Research
How Cura 1T turns evaluation traces into better training data
Cura’s recursive self-improvement loop diagnoses failed trajectories, synthesizes targeted examples, and tests whether each training update preserves previously learned capabilities.
Healthcare model development produces many kinds of evidence: a clinical answer, a diagnostic conversation, an EHR tool call, and the evaluation record attached to each. Cura 1T uses those records to guide the next training mixture. A failed trajectory becomes the starting point for a specific data-construction decision.
Cura is fine-tuned from Kimi-K2.6 through recursive self-improvement. Its research report follows capability development across patient communication, clinical reasoning over text and images, interactive diagnosis, and healthcare tool use.
Translate a failure into a training intervention
The loop begins with a plan that names a target behavior, evaluation metrics, and a data recipe. After training, the harness collects evaluation trajectories and groups their failures by cause. The refinement stage then produces examples aimed at the missing capability.
Different causes lead to different interventions. A reasoning error can guide the construction of a new clinical case. A knowledge gap can lead to examples grounded in retrieved material. A tool-use failure can identify a schema or workflow requirement that the next mixture should exercise. This structure keeps the relationship between observed behavior and proposed training data visible.
Preserve useful behavior through each round
The data mixture also contains retention anchors: examples for capabilities the current model already handles. Review considers the target metric alongside behavior carried forward from earlier rounds. Accepted refinements accumulate into the consolidated mixture used for Cura 1T.
The report includes an illustrative referral task in a synthetic EHR environment. Later model traces satisfy the required referral priority and structured resource fields. The example shows why an execution trace is useful: it lets researchers inspect the specific contract the model learned to follow.
Evaluate across different healthcare tasks
In the reported MedAgentBench evaluation, task success rises from 84.7% for the base model to 94.0% for Cura 1T. HealthBench Professional rubric scores rise from 50.3 to 66.2, and HealthBench Hard scores rise from 22.2 to 36.8. The report also evaluates medical text and image questions, simulated diagnosis, and out-of-domain capabilities.
These evaluation records make the development process inspectable at several levels: the individual failure, the data intervention, the candidate checkpoint, and the consolidated model.
Research by Haolin Chen, Weiran Yao, and the Actava.ai team. Explore the model and evaluation results on the Cura 1T page.
Source: Cura 1T: Healthcare Foundation Model via Recursive Self-Improvement.