Low-rank subspace recovery¶
Operate in the low-rank subspace of the safety alignment delta.
| Method | Mechanism |
|---|---|
| LOX | Low-rank extrapolation via SVD |
| LSSF | Low-rank safety subspace fusion |
| Safety vector restore | Truncated SVD of safety vector |
Operate in the low-rank subspace of the safety alignment delta.
| Method | Mechanism |
|---|---|
| LOX | Low-rank extrapolation via SVD |
| LSSF | Low-rank safety subspace fusion |
| Safety vector restore | Truncated SVD of safety vector |