Cura 1T: Healthcare Foundation Model via Recursive Self-Improvement

Haolin Chen; Leon Qi; Steve Brown; Deon Metelski; Tao Xia; Joonyul Lee; Qixuan Wang; Kevin Riley

cs.AIJul 15, 2026

Cura 1T: Healthcare Foundation Model via Recursive Self-Improvement

Authors: Haolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, +2 more

Organizations: actAVA AI Team

Abstract

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 $τ^2$ -Bench.

Figures & tables

Benchmark	Base	Cura 1T	$\Delta$
MedAgentBench	0.847	0.940	$+0.093$
HealthBench Professional	0.503	0.662	$+0.159$
HealthBench Hard	0.222	0.368	$+0.146$
MedXpertQA	0.569	0.655	$+0.086$
AgentClinic	0.754	0.796	$+0.042$

Table 1: Improvement of Cura 1T over its base model. Metrics differ by benchmark (rubric score for HealthBench, exact-letter pass@1 for MedXpertQA, task success for AgentClinic and MedAgentBench).

Figure 1: Left: recursive self-improvement loop for Cura 1T. A reviewer approves the plan before training and makes the keep, revert, or deploy decision after evaluation. Right: data refinement pipeline and training stack.

Figure 2: Cura 1T against frontier models and the Kimi-K2.6 base across six healthcare benchmarks: MedAgentBench ( Jiang et al., 2025 ) , HealthBench Professional and Hard ( Arora et al., 2025 ; OpenAI, 2026 ) , MedXpertQA text and multimodal ( Zuo et al., 2025 ) , and AgentClinic ( Schmidgall et al., 2024 ) .

Model	Intervention	Overall	Decision
Kimi-K2.6	$-$	0.847	$-$
Round 1	Behavior calibration	0.943	keep
Round 2	Behavior calibration + retention	0.967	keep
Round 3	Harness bug fix	0.973	keep
Cura 1T	Consolidated data mixture	0.940	release
Claude Opus 4.8	$-$	0.937	$-$

Table 2: MedAgentBench task success. Bold marks the best score and underline the second-best distinct score.

Model	Intervention	Professional	Hard	Decision
Kimi-K2.6	$-$	0.503	0.222	$-$
Round 1	Behavior calibration	0.601	0.332	revert
Round 2	Behavior calibration, cleaned	0.634	0.372	keep
Cura 1T	Consolidated data mixture	0.662	0.368	release

Table 3: HealthBench rubric scores at

T{=}1.0

. Bold marks the best score and underline the second-best distinct score in each column.

Model	Intervention	Text	Multimodal	Overall	Decision
Kimi-K2.6	$-$	0.484	0.672	0.569	$-$
Round 1	Reasoning correction	0.447	0.656	0.541	revert
Round 2	Reasoning correction, extended	0.454	0.657	0.545	revert
Round 3	Knowledge injection + retention	0.521	0.703	0.603	keep
Round 4	Data mixture curation	0.560	0.728	0.636	keep
Round 5	Data mixture curation, capped	$-$	$-$	0.440	revert

Table 4: MedXpertQA split and overall pass@1 at

T{=}1.0

. The overall score weights 2,450 text questions and 2,000 multimodal questions. Bold marks the best score and underline the second-best distinct score in each column.

Model	Intervention	MedQA	MedQA Ext	NEJM	NEJM Ext	Overall	Decision
Kimi-K2.6	$-$	0.869	0.827	0.400	0.567	0.754	$-$
Round 2	Behavior calibration + retention	0.841	0.841	0.800	0.717	0.807	keep
Cura 1T	Consolidated data mixture	0.879	0.850	0.800	0.625	0.796	release
Claude Opus 4.8	$-$	0.841	0.874	0.800	0.608	0.794	$-$
GPT-5.5	$-$	0.832	0.808	0.467	0.358	0.684	$-$

Table 5: AgentClinic pass@1 by subset under the tool-native protocol. Bold marks the best score and underline the second-best distinct score in each column.

