Cura 1T: Specialized Model for Agentic Healthcare

actAVA AI, Haolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, Frank Wang, Weiran Yao

Cura 1T uses a human-gated self-evolution loop to curate training data for specialized healthcare agentic tasks.

How can we specialize a general-purpose LLM for healthcare by treating data mixture and iterative self-evolution as the primary drivers of performance?

Healthcare models often fail across the diverse requirements of patient care, clinical reasoning, and electronic health record (EHR) tool use. A narrow update for one task frequently degrades performance on others, making it difficult to build a single, reliable clinical assistant. Cura 1T addresses this by treating data mixture construction as the primary object of search. A training agent plans experiments, evaluates model trajectories, and synthesizes targeted data to repair specific failures while using retention anchors to preserve existing capabilities. Across five of six healthcare benchmark panels, Cura 1T ranks at or near the top, consistently outperforming its base model while maintaining general reasoning performance.

Paper Primer

The core mechanism is a closed-loop self-evolution process where an LLM agent iteratively refines the training data mixture. The agent categorizes benchmark failures, synthesizes targeted correction rows, and merges them with retention anchors to prevent catastrophic forgetting of previously solved tasks.

Cura 1T achieves superior performance across clinical benchmarks compared to its base model.

The model improved task success on MedAgentBench from 0.883 to 0.940 and increased HealthBench Professional rubric scores from 0.503 to 0.662. The model ranks at or near the top on five of six healthcare benchmark panels.

Healthcare specialization does not erode general reasoning capabilities.

Cura 1T maintains performance on out-of-domain benchmarks including AIME, GPQA-Diamond, and $\tau$²-Bench. The model remains on par with frontier comparators on general reasoning and agentic tasks.

Why is a self-evolution loop necessary instead of just training on more medical data?

Healthcare training data is sparse and unevenly distributed across use cases. A generic update often causes "forgetting" of existing behaviors; the loop allows the model to target specific failure modes—like brittle EHR tool calls or omitted rubric items—without degrading other capabilities.

What is the role of the "retention anchor" in this training process?

Retention anchors are a set of examples for capabilities the model already handles correctly. By including these in every new data mixture, the training process ensures that targeted repairs for new tasks do not overwrite previously validated clinical knowledge or reasoning patterns.

Cura 1T demonstrates that for high-stakes domains like healthcare, the primary bottleneck is not model architecture or hyperparameter tuning, but the systematic curation of a balanced, failure-aware training data mixture.

Healthcare Specialization via Self-Evolution

We introduce Cura 1T, a self‑evolving healthcare LLM that tackles the data‑mixture problem.

Healthcare LLMs must simultaneously master patient‑facing dialogue, expert clinical reasoning, and complex workflow execution, yet progress on each sub‑task has historically been isolated.

Healthcare specialization means tailoring a language model to handle the distinct, high‑stakes tasks of patient interaction, clinical reasoning, and workflow automation within medicine.

**Figure 1** Performance of Cura 1T, frontier models, and the Kimi-K2.6 base across six healthcare benchmark panels: MedAgentBench (Jiang et al., 2025), HealthBench Professional and Hard (Arora et al., 2025; OpenAI, 2026), MedXpertQA (Zuo et al., 2025), and AgentClinic (Schmidgall et al., 2024).

Table 1 quantifies Cura 1T’s gains over the Kimi‑K2.6 base: the largest $\Delta$ is +0.159 on HealthBench Professional, and the smallest is +0.042 on AgentClinic, confirming that the loop benefits all three use‑case families.

Healthcare specialization is fundamentally a data‑mixture problem: success hinges on curating a balanced training set that addresses heterogeneous failures across clinical, conversational, and agentic tasks.

Prior Work in Healthcare AI

Key prior work on healthcare AI, LLM post‑training, and self‑evolution methods.

Healthcare language‑model research now spans three strands: patient‑care models (e.g., Med‑PaLM, Baichuan‑M4), clinical‑reasoning benchmarks (HealthBench, MedQA, MedXpertQA), and agentic evaluation suites that test multi‑turn, policy‑driven workflows such as AgentClinic and HealthAgentBench.

Post‑training pipelines typically start with supervised fine‑tuning (SFT), add reinforcement learning (RL), and may incorporate specialized fine‑tuning tricks like rejection‑sampling, STaR, or Self‑distillation fine‑tuning (SDFT) to keep updates close to the base model while preserving chain‑of‑thought data formats and using low‑rank adapters for efficiency.

