Safin-1: Safety from Within through Memory-Native State Evolution
Safin-1 tests a model architecture where safety behavior is carried in internal memory states, not just added after training.
The paper introduces MARCH, a routing system that keeps structured memory and retrieves relevant history during long interactions. Its “Safety State” is used to adapt the model on downstream safety tasks without repeatedly changing the backbone. The authors report gains across safety, long-context understanding, retrieval, efficiency, and general capability evaluations, while framing the work as an early architectural exploration. Source: HF Daily Papers' note.
The paper introduces MARCH, a routing system that keeps structured memory and retrieves relevant history during long interactions. Its “Safety State” is used to adapt the model on downstream safety tasks without repeatedly changing the backbone. The authors report gains across safety, long-context understanding, retrieval, efficiency, and general capability evaluations, while framing the work as an early architectural exploration. Source: HF Daily Papers' note.
score 5