Neuron / cluster recovery¶
Edit individual neurons or neuron clusters to restore safety behaviour.
| Method | Mechanism |
|---|---|
| NLSR | Neuron-level safety realignment |
| MSCP | Multi-level safety continual projection |
| Antidote v1 | WANDA importance pruning |
| Antidote v2 | Layer-adaptive WANDA + utility floor |
| PKE | Parameter knowledge editing for safety neurons |
| SafeReAct | Reactivate dormant safety neurons via probe inputs |
| SCRUB (recover) | Teacher-student distillation recovery (training-based) |