PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2607.15314v2cs.AI

TL;DR

Healthcare AI lacks specialized agentic models that jointly support patient care, multimodal clinical reasoning, interactive diagnosis, and EHR workflows. Cura 1T addresses this gap with a human-gated recursive self-improvement loop, ranking at or near the top across healthcare benchmarks while preserving strong out-of-domain reasoning and agentic performance.

  • Problem

    Healthcare requires models spanning patient care, multimodal clinical reasoning, and long-horizon agentic workflows, but suitable training data and supervision remain fragmented and limited.

  • Method

    Cura 1T uses a human-gated recursive self-improvement loop that plans target behaviors, trains, evaluates failures, and refines targeted data mixtures.

  • Results

    Cura 1T is strongest on five of six healthcare benchmark panels and ranks second on the remaining MedXpertQA multimodal tasks.

  • Takeaways & Limitations

    Cura 1T remains competitive on out-of-domain reasoning and agentic benchmarks, with no obvious erosion of general capability from continual learning.

Abstract

from arXiv · show

Healthcare AI agents handle patient consultation, clinical reasoning over text and images, interactive diagnosis, and electronic health record (EHR) tool use, yet specialized agentic models that cover these use cases together remain limited. These capabilities fail in different ways, and a narrow update for one task can degrade another. We present Cura 1T, a healthcare-specialized LLM built on the open-weight Kimi-K2.6 and trained through a human-gated recursive self-improvement (RSI) loop. Specifically, in each round, the RSI harness plans a target capability, trains the model, evaluates benchmark trajectories, and refines the data mixture from observed failures with targeted synthetic and curated examples rather than a single generic medical-data update. Across the healthcare evaluation suite, Cura 1T ranks at or near the top among frontier baselines while remaining competitive on out-of-domain reasoning and agentic benchmarks.

1 Introduction

Cura 1T is a healthcare-specialized LLM designed to unify patient care, multimodal clinical reasoning, and healthcare-agentic EHR use through human-gated recursive self-improvement. It achieves leading healthcare benchmark performance while preserving strong out-of-domain reasoning and agentic capability.

  • Motivation: Healthcare deployment demands clinician- and patient-facing guideline adherence, expert reasoning across text and images, and electronic-health-record workflow use.These use cases recur throughout the report and require a broader capability profile than answering difficult clinical questions alone.
  • Problem: Healthcare model building is a data-construction problem because supervision is scarce, sensitive, fragmented across sources, and difficult to verify with simple answer keys.Examples span guidelines, exams, patient interactions, images, and electronic-health-record workflows, while failure modes are unevenly distributed.
  • Method: Cura 1T is trained through a human-gated recursive self-improvement loop that plans capabilities, trains the model, evaluates trajectories, diagnoses failures, and rebuilds the data mixture.The approach uses targeted synthetic and curated examples rather than a single generic medical-data update.
  • Results: Five of six healthcare benchmark panels rank Cura as strongest, while it ranks second on the remaining MedXpertQA multimodal tasks.The same model consistently improves over Kimi-K2.6 across patient-care, clinical-reasoning, and healthcare-agentic evaluations.
  • Results: Cura 1T preserves strong performance on out-of-domain math, scientific-reasoning, and agentic benchmarks, indicating no obvious erosion of general reasoning or agentic capability.These results assess whether continual learning compromises capabilities beyond healthcare.

2 Related Work

Prior healthcare AI work spans medical question answering, continuous-care systems, and agentic clinical capabilities. Related post-training and recursive self-improvement methods provide foundations for adapting models through demonstrations, rewards, self-generated data, critique, and metric-driven optimization.

  • Healthcare AI: Healthcare language-model research spans patient care, clinical reasoning, and agentic clinical systems.Med-PaLM targets licensing-exam questions, while Baichuan-M4 adds continuous-care capabilities including tool use, patient memory, action constraints, evidence retrieval, and multimodal perception.
  • LLM post-training: LLM post-training combines supervised adaptation, reinforcement learning, and self-training.SFT commonly provides task-specific demonstrations and a cold start; RL optimizes against task-level rewards, followed by distillation or consolidation, while rejection sampling retains high-scoring generations.
  • Recursive self-improvement and auto-research: Recursive self-improvement methods use critique, task feedback, or metric-based optimization to improve models and language-model pipelines.Self-Refine and Reflexion improve subsequent attempts through critique or feedback, while OPRO, DSPy, and TextGrad optimize prompts, programs, or pipeline components.

