How it works

SAVeR2 defines four fine-grained reward signals for safety and uses gradient projection to stop safety updates from overwriting reasoning skills.