ShuohaoGao-THU/Salus-7B
Salus-7B by ShuohaoGao-THU is a 7.6 billion parameter medical language model built on Qwen2.5-7B-Instruct, specifically fine-tuned for sequential diagnostic testing in complex clinical cases. It employs supervised fine-tuning and Group Relative Policy Optimization to decompose diagnostic reasoning into differential diagnosis, strategic control, and workup proposal. This specialized model is optimized for evidence-gathering processes in medical diagnosis, achieving 83.64% Top-1 diagnostic accuracy on the CompDiag-Bench.
Loading preview...
Salus-7B: Specialized Medical Diagnosis Model
Salus-7B is a 7.6 billion parameter medical language model developed by ShuohaoGao-THU, based on the Qwen2.5-7B-Instruct architecture. It is specifically designed for sequential diagnostic testing in complex clinical scenarios, as detailed in the ICML 2026 paper "Salus: Strategic Diagnostic Testing for Complex Diagnosis via Multi-Agent Reinforcement Learning".
Key Capabilities
This model is optimized for three distinct functional roles in diagnostic reasoning:
- Differential Reasoner: Generates a differential diagnosis based on patient records.
- Strategic Controller: Decides whether to request more evidence (auxiliary examinations) or finalize a diagnosis.
- Workup Proposer: Recommends specific auxiliary examinations when more evidence is needed.
Salus-7B was trained using supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO), which calibrates evidence-seeking behavior and discourages premature diagnostic closure. It achieves 83.64% Top-1 diagnostic accuracy on complex cases in the CompDiag-Bench, outperforming DeepSeek-V3.2 (71.38%) and GPT-5.2 (80.30%) in this specific benchmark.
Important Considerations
- Specialized Use: Salus-7B is not a general-purpose chat model; its performance is optimized for the three fixed diagnostic instructions, primarily in Chinese.
- Research Only: This model is released for research purposes and is not validated for autonomous clinical use. It should not replace professional medical judgment.
- Limitations: The model may produce incorrect or incomplete diagnoses and recommendations. Users are responsible for evaluating its suitability and safety in their specific context.