Optimal intervention-layer scope

Determine the optimal layer bands and widths for the learned steering and hypernetwork interventions, and establish whether method-specific optimization of intervention scope narrows the performance gap between them.

Background

The experiments evaluate learned steering and the hypernetwork at only two scopes: a narrow middle-layer band and the full model depth. The results show that the narrow band outperforms full-depth intervention, but they do not establish that the selected band is optimal.

The authors identify a complete sweep over layer bands and widths for each method as necessary future work, noting that the observed gap between methods could diminish if each method were evaluated at its own best scope.

References

The learned steering and hypernetwork was evaluated at two scopes, so we can only establish that a narrow band beats full depth, not that our selected band is optimal. Our layer choices should be read as a working configuration rather than a tuned one. A full sweep over bands and widths per method is left to future work, and the gap between methods may narrow once each is given its own best scope.

— Predicting Steering Vectors and Adapter Weights for Few-Shot Author-Style Transfer  (2610.03163 - Popp et al., 2 Oct 2026) in Section “Limitations”