confidence intervals from patient-level bootstrap resampling given for balanced accuracy. Predictive values and F1 depend on the test-partition prevalence, which is 11.5% for glaucoma and 36% for heart disease.
Figure 2: A general agent grows into a medical diagnostic system. a , b , Test balanced accuracy across improvement rounds for glaucoma (REFUGE2) and heart disease (MITEA). Lines show the mean over five trajectories and shading shows 95% confidence intervals. Horizontal references reproduce MedAgent-Pro, the best general multimodal model and the task-specific VisionUnite model as reported by MedAgent-Pro Wang et al. (2025b) . Annotations mark the registration round of representative tools. c , Cumulative number of stable task-specific tools across both tasks by capability category. General execution and retrieval utilities present in the initial agent are excluded from the count.
Figure 3: MedRSI rediscovers expert-designed clinical capabilities. a , Timeline of the 16 stable tools acquired on the public tasks. Open circles mark entry into the experimental pool, filled squares mark registration into the stable agent and bar colour indicates the capability category. b , Coverage of the five expert-designed capability categories in MedAgent-Pro across improvement rounds. c , Correspondence between expert-designed categories and tools discovered by MedRSI at round 20. Cell values give the number of discovered tools that execute the corresponding expert function on held-out cases.
Tool
Task
Category
Route
Rd.
Resource
Quality
Gain
Fundus quality and disc crop
G
Image prep.
C
2
rule based
0.98
+0.9
Optic disc localization
G
Segmentation
M
3
600 images
0.94
+2.4
Optic cup and disc segment.
G
Segmentation
M
4
600 masks
0.87
+7.6
Vertical cup-to-disc ratio
G
Measurement
T
5
composition
0.91
+6.1
Rim width by ISNT sector
G
Measurement
T
7
composition
0.84
+2.2
Peripapillary atrophy grading
G
Prediction
M
9
600 images
0.81
+1.7
Table 3: Inventory of the stable tools acquired on the public tasks. Route indicates code generation (C), tool composition (T) or model development (M). Round gives the improvement round at which the tool moved from the experimental pool into the stable registry. Intrinsic quality is the Dice coefficient for segmentation, the Pearson correlation against the reference indicator for measurement, and discovery-cohort balanced accuracy for prediction and synthesis. Mean trial gain is the paired change in task balanced accuracy across the three trial cohorts. All values are means over five trajectories.
Family
Method
bAcc (95% CI)
F1
AUC
Sens.
Spec.
General
GPT-4o
68.4 (65.1–71.7)
60.9
72.6
61.3
75.5
Single modality
Fundus only
78.1 (75.2–81.0)
72.9
84.9
73.2
83.0
OCT only
80.6 (77.8–83.4)
75.9
87.2
76.9
84.3
Visual field only
75.3 (72.1–78.5)
69.4
81.7
70.1
80.5
Multimodal
Late fusion
83.9 (81.2–86.6)
80.0
90.4
80.6
87.2
MedAgent-Pro
79.2 (76.4–82.0)
74.2
85.8
74.5
83.9
Table 4: Multimodal glaucoma diagnosis. All values are percentages. Single-modality rows use the same reasoning agent restricted to one input. Late fusion combines the three single-modality probabilities with a fitted logistic model. The task-specific fusion model is trained by the study team on the same labelled training eyes with a development budget matched to the total budget consumed by MedRSI. Values are means over five runs with 95% confidence intervals from patient-level bootstrap resampling.
Method
nMAE (95% CI)
MAE
Bias
r
ICC
Within 5 pts
GPT-4o
1.00 (0.92–1.08)
9.3
−4.1
0.34
0.29
31.2
MedAgent-Pro
0.85 (0.78–0.92)
7.9
−3.2
0.52
0.48
42.7
Noncontrast predictor applied directly
0.92 (0.84–1.00)
8.6
−3.8
0.44
0.40
37.5
Contrast-only predictor
0.74 (0.67–0.81)
6.9
−2.1
0.66
0.63
51.0
Synthetic-only predictor
0.79 (0.72–0.86)
7.3
−2.6
0.61
0.58
47.9
Task-specific regressor
0.66 (0.60–0.72)
6.1
−1.7
0.74
0.71
58.3
Table 5: Ejection fraction prediction from contrast echocardiography. Normalized mean absolute error (nMAE) divides the absolute error by the training-set standard deviation of ejection fraction, and lower values indicate better prediction. MAE and bias are given in ejection fraction points against the cardiac magnetic resonance reference. The Pearson correlation and the intraclass correlation coefficient (ICC) assess agreement with the reference. Within 5 points is the percentage of test studies whose predicted ejection fraction falls inside the prespecified clinical tolerance. Values are means over five runs with 95% confidence intervals from patient-level bootstrap resampling.
