Cura 1T: Specialized Model for Agentic Healthcare

Specialized LLM for clinical workflows trained via a human-gated self-evolution loop to improve patient consultation, multimodal clinical reasoning, interactive diagnosis, and EHR tool use. Iteratively refines targeted synthetic and curated data based on benchmark failures to raise specific capabilities without broad regressions.

Visit Website

Introduction

Healthcare tasks pull in conflicting demands: safe, empathetic patient conversation; rigorous stepwise clinical reasoning over text and images; interactive diagnostic probing; and correct use of EHR tools. That tension makes one-shot medical fine-tuning risky—improving one skill can degrade another. Cura 1T's core insight is to treat capability development as a planned, iterative loop where a training agent proposes target capabilities, generates and curates data, evaluates trajectories on benchmarks, and updates the data mixture based on observed failure modes.

Key Capabilities

Targeted capability growth so what: each training round focuses on a narrowly defined clinical ability (e.g., differential diagnosis with imaging prompts), which reduces collateral performance loss on unrelated tasks and yields measurable gains on specific benchmarks.
Human-gated loop so what: human reviewers gate the synthetic data and trajectory selection, keeping automated evolution aligned with clinical safety constraints and reducing harmful regression risk.
Multimodal and agentic behavior so what: supports text+image clinical reasoning and interaction with EHR-like tool interfaces, making the model usable for both consultation-style dialogs and workflow automation.
Benchmark-driven refinement so what: failures on held-out evaluation trajectories directly inform the next data mixture, accelerating correction of real-world weaknesses rather than broad, unfocused updates.

Who it's for and trade-offs

Great fit if you need a model tailored for clinical workflows that must balance conversational safety, reasoning fidelity, and tool-mediated actions—teams building clinical assistants, diagnosis support tools, or EHR automation pipelines. Look elsewhere if you need an off-the-shelf, fully validated production medical AI (Cura 1T focuses on capability development and evaluation rather than regulatory certification), or if you require a lightweight general-purpose LLM for non-clinical tasks.

Methodological note

The paper emphasizes a data-centered training regime: a planning agent specifies capability goals, generates synthetic examples and collects curated cases, automated evaluations run patient-to-disposition trajectories, and humans gate dataset updates. This loop is presented as a way to iteratively concentrate training signal where the model fails, contrasting with single-shot medical fine-tuning approaches.

Back

Information

Websitearxiv.org
OrganizationsactAVA AI
AuthorsHaolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, Frank Wang, Weiran Yao
Published date2026/07/15

More Items

Large Language Model Papers2026

Language Models Can Control Their Own Attention

Namgyu Ho, Huzama Ahmad +4KAIST AI, Google DeepMind

Introduces Declarative Attention (DA), a zero-shot protocol that has LMs declare which parts of long context to attend to during chain-of-thought, letting the runtime build dynamic attention masks and skip most KV-cache reads. Produces large token savings (up to ~52% on Gemma-4-31B) with modest accuracy loss.

llm long-horizon vllm gemma qwen+5

Natural Language Processing Papers2026

It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning

Runpeng Dai, Kaili Huang +2University of North Carolina at Chapel Hill, Apple

Generates compact keyword sets for both queries and items with LLMs and matches them directly via an inverted index. Uses supervised fine-tuning to align keyword spaces, then alternates GRPO-based reinforcement learning on query- and item-side generators to co-evolve representations and maximize retrieval F1 while staying compatible with keyword-based infrastructure.

retrieval LLM RL sft benchmark+2

Large Language Model Papers2026

SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers

Shaowen Wang, Ge Zhang +7

Studies looping shared transformer layers in Mixture-of-Experts models under matched budgets and proposes SMELT: loop the middle half twice while matching per-token FLOPs, non-embedding parameters, and KV cache. Shows 6.8–18.0% training-FLOPs savings on the compute-optimal frontier, stronger downstream gains on code and long-context tasks.

moe transformers llm research benchmark+2