HuggingFace Papers
7/10 signal
Cura 1T: Specialized Model for Agentic Healthcare
agenticresearchmodelseval
Summary
This paper introduces Cura 1T, a healthcare-specialized LLM based on Kimi-K2.6, trained to handle patient care, clinical reasoning, and agentic tasks. The model is developed using a human-gated self-evolution loop where a training agent plans experiments, trains the model, evaluates failures on benchmarks, and refines the data mixture accordingly. This data-centric approach focuses on synthesizing targeted examples to fix specific failure modes rather than using a single generic medical dataset. Across a suite of six healthcare benchmark panels, Cura 1T ranks as the strongest model on five and second on the sixth, while remaining competitive on out-of-domain reasoning and agentic tasks.
Problem
Frontier language models can answer clinical questions but are not robust enough for broad healthcare deployment, which requires a combination of patient care communication, expert clinical reasoning, and agentic workflow execution (e.g., using electronic health records). These capabilities fail in different ways, and training for one can degrade performance in another. This is primarily a data-construction challenge, as high-quality, comprehensive healthcare training data is scarce, fragmented, and difficult to obtain.
Method
Cura 1T, a 1-trillion parameter model, is post-trained from Kimi-K2.6 using low-rank adapters (LoRA) within a human-gated self-evolution loop. In each round, an LLM agent plans a training run, trains a candidate model using a sequence of Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and Self-Distillation Fine-Tuning (SDFT), and evaluates it. The agent then analyzes failures on benchmarks like MedAgentBench, HealthBench, MedXpertQA, and AgentClinic to synthesize targeted data corrections (e.g., knowledge injection, reasoning correction, behavior calibration) and curate the next data mixture.
Details
Healthcare Benchmark Performance vs. Base Model:
| Benchmark | Base (Kimi-K2.6) | Cura 1T | Δ |
|---|---|---|---|
| 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 |
MedAgentBench (Task Success):
- Cura 1T achieves a task success rate of 0.940, outperforming Claude Opus 4.8 (0.937), Gemini 3.1 Pro (0.913), and GPT-5.5 (0.894).
- The self-evolution loop improved the score from a base of 0.883 to 0.973 in a benchmark-specific round by targeting brittle execution of EHR writes with synthetic tool-use trajectories.
HealthBench (Rubric Score):
- On HealthBench Professional, Cura 1T scores 0.662, the highest among evaluated models.
- On HealthBench Hard, Cura 1T scores 0.368.
- An early training round that improved aggregate scores was reverted because it caused significant degradation on a subset of tasks (e.g., a 0.508 drop on Hard for that subset), highlighting the need for targeted, clean data mixtures.
MedXpertQA (pass@1):
| Model | Text | Multimodal | Overall |
|---|---|---|---|
| Cura 1T | 0.600 | 0.722 | 0.655 |
| GPT-5.5 | 0.596 | 0.771 | 0.675 |
| Claude Opus 4.8 | 0.562 | 0.710 | 0.628 |
| Kimi-K2.6 (Base) | 0.484 | 0.672 | 0.569 |
- Cura 1T is the second-strongest model overall. The improvement was driven by adding closed-book clinical knowledge and retention examples, as reasoning-only corrections were found to be unstable and were reverted.
AgentClinic (pass@1):
- Using a tool-native harness, Cura 1T achieves an overall pass@1 of 0.796, outperforming Claude Opus 4.8 (0.794) and GPT-5.5 (0.684).
- The key improvement came from training on interactive trajectories with retention anchors, which taught the model to complete a clinical workup before making a premature diagnosis.
Out-of-Domain Performance:
- On reasoning benchmarks (AIME 2025, AIME 2026, GPQA-Diamond), Cura 1T performs on par with frontier models.
- On agentic benchmarks (τ²-Bench), Cura 1T is on par with comparators on the airline domain and surpasses publicly reported scores on the retail and telecom domains, indicating that healthcare specialization did not erode general capabilities.
What's new
The primary contribution is the methodology: a human-gated, data-centric self-evolution loop for specialized model training. Unlike prior self-improvement work that optimizes prompts, code, or hyperparameters, this loop treats the training data mixture as the primary object of search. An agent analyzes specific model failures on evaluation benchmarks and synthesizes targeted data to address them, enabling robust improvement in a data-scarce domain like healthcare.
Limitations
The authors state that results are bounded by constraints on compute, data, and other resources. Future work should broaden clinical behavior coverage, address long-horizon agentic tasks, and explore full-parameter updates. The paper also includes a safety warning that Cura 1T is a research model and not a medical service, and strong benchmark scores do not guarantee safety for clinical use.
Conclusion
Cura 1T demonstrates that a specialized healthcare LLM can be effectively trained via a self-evolution loop that focuses on data mixture curation. By analyzing benchmark failures and synthesizing targeted data, this method improves performance across patient care, clinical reasoning, and agentic healthcare tasks. The model achieves state-of-the-art or near state-of-the-art results on a suite of healthcare benchmarks while preserving general reasoning and agentic capabilities on out-of-domain evaluations.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.