3 actAVA Cura

Cura 1T is a healthcare-specialized model built on Kimi-K2.6 for patient care, clinical reasoning, and healthcare agentic tasks. Its recursive self-improvement loop uses human-gated planning, staged training, evaluation, and failure-driven data refinement to improve targeted capabilities while preserving solved behavior.

  • 3 actAVA Cura: Cura 1T is post-trained on Kimi-K2.6 for patient care, clinical reasoning, and healthcare agentic tasks, with a 256K context window and native text-plus-vision input.Patient care targets clear, safe responses aligned with physician rubrics, while clinical reasoning spans expert medical questions over text and images.
  • Human-gated control: Human review gates the plan and the keep, revert, or deploy decision after evaluation, while kept rounds continue training and contribute validated data refinements.The protocol evaluates candidate trajectories and uses evidence for a keep-or-revise decision; reverted rounds remain in the experimental record.
  • Recursive self-improvement: The RSI harness adds failure analysis, targeted data synthesis, and mixture curation to the closed-loop search over training configurations and behaviors.This data-first design changes mixture decisions to address missing capabilities while preserving previously solved behavior.
  • Training loop: Each round uses SFT as a low-cost screen, then RL for reward-driven improvement and SDFT as the continual-learning step.SFT checks whether the proposed data mixture and hyperparameters train stably and move target metrics in the intended direction before RL and SDFT.
  • Data refinement: The refinement pipeline categorizes failures, synthesizes targeted trajectories, and validates new rows for format, safety, PII risk, duplicates, evaluation overlap, and repair coverage.Reusable skills include retention anchoring, reasoning correction, knowledge injection, behavior calibration, task-specific repair, and mixture curation.

4 Experiments

Cura 1T is evaluated across healthcare benchmarks spanning patient-facing responses, clinical reasoning, interactive diagnosis, and EHR tool use, with targeted refinement addressing task-specific failures. It improves substantially on healthcare tasks while preserving competitive performance on out-of-domain reasoning and agentic benchmarks.

  • Healthcare evaluation: The evaluation suite covers MedAgentBench, HealthBench Professional and Hard, MedXpertQA, and AgentClinic across patient-facing quality, clinical reasoning, diagnosis, and EHR tool use.MedXpertQA reports text, multimodal, and overall pass@1; AgentClinic reports subset and overall pass@1 under a simulated clinic harness.
  • HealthBench: 0.634 on HealthBench Professional and 0.372 on HealthBench Hard are reached by the clean behavior mix, whereas the broader behavior-correction round is reverted after degraded-subset declines.The degraded subset falls by 0.252 on Professional and 0.508 on Hard relative to base-model performance.
  • MedXpertQA: 0.655 overall pass@1 is reached on MedXpertQA after consolidation, exceeding the base and Claude Opus 4.8 while remaining second to GPT-5.5.Knowledge injection + retention reaches 0.603 overall pass@1, and mixture refinement reaches 0.636 before consolidation.
  • AgentClinic: 0.807 overall pass@1 is achieved by Interactive trajectory + retention on AgentClinic, while consolidated Cura 1T reaches 0.796 and matches the best reported NEJM score.The largest gains occur on the NEJM and NEJM-Ext subsets.
  • MedAgentBench: 0.973 task success is obtained on MedAgentBench after RSI rounds, while released Cura 1T scores 0.940 after consolidation and exceeds the strongest frontier reference at 0.937.Consolidation costs 0.033 on this task while requiring the model to hold scores on other evaluation panels.
  • Out-of-domain evaluation: Cura 1T is on par with frontier comparators on each reported out-of-domain reasoning benchmark and τ 2-airline, while surpassing the mThe supplied passage truncates the comparison for τ 2-retail and τ 2-telecom.

5 Conclusion · A Appendix

Cura 1T is presented as a healthcare-specialized model built through recursive self-improvement from Kimi-K2.6. It leads GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro on five of six key clinical and operational benchmark panels.

