Summary (Overview)

• Cura 1T is a 1-trillion-parameter healthcare-specialized LLM post-trained from the Kimi-K2.6 base using a human-gated self-evolution loop that treats the data mixture as the primary object of search.
• The model achieves state-of-the-art or near-top performance across five healthcare benchmark panels: MedAgentBench, HealthBench Professional, HealthBench Hard, MedXpertQA, and AgentClinic.
• The self-evolution loop automates the cycle of planning, training, evaluating, and refining the data mixture, using targeted synthetic data construction (e.g., tool-use trajectories, behavior calibration, knowledge injection, retention anchors) rather than a single generic medical-data update.
• Cura 1T improves over the base model by as much as +0.159 on HealthBench Professional and +0.093 on MedAgentBench, while preserving strong out-of-domain performance on AIME, GPQA-Diamond, and τ₂-Bench.
• The central contribution is a repeatable training loop that identifies missing capabilities from graded failures, synthesizes corrective data, and curates a mixture that avoids unintended forgetting.

Introduction and Theoretical Foundation

The paper addresses the challenge of building a single LLM that can handle three core healthcare use cases:

1. Patient care – clinician- and patient-facing responses that follow physician guidelines.
2. Clinical reasoning – expert medical question answering across text and images.
3. Healthcare agentic tasks – multi-turn diagnostic workups, EHR tool use, and long-horizon workflows.

These tasks have distinct failure modes (rubric omissions, missing facts, brittle reasoning, premature diagnosis, malformed tool calls), and a narrow update for one task can degrade another. The authors argue that the primary bottleneck is not a training algorithm or hyperparameter configuration, but data construction: healthcare training signal is scarce, fragmented across guidelines/exams/patient interactions/images/EHR workflows, and often cannot be checked by a simple executor.

The theoretical foundation combines:

• Low-rank adaptation (LoRA) (Hu et al., 2021) for parameter-efficient training of a 1T model.
• Self-distillation fine-tuning (SDFT) (Shenfeld et al., 2026), which anchors updates to the model’s own on-policy trajectories to avoid damaging native reasoning patterns.
• Self-evolution and auto-research patterns (e.g., Self-Refine, Reflexion, autoresearch), but with the data mixture as the search object rather than prompts, code, or hyperparameters.

Methodology

The training is built on a thin adapter-training stack over Kimi-K2.6 (256K context, text + vision input). The core innovation is the human-gated self-evolution loop (Figure 2 in the paper):

Self-Evolution Loop Components

Training Stack

• SFT – low-cost prerequisite to verify mixture and hyperparameters.
• RL – reward-driven improvement stage.
• SDFT – final consolidation step. The SDFT objective is:

$$\mathcal{L}(\theta) = D_{\text{KL}}(\pi_\theta(\cdot|x) \parallel \pi(\cdot|x,c)) = \mathbb{E}_{y \sim \pi_\theta(\cdot|x)}\left[\log \frac{\pi_\theta(y|x)}{\pi(y|x,c)}\right]$$

where $x$ is the prompt, $c$ is privileged context (e.g., an intervened trajectory, reference behavior, verified knowledge), and $\pi_\theta(\cdot|x)$ is the student distribution. The student sees only the original prompt after training, which keeps updates anchored to the model’s own producible trajectories.

Data Refinement Skills

The agent uses five specialized synthesis skills:

• Retention Anchor: Add examples of already-solved behaviors to prevent forgetting.
• Reasoning Correction: Edit failed reasoning traces and seed new questions.
• Knowledge Injection: Identify missing clinical knowledge, ground with retrieved documents, synthesize QA rows.
• Behavior Calibration: Compare desired vs. observed behavior, synthesize rubrics, generate conditioned responses.
• Other: Task-specific failures (e.g., agentic workflow errors).
• Data Mixture Curation: Merge, deduplicate, cap, and reweight the training set.

The loop runs separately for each capability and is kept or reverted based on evaluation. The final Cura 1T model is trained from the consolidated mixture accumulated across successful rounds.

Empirical Validation / Results

Healthcare Benchmarks

The paper evaluates Cura 1T on five benchmark panels. Below are the key tables.

Table 1: Improvement from Kimi-K2.6 base

Table 3: MedAgentBench task success

Table 4: HealthBench rubric scores at T=1.0

Table 5: MedXpertQA split and overall pass@1 at T=1.0

Table 6: AgentClinic pass@1 by subset under tool-native protocol

Out-of-Domain Evaluation

Cura 1T was evaluated on:

• Reasoning: AIME 2025, AIME 2026, GPQA-Diamond – performance on par with frontier models.
• Agentic: τ₂-Bench (airline, retail, telecom) – matches or surpasses reported scores, particularly on retail and telecom.

Evolution Path Summary

The evolution map (Figure 3) shows that of the 11 attempted interventions (across all benchmarks), 7 were retained and 4 were reverted due to subset degradation, insufficient factual coverage, non-termination, or premature diagnosis. The final Cura 1T model uses the consolidated mixture from retained rounds.

Theoretical and Practical Implications

• Data mixture is the primary search object: The central lesson is that healthcare specialization is a data-mixture problem before a hyperparameter-search problem. The self-evolution loop succeeded because it used evaluation failures as evidence for what data was missing.
• Different failure modes require different data actions: HealthBench needed behavior calibration (cleaner principle-rubric data), MedXpertQA needed knowledge injection with retention, MedAgentBench needed tool-use trajectories, and AgentClinic needed interactive trajectory training. A single generic medical-data update would not address these diverse failure modes.
• Retention anchors are critical: Adding examples of already-solved behaviors prevents targeted repairs from causing unintended forgetting. This was especially important for MedXpertQA, where reasoning correction alone degraded performance until retention anchors were added.
• SFT as a screening step: The use of SFT before RL/SDFT to verify mixture and hyperparameters at low cost is a practical design choice for large-model training.
• Out-of-domain preservation: The healthcare gains did not collapse general reasoning or agentic ability, suggesting that the self-evolution loop, with retention anchors, successfully avoids catastrophic forgetting.

Conclusion

Cura 1T demonstrates a repeatable, data-centric self-evolution loop for building a healthcare-specialized LLM. The key contributions are:

1. A human-gated loop that plans, trains, evaluates, and refines the data mixture using specialized synthesis skills (reasoning correction, knowledge injection, behavior calibration, retention anchors).
2. State-of-the-art results on five healthcare benchmarks, with the largest gains from tool-use trajectories (MedAgentBench), cleaner behavior data (HealthBench), knowledge+retention (MedXpertQA), and interactive trajectories (AgentClinic).
3. Preserved out-of-domain performance on math, science, and general agentic benchmarks.
4. Empirical evidence that a single generic medical-data update is insufficient; the data mixture must be adapted to the specific failure modes of each task.

Limitations: The results are bounded by current compute, data, and training regime. Future work should broaden clinical behavior coverage, target long-horizon agentic tasks, and explore full-parameter updates.

Safety and clinical use: Cura 1T is a research model, not a medical service. Strong benchmark scores do not establish safety for unsupervised clinical use.