Separate task effects from selection-rule effects in layer ablations

Determine whether the differing depth profiles observed for the proposed graded-contrast weight-editing method and CreativityNeuro’s set-difference method arise from the behavioral tasks or from the respective weight-selection rules by conducting layer ablations with the set-difference mask.

Background

The paper finds that the proposed edit’s effect is concentrated relatively early in the network, whereas CreativityNeuro reports an effect concentrated in later layers. The authors attribute this difference mainly to the tasks, but note that the relevant ablation for CreativityNeuro’s set-difference mask was not performed. Consequently, the source of the contrasting depth profiles remains unresolved.

References

We have not run the layer ablation with the set-difference mask, which would separate task from rule directly.

— Resisting Harmful Instructions through Training-Free Weight Editing  (2610.21189 - Ebrahimpour-Boroojeny et al., 15 Oct 2026) in Appendix A.3, “Where in depth the edit acts”