Skip to content

Cura 1T: Specialized Model for Agentic Healthcare

🕒 Published (v1): 2026-07-15 00:00 UTC · Source: HuggingFace · link

Why this paper was selected

1T-param agentic healthcare model covering reasoning + workflow execution; applied medical AI

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Cura 1T is a 1-trillion-parameter healthcare-specialized LLM post-trained on Kimi-K2.6 via a human-gated self-evolution loop in which an LLM agent iteratively diagnoses failure modes, synthesizes targeted training data, and curates the post-training mixture. It ranks first on 5 of 6 evaluated healthcare benchmark panels while preserving out-of-domain reasoning and agentic performance. The key novelty is treating the data mixture—not the training algorithm or hyperparameters—as the primary optimization target.

Problem

Frontier LLMs handle individual healthcare sub-tasks (clinical QA, patient-facing dialogue, EHR tool use) in isolation, but no single specialized model covers all three simultaneously without capability interference: a targeted data update for one task routinely degrades another. Healthcare post-training is further constrained by scarce, heterogeneous supervision signal, making naive mixture scaling insufficient.

Method

Cura 1T is built by fine-tuning Kimi-K2.6 with LoRA adapters through a closed self-evolution loop. Each round proceeds in four steps:

  1. Plan — A training agent defines target behaviors, benchmarks, metrics, and a candidate data recipe; a human approves the plan before training begins.
  2. Train — The candidate adapter is trained through an SFT → RL → SDFT stack. SFT acts as a fast sanity screen. RL optimizes against task-level reward signals. SDFT (self-distillation fine-tuning) consolidates the final adapter via the KL objective: $\(L(\theta) = \mathbb{E}_{y \sim \pi_\theta(\cdot|x)}\left[\log\frac{\pi_\theta(y|x)}{\pi(y|x,c)}\right]\)$ where \(c\) is privileged context (corrected trajectory, retrieved knowledge, or rubric) visible only to the teacher, keeping student updates anchored to on-policy traces.
  3. Evaluate — Benchmarks run, and graded trajectories are collected.
  4. Refine — The agent applies five targeted data-synthesis skills: Reasoning Correction (edits failed traces with a frontier model, then generates pattern-matched variants), Knowledge Injection (retrieval-grounded closed-book QA), Behavior Calibration (rubric-conditioned response synthesis), Retention Anchors (examples of already-solved behaviors to prevent forgetting), and Other (task-specific agentic workflow failures). A validation gate filters for format, safety/PII, deduplication, and repair coverage before a human issues a keep/revert/deploy decision.

The loop runs independently per capability; kept rounds accumulate into a consolidated mixture used for the final Cura 1T checkpoint.

Key Contributions

  • A data-first self-evolution loop for healthcare post-training where an LLM agent optimizes the training mixture rather than prompts, code, or hyperparameters.
  • SDFT integration as the final training stage to anchor updates to on-policy traces, mitigating damage to long medical reasoning chains from off-policy supervision.
  • Human-gated review at plan approval and keep/revert/deploy decision points, balancing automation with safety in high-stakes domain adaptation.
  • A suite of targeted data-synthesis skills (Reasoning Correction, Knowledge Injection, Behavior Calibration, Retention Anchors) that decompose failure modes into addressable data interventions.
  • Demonstrated cross-capability consolidation: capability-specific evolution paths are merged into a single model that improves uniformly across patient care, clinical reasoning, and agentic EHR tasks without eroding general math/science/agentic benchmarks.

Results

  • MedAgentBench (EHR FHIR tool-use): Cura 1T 0.940 vs. base 0.847 (+0.093); best frontier (Claude Opus 4.8) 0.937.
  • HealthBench Professional (physician rubric): Cura 1T 0.662 vs. base 0.503 (+0.159); best among all evaluated models.
  • HealthBench Hard: Cura 1T 0.368 vs. base 0.222 (+0.146); best among evaluated models.
  • MedXpertQA overall pass@1: Cura 1T 0.655 vs. base 0.569 (+0.086); second to GPT-5.5 (0.675); beats Claude Opus 4.8 (0.628).
  • AgentClinic overall pass@1: Cura 1T 0.796 vs. base 0.754 (+0.042); Claude Opus 4.8 reaches 0.841.
  • Evolution path peak before consolidation: MedAgentBench 0.973 (Round 3), MedXpertQA 0.636 (Round 4); consolidation slightly reduces per-benchmark peaks while maintaining cross-task balance.
  • Out-of-domain benchmarks (AIME, GPQA-Diamond, \(\tau^2\)-Bench): competitive with base, indicating no obvious regression.

Limitations

  • Consolidation cost: per-capability peaks (e.g., MedAgentBench 0.973) drop after cross-benchmark mixture consolidation (0.940), revealing residual tension between tasks that the loop partially but not fully resolves.
  • Human bottleneck: plan approval and keep/revert/deploy decisions require domain-expert human review, limiting throughput and making the loop less autonomous than fully automated alternatives.
  • Base-model ceiling: Cura 1T remains second to GPT-5.5 on MedXpertQA and to Claude Opus 4.8 on AgentClinic, suggesting the base model (Kimi-K2.6) constrains the ceiling on some tasks.
  • LoRA-only updates: training only low-rank adapters rather than full weights may limit the magnitude of achievable behavioral change, particularly for deeply embedded failure modes.
  • No out-of-domain healthcare evaluation: generalization to healthcare settings not covered by the five evaluation benchmarks (e.g., radiology report generation, surgical planning) is unassessed.
  • Reversion overhead: several evolution rounds were reverted (Behavior Correction, Reasoning Correction, Data Mixture Tuning), indicating the loop is still empirically brittle and expensive to navigate.

Relevance to Agentic AI / LLM Agents

Cura 1T instantiates an LLM-agent-as-trainer paradigm in which an agent plans, executes, and evaluates post-training experiments end-to-end, extending the auto-research trend (AI Scientist, autoresearch) to the healthcare domain. Its EHR tool-use capability (MedAgentBench FHIR calling) is a direct instance of long-horizon agentic task completion under structured tool schemas, making the model itself a deployable healthcare agent. The self-evolution loop's architecture—failure analysis → targeted synthesis → mixture curation → human gate—is a reusable template for capability-controlled agent specialization in any low-data, high-stakes domain. The reverted rounds and consolidation cost also provide concrete empirical data on capability interference, a central unsolved problem in multi-task agentic fine-tuning.