Self‑evolution research began with Self‑Refine and Reflexion, progressed through prompt‑ and program‑optimizing systems (OPRO, DSPy, TextGrad), and matured in open‑source autoresearch loops that edit training code; Cura adopts this closed‑loop discipline but redirects the optimization target toward the post‑training data mixture.

The Cura 1T Architecture

How Cura 1T iteratively refines its data to master clinical reasoning.

Cura 1T sits on top of the Kimi‑K2.6 foundation and adds a trillion‑parameter head that can read long multimodal inputs, enabling it to answer patient‑facing queries, solve complex diagnostic reasoning, and orchestrate multi‑turn clinical workflows.

How does Cura 1T differ from simply fine‑tuning a large language model on medical data?

Fine‑tuning changes the whole parameter set and quickly forgets previously mastered behavior. Cura 1T freezes the trillion‑parameter core and only trains lightweight adapters, then runs an agent‑managed self‑evolution loop that explicitly preserves retained capabilities while targeting failures.

The loop works like a chef who tastes a dish, notes the missing spice, adds a pinch, and re‑cooks—each iteration diagnoses a failure, synthesizes a targeted data “spice”, and retrains the model on the enriched mixture.

Plan: Agent flags the failure and proposes a “Reasoning Correction” skill.

Refine: The frontier model rewrites the chain‑of‑thought, inserting a missing differential diagnosis step.

Data Mixture Curation: The corrected trace is added, the original failing example is kept as a negative, and a Retention Anchor example for “common cold” is inserted.

Train: SFT verifies the mixture is well‑formed, RL explores policy improvements, SDFT distills the new behavior while preserving prior knowledge.

Evaluate: Benchmarks show the pneumonia case now succeeds; the common‑cold behavior remains unchanged.

The loop can repair a specific reasoning gap without erasing previously learned medical facts, because the Retention Anchor explicitly protects them.

Instead of updating the entire trillion‑parameter backbone, a thin set of adapters is inserted; they are cheap to train, fast to iterate, and keep the core knowledge intact.

Why not replace the three‑stage stack with a single end‑to‑end RL run?

End‑to‑end RL on a trillion‑parameter model would be prohibitively expensive and would quickly erase previously learned medical knowledge. The three‑stage stack isolates cheap supervision (SFT), targeted policy improvement (RL), and anti‑forgetting distillation (SDFT), making the loop both compute‑efficient and safe.

**Figure 2** Left: Agent-managed self-evolution loop for Cura 1T. Human review gates the plan before training and the keep, revert, or deploy decision after evaluation. Right: Data refinement pipeline and training stack.

Table 2 enumerates the twelve concrete components of the loop, from the initial planning specifications to the final human‑review gate, making the workflow fully auditable.

Empirical Performance and Benchmarks

Evaluation shows Cura 1T outperforms baselines across clinical benchmarks.

Healthcare specialization is a data‑mixture problem; Cura 1T iteratively refines its training data via an agent‑managed self‑evolution loop to master clinical reasoning and tool use.

Cura 1T attains the highest Professional score on HealthBench, reaching 0.662.

HealthBench Professional benchmark shows Cura 1T scoring 0.662, surpassing the previous best of 0.634.

A benchmark that checks whether a model can correctly invoke FHIR‑based EHR tool calls to complete a clinical workflow.

Evaluates open‑ended clinical answers against physician‑authored rubrics, rewarding inclusion of required medical points.

Measures exact‑letter accuracy on a mix of text‑only and multimodal medical questions, reporting pass@1 for each modality and an overall weighted score.

**Figure 3.** Evolution 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.

Table 3 records MedAgentBench success across self‑evolution rounds; the consolidated Cura 1T model scores 0.940, slightly below the round‑3 frontier of 0.973 but well above the base 0.883.

Table 4 presents HealthBench rubric scores; Cura 1T achieves the top Professional score (0.662) while staying close to the best Hard score (0.368), confirming balanced improvement.

Table 5 summarizes MedXpertQA results; after knowledge injection and mixture refinement, Cura 1T reaches an overall pass@1 of 0.655, outperforming the base model and Claude Opus 4.8.

Table 6 details AgentClinic outcomes; interactive trajectory + retention raises overall pass@1 to 0.807, and Cura 1T preserves most of that gain with a final score of 0.800.

Tables 8–12 illustrate specific correction cases (referral intent, drug identification, coding, response progression, and multi‑turn diagnosis), each showing how the self‑evolution loop repairs failures while retaining prior capabilities.

