Multi-layer steering with layer-incremental vectors
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — introduced
An extension of single-layer steering that applies a 'layer-incremental' persona vector (the difference between consecutive layers' persona directions) at every layer, avoiding cumulative double-counting of overlapping directional information. More effective than single-layer steering at suppressing trait acquisition during preventative steering, and better preserves MMLU accuracy during post-hoc steering.
If steering at one layer helps, steering at every layer sounds like it should help more — except simply repeating the same correction at every layer risks double-counting signal that has already propagated forward from earlier layers, rather than adding anything new. This method's fix is to steer with only each layer's incremental contribution instead of the full vector, layer by layer.
What that layer-incremental adjustment buys for the paper's two steering-based mitigations differs by direction: for one, it closes a gap where single-layer intervention wasn't fully suppressing an unwanted trait; for the other, it achieves comparable suppression while doing noticeably less damage to the model's general capability.
What it is
Multi-layer steering extends the paper's single-layer intervention, applied only at "the layer identified as most informative for a given trait" (persona-vectors, §"J.3 Steering across all layers", p. 46), to every layer of the model at once. Because persona vectors are extracted from the residual stream at each layer separately, naively injecting the same persona vector at every layer would double-count overlapping directional signal that has already propagated forward from earlier layers' injections. The fix is a layer-incremental vector:
$$\vec{v}_\ell^{\text{incremental}} = \vec{v}_\ell - \vec{v}_{\ell-1}$$
"where $\vec{v}_\ell$ denotes the persona direction at layer $\ell$. This formulation ensures that the steering applied at each layer reflects only the incremental contribution of that layer's representation, avoiding redundant or excessive influence" (persona-vectors, §"J.3 Steering across all layers", p. 46).
What it fixes for preventative steering
Single-layer Preventative steering does not always fully block trait acquisition, particularly on datasets intentionally designed to elicit the trait (persona-vectors, §"5.2 Preventative steering limits behavioral shifts during finetuning", p. 8). Applying layer-incremental vectors throughout training instead pushes trait acquisition down to near-baseline levels even on the hardest datasets, achieving comparable suppression with smaller coefficients and with no additional MMLU cost relative to ordinary finetuning (persona-vectors, §"J.3 Steering across all layers", p. 46-47). This means single-layer preventative steering's shortfall was one of incomplete suppression, not of capability damage; the multi-layer version closes exactly that specific gap rather than trading one problem for another.
The same fix, a different payoff for post-hoc steering
Applied to post-hoc rather than preventative steering, the identical layer-incremental construction pays off along a different axis: all-layer inference-time steering achieves trait suppression similar to single-layer steering, but with better preserved MMLU accuracy (persona-vectors, §"J.3 Steering across all layers", p. 46-47). One technical fix, avoiding redundant, double-counted directional signal across layers, therefore strengthens preventative steering along the suppression axis and post-hoc steering along the capability-preservation axis, depending on which direction along the vector the intervention is pushing.