Full orthogonal projection can be too aggressive. Scale the edit:
W' = W - α · (W projected onto span{r})Or equivalently blend:
W' = (1-α)W + α · W_ablatedChoosing α#
| α | Effect |
|---|---|
| 0.25 | Mild — fewer capability hits, refusals may remain |
| 0.50 | Balanced starting point for experiments |
| 1.00 | Full ablation — maximum refusal reduction, highest damage risk |
Sweep α ∈ {0.25, 0.5, 0.75, 1.0} on a small eval suite before committing.