Whole-model recovery¶
Restore safety by editing the entire model's weights. These methods treat the model as a single unit.
| Method | Mechanism |
|---|---|
| Task arithmetic | Add safety task vector v = aligned - base |
| SOMF merge | Subspace-oriented model fusion |
| Pre-post merge | Interpolate toward pre-fine-tuning checkpoint |
| WiSE-FT | Weight-space interpolation toward aligned model |