Search papers, labs, and topics across Lattice.
The paper introduces Safin-1, a family of foundation models designed to enhance safety through an intrinsic mechanism called Safety from Within, which leverages memory routing and state evolution. By employing a novel architecture known as Memory-Anchor Routing across Context History (MARCH), Safin-1 enables models to maintain structured memory states and adapt capabilities dynamically during interactions. Experimental results demonstrate that this approach leads to significant safety improvements in downstream tasks, reframing memory as an active component for evolving model behavior rather than a passive record.
Safety in AI can now evolve from within the model itself, drastically enhancing long-term interaction capabilities without external constraints.
Long-horizon complex tasks require foundation models to accumulate information, maintain internal states, and adapt over extended interactions. Safety should be an intrinsic property of the model itself, rather than a behavioral constraint relying solely on external safeguards or post-hoc alignment such as supervised fine-tuning. This motivates Safety from Within, where safety-relevant capabilities are represented and invoked through the model's native computation. We present Safin-1, a family of foundation models realizing this principle through memory routing and state evolution. Safin-1 is built on Memory-Anchor Routing across Context History (MARCH), a network architecture that maintains structured memory states and selectively retrieves relevant historical information through content-conditioned routing. It supports test-time adaptation of persistent capability states without repeatedly modifying the backbone, enabling controlled specialization over a shared foundation. We investigate this interface on downstream safety tasks through a Safety State, demonstrating effective state-based adaptation with substantial safety improvements. More broadly, the routed-state interface unifies contextual memory and persistent capability adaptation within the model's native computation, reframing memory from a passive record of prior context into an active substrate for maintaining and evolving model behavior. Evaluations across general capabilities, long-context understanding, retrieval, and efficiency further validate Safin-1. These findings provide a path toward safety as a state-native and adaptively maintainable capability. This work is only an initial architectural exploration of Safety from Within, and substantial further work is needed to realize this broader vision.