Figure 4: Invention of new clinical models on previously unseen tasks. a , Test balanced accuracy for multimodal glaucoma diagnosis across improvement rounds, starting from MedAgent-Pro. b , Normalized mean absolute error of ejection fraction from contrast echocardiography across improvement rounds. Lines show means over five runs and shading shows 95% confidence intervals. Horizontal references show MedAgent-Pro, the general multimodal model and the task-specific model trained on the same data. Annotations mark the registration of the tools that produce each change. c , Development record of the multimodal glaucoma network. d , Development record of the generative pipeline for contrast echocardiography, in which a translation model expands the contrast training set and a predictor is retrained on real and synthetic studies.
Configuration
bAcc
F1
AUC
Cost
Severe
Tools
Discard
Round 30
Generic RSI
77.4
61.8
83.1
9.2
6.4
45
0
72.7
Uniform failure priority
88.7
74.6
92.8
6.0
4.0
15
61
88.3
Immediate registration
82.4
67.2
87.4
7.6
5.1
42
0
76.9
Single trial cohort
90.6
77.9
94.1
4.7
3.2
31
24
84.7
Without failure clustering
90.2
77.1
93.6
3.9
2.6
14
66
90.1
Code generation only
79.8
63.4
85.2
7.1
5.3
9
58
79.4
Table 6: Ablation of the improvement mechanisms on the glaucoma trajectory. Test balanced accuracy, F1 and AUC are measured at round 20, and long-horizon balanced accuracy at round 30. Clinical cost is the mean consequence-weighted error per 100 test cases. Severe errors are errors graded serious or critical as a percentage of all test cases. Stable tools are counted at round 20. Discard rate is the percentage of experimental tools removed before registration. All configurations share the initial agent, patient partitions, candidate development budget and random seeds. Values are means over five runs.
Figure 5: Clinical-cost prioritization and slow registration are required for medical recursive self-improvement. a , Training and test balanced accuracy for the complete agent and the uniform priority ablation on the glaucoma trajectory. b , Mean clinical cost per 100 test cases. c , Residual test errors at round 20 by consequence grade. d , Test balanced accuracy over 30 rounds for the complete agent, immediate registration and generic recursive self-improvement, which combines uniform priority with immediate registration. e , Number of stable tools over the same horizon. f , Paired balanced accuracy change of one candidate tool across its discovery cohort and three trial cohorts. The dotted line marks the registration margin δ .
Figure 6: Sensitivity of MedRSI to its settings, its reasoning model and the patient population. a , Balanced accuracy at round 20 across values of the clinical weight λ , with the mean clinical cost printed above each point. b , Balanced accuracy at round 30 across the number of trial cohorts K , with the number of registered tools printed above each point. c , Glaucoma balanced accuracy at rounds 0 and 20 for four reasoning models, with the MedAgent-Pro reference. d , Sensitivity by glaucoma stage for MedAgent-Pro, the uniform priority ablation and the complete agent. e , Agent balanced accuracy against the optic cup Dice coefficient of successive versions of the self-trained segmentation model. f , Mean trial gain of every candidate that reached the experimental pool on the public tasks, separated by registration outcome, with horizontal lines at the medians.
Task
Rd.
Cand.
Pool
Reg.
GPU h
Tok. (B)
Days
h/tool
Glaucoma (REFUGE2)
20
64
22
8
96
0.9
2.1
12.0
Heart disease (MITEA)
20
54
19
8
116
1.0
2.4
14.5
Multimodal glaucoma
12
41
14
6
174
0.7
3.2
29.0
Contrast echo
12
38
13
5
265
0.6
4.6
53.0
Total
—
197
68
27
651
3.2
12.3
24.1
Table 7: Discovery activity and computational budget. Candidates counts the development attempts the builder completed, pool counts those that passed the discovery threshold, and registered counts those that survived their trials. Accelerator hours are measured on NVIDIA A100 80 GB devices and cover candidate model training and evaluation. Reasoning tokens cover diagnosis, reflection, tool building and judging. Values are means over five runs for the public tasks and over five runs for each additional task.