  • 5 Conclusion: Cura 1T is built through a recursive self-improvement loop.The model is described as healthcare-specialized.
  • 5 Conclusion: Cura 1T is post-trained from Kimi-K2.6.
  • 5 Conclusion: Cura 1T is described as the strongest healthcare LLM today.
  • 5 Conclusion: 5 of the six benchmark panels are led by Cura 1T over GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro.These panels are characterized as the ones that matter most in clinical and operational work.
  • 5 Conclusion: Healthcare requires clinician-grade patient communication.
  • 5 Conclusion: Healthcare requires expert reasoning over clinical text and images.
  • 5 Conclusion: Healthcare requires reliable execution against the real systems that payers and providers run.

A.1 Training Hyperparameters

Table 7 reports Cura 1T’s final training hyperparameters. The values reflect final consolidation across capability-varying screening runs, not capability-specific screening settings or mixture proportions.

  • Hyperparameters vary by capability across the screening runs.
  • The table reports final consolidation rather than capability-specific screening runs or mixture proportions.
  • Table 7 reports the final training hyperparameters for Cura 1T.

A.2 Harness Implementations · A.3 Case Studies · A.3.1 MedAgentBench: Referral ServiceRequest

The harness implementations replace ambiguous control protocols with native healthcare tools, while the referral case study shows that iterative correction resolves exact FHIR-contract failures and earns full reward.

  • A.2 Harness Implementations: AgentClinic uses native order_test and submit_diagnosis tools, while history-taking remains ordinary assistant dialogue passed through the patient simulator.Test results enter trajectories only through executed order_test calls, eliminating ambiguity between requested and reported findings.
  • A.2 Harness Implementations: MedAgentBench exposes FHIR reads, writes, and final submission through fhir_get, fhir_post, and finish.Reads execute against the live FHIR server; writes are logged for grading without mutating the server, and episodes permit up to eight turns.
  • A.3 Case Studies: The case study requires an orthopedic surgery referral ServiceRequest for synthetic patient S2016972 with SNOMED 306181000000106, active status, order intent, stat priority, exact subject reference, and an SBAR note.The requested timestamp is 2023-11-13T10:15:00+00:00.
  • A.3.1 MedAgentBench: Referral ServiceRequest: Reward 1 follows the Round 1 tool-use response, which submits fhir_post with a ServiceRequest, stat priority, and SNOMED 306181000000106.The result passage records the initial referral submission and reward.
  • A.3.1 MedAgentBench: Referral ServiceRequest: Later responses preserve priority=stat, exact SNOMED coding, subject, timestamp, and SBAR note, satisfying the referral’s schema and content constraints.These responses correct the base model’s failure to meet the exact FHIR contract despite understanding the referral intent.
  • A.3.1 MedAgentBench: Referral ServiceRequest: Under the corrected harness, the response combines priority=stat with exact coding and finish([]), receiving Reward 1.The harness bug fix confirms the correction under the intended evaluation protocol.

A.3.2 HealthBench Professional: EROSTIN 10MG

This example tests whether the model identifies EROSTIN 10MG as ebastine while avoiding confusion with estradiol or Elestrin. Cura 1T preserves the corrected identification, although the behavior-correction round was reverted because of subset-level degradation.

  • Task and criterion: The criterion requires identifying ebastine as EROSTIN 10MG’s active ingredient and avoiding confusion with estradiol or Elestrin.The prompt asks what medication EROSTIN 10MG is.
  • Base response: The base model declines to identify Erostin 10 mg from the brand name alone.This is characterized as a drug-identification failure rather than a reasoning failure.
  • Behavior correction: Behavior correction correctly identifies EROSTIN 10 mg as an antihistamine containing ebastine, but the round is reverted for subset-level degradation.The correction fixes the example locally but is not retained globally.
  • Retention and final model: Cura 1T identifies EROSTIN 10 mg as an antihistamine containing ebastine as its active ingredient.The intermediate retention check also preserves the corrected active ingredient.

A.3.3 HealthBench Professional: Sciatica Coding Regression

The sciatica-coding trace shows a subset-level regression: behavior correction changed the correct unspecified-side code M54.30 to the incomplete parent category M54.3, whereas Cura 1T restored billable laterality-specific ICD-10-CM coding. This degradation contributed to reverting Round 1 despite its higher overall score.

