source: arxiv artificial intelligence: cura 1t: specialized model for agentic healthcare
level: research
cura 1t is a large language model specialized for healthcare tasks. it handles patient consultation, clinical reasoning with text and images, interactive diagnosis, and electronic health record tool use. the model is built using a human-gated self-evolution loop. in each round, a training agent picks a target capability, trains the model, checks benchmark results, and adjusts the data mix based on failures. this approach uses synthetic and curated examples to fix specific weaknesses instead of applying a single broad medical-data update.
the self-evolution loop is data-centered. it avoids the common problem where improving one task hurts performance on another. by focusing on observed failures, the model gets better at multiple healthcare use cases without trade-offs. the process repeats over several rounds, each time refining the training data to address gaps. this method keeps the model aligned with real-world clinical needs and reduces the risk of degrading existing skills.
in evaluations across a healthcare benchmark suite, cura 1t ranks at or near the top compared to frontier models. it shows strong results in tasks that require both medical knowledge and practical workflow execution. the model's design makes it suitable for agentic healthcare applications where reliability and safety are critical. the self-evolution framework could also apply to other specialized domains needing continuous, targeted improvement.
why it matters: this approach shows how to build safer, more reliable ai for healthcare by fixing specific failures without breaking other capabilities, which is crucial for real-world clinical use.
source: arxiv artificial intelligence: cura 1t: specialized model for agentic healthcare