Specificity of the harmful single-head intervention effect

Determine whether the wrong-direction effect produced by single-head L11 intervention in RACE×RELIG requires the paired group’s donor activation or is also produced by coherent identity donors from other same-type demographic groups.

Background

For RACE×RELIG, replacing the L11 single-head activation produces a consistent shift away from the target group’s truth. Donor-control experiments indicate that this is not merely generic activation corruption: dimension-shuffled donors perturb the output more strongly but do not produce systematic directional drift.

Donors from other same-type identities show weak trends in the same direction, but each individual comparison is nonsignificant. The direct contrast between the paired donor and the mean of the alternative donors is also suggestive but unresolved, leaving the identity specificity of the backfire open.

References

Whether it further requires the paired group's donor is suggestive but unsettled: donors from other same-type identities trend the same way weakly (each n.s.), and the direct paired contrast against their mean reaches $t=-2.0$, exact $p=0.073$.

Readable, Faithful, Used: Three Dissociable Properties of Demographic Identity in a Language Model  (2608.18768 - Robbani, 19 Aug 2026) in Section 5.2, item 2 of the enumerated findings