steering-block-selection

Rank layers by gradient/attribution scores or an activation signal like probe accuracy, or start with mid-to-late layers as a heuristic. If a single layer shows no effect, intervene on 3-5 layers instead of one.

A single-site intervention on a deep stack often gets denoised by downstream normalization and attention, producing a false negative. Widen the intervention to 3-5 layers or re-screen layers by gradient or activation signals.

Choose the component based on what the claim is about. Circuit discovery studies generally target attention heads plus MLP, while broader behavioral steering often targets the residual stream.

Never copy a raw index across models of different depth. Scale by relative depth first (layer 16 of 32 becomes layer 24 of 48), then check neighboring layers to confirm the site.

Intervening on the whole stack causes off-distribution collapse, so the observed effect is collapse-driven and the localization claim becomes untestable. Use spaced intervals every 2-3 layers when sweeping the stack.

Re-sweep the coefficient whenever the site set changes, because the coefficient plateau is site-dependent. Lock the site selection first, then tune the coefficient.