Figure 3: Improvement map from the base model to Cura 1T. Values are changes from benchmark-specific bases; solid and dashed red arrows mark retained and reverted interventions.

Figure 4: Out-of-domain evaluation results for Cura 1T.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Role in the loop	Output
Plan	Define target behavior, metrics, data recipe, and candidate hyperparameters.	Training specifications
Train	Run SFT to verify the mixture and hyperparameters, RL to improve the policy against reward signals, and SDFT to consolidate the final model.	Candidate model
Evaluate	Run benchmark harnesses and collect graded trajectories and failure summaries.	Failed trajectories and metrics
Refine	Categorize failures, synthesize targeted data, curate the next mixture, validate candidate rows, and suggest next steps.	Validated data refinement
Reasoning Correction	Use a corrected reasoning pattern to synthesize new cases in different clinical contexts.	Reasoning rows
Knowledge Injection	Use verified knowledge to synthesize new questions that require the missing concept.	Source-grounded knowledge rows

Appendix

Table 6: Components of the recursive self-improvement loop in Figure 1 . The table summarizes how evaluation records are converted into refined data mixtures and how SFT, RL, and SDFT divide the training step.

Configuration	Cura 1T
Algorithm	SFT $\rightarrow$ RL $\rightarrow$ SDFT
Base model	Kimi-K2.6
Learning rate	$3\times 10^{-4}$ , linear schedule
Batch size	128
Sequence limit	256K tokens
Training length	6 epochs with early stopping

Appendix

Table 7: Training hyperparameters for Cura 1T.

Stage	Response excerpt	Result
Base	fhir_post(ServiceRequest, priority= routine , code=306181000000106, display=...)	Reward 0
Round 1 (Behavior calibration)	fhir_post(ServiceRequest, priority=stat, code=306181000000106)	Reward 1
Round 2 (Behavior calibration + retention)	priority=stat ; exact SNOMED coding, subject, timestamp, and SBAR note are preserved.	Reward 1
Round 3 (Harness bug fix)	Under the corrected harness: priority= stat with exact coding and finish([]) .	Reward 1

Appendix

Table 12

Stage	Response excerpt	Result
Base	“I cannot definitively identify what drug ‘Erostin 10 mg’ is based on the brand name alone.”	Reward 0
Behavior calibration	“EROSTIN 10 mg is an antihistamine medication containing ebastine.” The task is correct, although the round is reverted for subset-level degradation.	Reward 1
Retention check	“EROSTIN 10 mg is an antihistamine medication that contains ebastine as its active component.”	Reward 1
Cura 1T	“EROSTIN 10 mg is an antihistamine medication that contains ebastine as its active ingredient.”	Reward 1

Appendix

Table 13

Stage	Response excerpt	Result
Base	“ M54.30 —sciatica, unspecified side; M54.31 —right; M54.32 —left.”	Reward 1.000
Behavior calibration	“ M54.3 —sciatica, unspecified side,” while retaining M54.31 and M54.32 .	Reward 0.000
Cura 1T	Restores M54.30 , M54.31 , and M54.32 , and distinguishes ICD-10-CM from the WHO category.	Reward 0.973

Appendix

Table 14

Stage	Response excerpt	Result
Base	“A positive surgical margin signifies likely residual disease and the highest baseline recurrence risk … $\boxed{{\color[rgb]{0.7539,0.2227,0.168}\mathrm{F}}}$ .”	Reward 0
Round 1 (Reasoning correction)	“Option F combines … grade 3 histology and a positive margin. $\boxed{\mathrm{F}}$ .”	Reward 0
Round 2 (extended)	“The highest-risk breast-conservation patient … $\boxed{\mathrm{F}}$ .”	Reward 0
Round 3 (Knowledge injection + retention)	“Positive-margin cases … are not rescued by RT alone … Option D … $\boxed{\mathrm{D}}$ .”	Reward 1
Round 4 (Data mixture curation)	“Positive-margin cases … are not rescued by RT alone … Option D … $\boxed{\mathrm{D}}$ .”	Reward 1
Cura 1T	“Positive margins … require further surgery … Option D … $\boxed{\mathrm{D}}$ .”	Reward 1