Tool-augmented large language model (LLM) agents can orchestrate specialist classifiers, segmentation models, and visual question-answering modules to interpret chest X-rays. However, these agents still solve each case in isolation: they fail to accumulate experience across cases, correct recurrent reasoning mistakes, or adapt their tool-use behavior without expensive reinforcement learning. While a radiologist naturally improves with every case, current agents remain static. In this work, we propose Evo-MedAgent, a self-evolving memory module that equips a medical agent with the capacity for inter-case learning at test time. Our memory comprises three complementary stores: (1)~Retrospective Clinical Episodes that retrieve problem-solving experiences from similar past cases, (2)~an Adaptive Procedural Heuristics bank curating priority-tagged diagnostic rules that evolves via reflection, much like a physician refining their internal criteria, and (3)~a Tool Reliability Controller that tracks per-tool trustworthiness. On ChestAgentBench, Evo-MedAgent raises multiple-choice question (MCQ) accuracy from 0.68 to 0.79 on GPT-5-mini, and from 0.76 to 0.87 on Gemini-3 Flash. With a strong base model, evolving memory improves performance more effectively than orchestrating external tools on qualitative diagnostic tasks. Because Evo-MedAgent requires no training, its per-case overhead is bounded by one additional retrieval pass and a single reflection call, making it deployable on top of any frozen model.
Weixiang Shen, Bailiang Jian, Jun Li +6
Technical University of Munich (TUM) · TUM University Hospital · LMU Munich +3
Healthcare spans high-stakes communication, expert reasoning, and workflow execution, yet specialized language models that cover these use cases together remain limited. A healthcare model must handle patient consultation, clinical reasoning over text and images, interactive diagnosis, and electronic health record (EHR) tool use. These capabilities fail in different ways, and a narrow update for one task can degrade another. We present Cura 1T, a healthcare foundation model trained through recursive self-improvement (RSI). In each RSI round, the RSI harness runs the current model on healthcare benchmarks, evaluates the trajectories to locate capability gaps, and refines the training mixture by synthesizing training data. On 6 healthcare benchmarks, Cura 1T scores highest on MedAgentBench, HealthBench Professional, HealthBench Hard, MedXpertQA text, and AgentClinic, and second on MedXpertQA multimodal. It preserves performances on out-of-domain reasoning and agentic benchmarks including AIME, GPQA-Diamond, and τ2-Bench.
Recursive self-improvement (RSI) lets a system improve the model-building machinery from its own failures, so every later model inherits the gain. Yet RSI has been validated almost exclusively on coding and formal benchmarks such as science QA and mathematics. This format bound limits RSI to improvement within a machine-checkable slice, not general capability where questions are open and correctness is settled by argument, replication, or measurement. We argue RSI must next operate across real, diverse scientific, engineering, and meta-scientific domains, not where formal evaluation is merely tractable. To that end we present MetaRSI-v1, where improvement is the scheduled composition of three typed operators over one unified paradigm. Data-RSI amplifies existing competence and marks its boundary; Harness-RSI edits a five-slot scaffold without touching weights; Model-RSI internalizes capability into parameters through bounded training. Sharing one loop kernel and artifact vocabulary, they make data, scaffold, and model changes composable rather than exclusive. A two-axis optimizer jointly decides operator order and each operator's proposal policy, while a meta-level policy revises the schedule across terms. We validate MetaRSI-v1 under the field's standard evaluations, on code and closed-form science, with no external teacher: the target model plays every role in its own loop. MetaRSI-v1 reframes self-improvement from a single-surface edit to a composition across the full model-production pipeline, opening two paths: a model route internalizing capability through training, and a harness route leaving weights untouched and thus extending self-improvement to any model reachable through an interface, with Data-RSI redefined as the shared substrate feeding both. The framework further yields refutable laws on where loops exist, how operators compose, and what supervision buys.