Model merging: the linearization point decides interference
A new method shows that the directional derivative computed at the receiver model predicts interference in model merging, halving accuracy error compared to a no-change baseline. At the shared backbone it fails, especially in vision tasks. LR-Merge u...