Appendix

Table 15

Stage	Response excerpt	Result
Base	“The most likely diagnosis is Chronic Eosinophilic Pneumonia .”	Reward 0
Reverted diagnostic attempt	“DIAGNOSIS READY: Pulmonary alveolar proteinosis,” despite normal anti-GM-CSF and nondiagnostic BAL/PAS results.	Reward 0
Behavior calibration + retention	“The diffuse ground-glass opacities … and macrophages in the airspaces … [are] classic for Desquamative Interstitial Pneumonia.”	Reward 1
Cura 1T	Requests biopsy review, receives “Extensive alveolar filling with pigment-laden macrophages,” and submits desquamative interstitial pneumonia .	Reward 1

Appendix

Table 16

Explore similar work

HealthAgentBench: A Unified Benchmark Suite of Realistic Agentic Healthcare Environments for Challenging Frontier AI Agents
Jun 30, 2026Qianchu Liu, Sheng Zhang, Guanghui Qin +16Agentic Benchmarks Memoryagentbench
Microsoft Research
Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards
Sep 21, 2026Kalash Shah, Kunal Singh, Snehan J +1Clinical Reasoning Training LLM Reasoning Strategies
Fractal AI Research
Healthcare AI GYM for Medical Agents
May 1, 2026Minbyul JeongClinical Reasoning Training Skillgym
Upstage AI

Jun 30, 2026, cs.AIJ/K move · Enter open · S save

HealthAgentBench: A Unified Benchmark Suite of Realistic Agentic Healthcare Environments for Challenging Frontier AI Agents

Qianchu Liu, Sheng Zhang, Guanghui Qin, Jeya Maria Jose Valanarasu+15

Microsoft Research

Agentic Benchmarks Memoryagentbench Clinical Reasoning Training Electronic Health Records Agentic Control

Open paper AskPDF arXiv

As AI agents become increasingly capable of complex, long-horizon reasoning, rigorous and holistic evaluation is essential for measuring progress toward real-world healthcare applications. We introduce HealthAgentBench, a suite of 54 agentic healthcare tasks across 7 categories each with its unique environment. The benchmark suite spans diverse workflows throughout the patient journey and a broad range of modalities. Each task is designed to replicate an end-to-end clinical workflow: given minimal instructions, an agent must explore raw healthcare data, operate within a complex environment, and execute multi-step solutions that go beyond naive prompting. A final task success rate is reported to provide a single, interpretable metric for HealthAgentBench overall performance for each agent. Evaluating frontier agents on HealthAgentBench, we find that overall task success rate remains low, underscoring the difficulty of the suite. The strongest and the most cost effective agent, Codex GPT-5.5, achieves only approximately 42% success rate. Beyond aggregate performance, HealthAgentBench reveals nuanced strengths and weaknesses across task categories. Frontier agents show promise in automatically developing research modeling pipelines over EHR data, but medical imaging remains especially challenging, particularly for Claude Code models, while Codex GPT-5.5 shows emerging capability. Tasks that combine large search spaces with compositional reasoning requirements remain difficult for all current agents. Together, these results suggest that HealthAgentBench provides a challenging and realistic benchmark with substantial room for future progress. We release our benchmark at https://github.com/microsoft/HealthAgentBench.

Abstract

Explore similar work

HealthAgentBench: A Unified Benchmark Suite of Realistic Agentic Healthcare Environments for Challenging Frontier AI Agents

Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards

Healthcare AI GYM for Medical Agents