  • Task and criterion: Correct ICD-10-CM sciatica coding requires M54.30 for unspecified side, M54.31 for right, and M54.32 for left; M54.3 is incomplete.The task distinguishes billable laterality codes from the parent category and from the WHO category.
  • Regression trace: Behavior correction reduced the unspecified-side code from M54.30 to M54.3 while retaining M54.31 and M54.32, changing reward from 1.000 to 0.000.The correction preserved the code family but dropped the final digit required for the billable unspecified-laterality code.
  • Cura 1T correction: Cura 1T restored M54.30, M54.31, and M54.32, distinguished ICD-10-CM from the WHO category, and corrected the incomplete recommendation.This matched trace illustrates the subset-level degradation that caused Round 1 to be reverted despite its higher overall score.

A.3.4 MedXpertQA Text-197: DCIS Radiotherapy Benefit

This MedXpertQA case asks which DCIS patient gains the greatest local-control benefit from radiotherapy, with the gold answer being option D: a 0.9 cm grade-3 DCIS after lumpectomy with a negative margin. Cura 1T corrected an initial positive-margin error by learning that positive margins require further surgery rather than simply increasing radiotherapy benefit, then retained the correct answer.

  • Task and criterion: The gold answer is D: a 0.9 cm grade-3 DCIS after lumpectomy with a negative margin.This patient is identified as gaining the greatest local-control benefit from radiotherapy.
  • Initial response: The base model selected option F, combining grade 3 histology with a positive margin, and received Reward 0.Its reasoning treated the positive margin as indicating the highest baseline recurrence risk.
  • Iterative correction: After reasoning correction and knowledge injection, the model selected option D and received Reward 1.The corrected reasoning recognized that positive-margin cases are not rescued by radiotherapy alone.
  • Mixture refinement: Retained rounds preserved option D by learning that positive margins require further surgery rather than simply increasing radiotherapy benefit.Mixture refinement and consolidation retained this clinical constraint.

A.3.5 AgentClinic NEJM-Ext 21: Desquamative Interstitial Pneumonia · B Contributors

The AgentClinic case tests evidence-led diagnosis of desquamative interstitial pneumonia, contrasting incorrect early diagnoses with Cura 1T’s retained multi-turn workup. The section also lists the paper’s contributors.

  • A.3.5 AgentClinic NEJM-Ext 21: Desquamative Interstitial Pneumonia: The 34-year-old patient’s gold diagnosis is desquamative interstitial pneumonia, based on tobacco use, dyspnea, dry cough, imaging, lavage, and biopsy findings.The biopsy showed pigment-laden macrophages alongside diffuse ground-glass opacities with peripheral consolidation.
  • A.3.5 AgentClinic NEJM-Ext 21: Desquamative Interstitial Pneumonia: The base model proposed Chronic Eosinophilic Pneumonia and received reward 0, reverting its diagnostic attempt.The recorded response was: “The most likely diagnosis is Chronic Eosinophilic Pneumonia.”
  • A.3.5 AgentClinic NEJM-Ext 21: Desquamative Interstitial Pneumonia: Cura 1T incorrectly declared pulmonary alveolar proteinosis despite normal anti-GM-CSF and nondiagnostic BAL/PAS results.The trajectory recorded “DIAGNOSIS READY: Pulmonary alveolar proteinosis.”
  • A.3.5 AgentClinic NEJM-Ext 21: Desquamative Interstitial Pneumonia: After requesting biopsy review and receiving extensive alveolar filling with pigment-laden macrophages, Cura 1T submitted desquamative interstitial pneumonia.This response used the discriminating pathology evidence to reach the gold diagnosis.
  • A.3.5 AgentClinic NEJM-Ext 21: Desquamative Interstitial Pneumonia: The retained interactive trajectory teaches gathering and using discriminating evidence before diagnosis, preserving evidence-led workup behavior.The case is explicitly characterized as multi-turn and dependent on evidence collection before diagnosis.
  • B Contributors: The contributors are Haolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, Frank Wang, and Weiran Yao.These names are listed under the B Contributors subsection.
Loading 2607.15314v2…