Lessons in Healthcare Specialization

We discuss why data‑mixture, not hyperparameter search, drives healthcare specialization and outline the limits of our self‑evolution approach.

Our experiments confirm that healthcare specialization is fundamentally a data‑mixture problem rather than a hyperparameter‑search problem. The agent‑managed self‑evolution loop addresses this by turning evaluation failures into evidence of missing data, then applying SFT to screen candidate mixtures, RL to improve the policy, and SDFT to consolidate the final round. This approach is crucial when training data are sparse and uneven, such as needing more rubric‑following behavior for patient care or more factual coverage for expert questions.

Benchmark analyses show that a single generic medical‑data update cannot fix all failures. HealthBench errors are dominated by omitted rubric items and verbosity‑driven grading, MedXpertQA mixes reasoning‑pattern failures with missing knowledge, and MedAgentBench reveals brittle FHIR writes, while AgentClinic requires task‑specific interactive trajectories. These observations justify the design in Section 3.2 to categorize failures before synthesizing new data.

Cura 1T builds a healthcare‑specialized LLM by iterating a self‑evolution loop starting from Kimi‑K2.6, where each round plans a trajectory, trains candidate adapters, evaluates, analyzes failed trajectories, and curates the next data mixture. The key contribution is a repeatable training loop that treats data‑mixture construction as the primary search object, not a bespoke benchmark prompt or dataset. Across the clinical suite, Cura 1T yields the strongest gains from setting‑specific data actions—tool‑use trajectories for MedAgentBench, cleaner behavior data for HealthBench, knowledge and reasoning repair with retention for MedXpertQA, and task‑specific interactive trajectories for AgentClinic.

Out‑of‑domain evaluations on AIME, GPQA‑Diamond, and $\tau$2‑Bench indicate that these healthcare gains do not compromise general reasoning or agentic ability. Nevertheless, our results are bounded by current compute, data, and resource constraints, limiting the scope of the training regime. Future work should broaden clinical behavior coverage, target long‑horizon agentic tasks, and explore full‑parameter updates as resources allow.

Treating data‑mixture construction as the core search object enables robust healthcare specialization without sacrificing general capabilities.

Training Configuration and Details

Full hyperparameters, harness APIs, and detailed case‑study trajectories are documented.

The appendix records the exact training configuration, the concrete tool‑call interfaces used for evaluation, and the step‑by‑step evolution of model behavior on representative clinical tasks.

Two harnesses expose the model to realistic clinical tool calls: AgentClinic for order‑test/diagnosis workflows and MedAgentBench for FHIR‑based interactions.

AgentClinic replaces free‑text routing with two deterministic functions—`order_test`(`test_name`) to request a finding and `submit_diagnosis`(diagnosis) to end the encounter—while preserving ordinary dialogue for history taking.

MedAgentBench provides three native operations—`fhir_get`, `fhir_post`, and finish—that interact with a live FHIR server; writes are logged for grading without mutating the server state.

The case‑study section details how Cura 1T iteratively refines its behavior across multiple interaction rounds, turning initial failures into correct clinical actions.

In the MedAgentBench referral task, the base model issues a routine‑priority request; after two tool‑use rounds it learns to emit a stat priority with the exact SNOMED code, and a harness bug fix finally yields a fully compliant submission, raising the reward from 0 to 1.

For the HealthBench drug‑identification query “EROSTIN 10 MG”, the base model cannot name the active ingredient; a behavior‑correction step supplies the correct antihistamine (ebastine), and Cura 1T retains this knowledge, achieving a reward of 1.

The Sciatica coding case shows the base model correctly listing laterality codes but then dropping the unspecified‑side digit after a correction; Cura 1T restores the full M54.30 code and preserves laterality distinctions, improving the reward from 0.973 back toward 1.

In the DCIS radiotherapy benefit question, early rounds mistakenly focus on positive margins as a benefit factor; later reasoning injects the clinical constraint that positive margins require further surgery, leading Cura 1T to select the correct answer (option D) and obtain a reward of 1.

The NEJM‑Ext 21 interstitial pneumonia case illustrates that the base model misdiagnoses the condition; after gathering imaging and biopsy evidence through tool calls, Cura 1T correctly identifies desquamative interstitial pneumonia, raising the reward from 0 to 1.

Contributors: Haolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, Frank Wang, Weiran Yao.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers