SafeMed-R1: Clinician-Audited Safety and Ethics for Medical LLMs
There's a new medical language model called SafeMed-R1 that's designed to boost safety and ethics in clinical settings. It uses a unique system called Clinical Trust Signals (CTS) to link reasoning instances to clinician evaluations and editing histories, making the reasoning process transparent. This model has achieved a macro-averaged accuracy of 79.6% across different clinical tests. In safety assessments, it shows the lowest risk level and reduces unsafe outputs by about 3 to 5% compared to its previous version. A study with 30 medication safety scenarios found that SafeMed-R1 matches the medical accuracy of PGY1 and PGY2 residents but outperforms them in safety, adherence to guidelines, and clinical relevance. You can check out this research on arXiv with the identifier 2605.28338.
Key facts
- SafeMed-R1 is a medical LLM trained with a Clinical Trust Signals pipeline.
- The pipeline links reasoning instances to clinician rubric scores and edit histories.
- SafeMed-R1 achieves 79.6% macro-averaged accuracy on clinical benchmarks.
- It reduces unsafe outputs by 3-5% relative to its baseline under adversarial testing.
- In a study of 30 medication safety vignettes, it matched PGY1 and PGY2 residents on correctness.
- It scored higher than residents on medication safety, guideline consistency, and clinical usefulness.
- The research is published on arXiv with ID 2605.28338.
- The model addresses governance requirements for auditable reasoning and ethics alignment.
Entities
Institutions
- arXiv