Preventative Steering

Novel method that steers the model toward an undesired persona direction during training to cancel out pressure imposed by the training objective

Neighborhood — ranked by edge-count

paper

concept

Preventative Steering During Training
related_to
Alternative to inference-time activation capping: applying persona steering during training to deeply anchor models; cited from Chen et al.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Preventative Promptingmethod0.802
Alternative to preventative steering: prepending a trait-eliciting system prompt to training samples to cancel out training pressure
steering (intervention on internals)concept0.773
General technique of modifying activations to control model behavior.
Stepwise steeringmethod0.766
Novel method that applies intervention only when the model begins a new thinking step (at the \n\n delimiter) rather than at every token
direction-based steeringconcept0.762
Paradigm of finding the right direction in activation space (e.g., linear steering).
Bidirectional Steeringconcept0.761
Ability to steer model behavior in two opposite semantic directions on a trait.
steering vectorsconcept0.758
A method for modifying model behavior by adding perturbation vectors to activations, used here to try to reduce eval awareness.
Pullback Steeringmethod0.758
The method of optimizing steering interventions in activation space to produce outputs that follow the behavior manifold, independent of the representation manifold.
Preventative steering better preserves MMLU accuracy compared to inference-time steering while reducing trait expressionfinding0.754
Key advantage of preventative over post-hoc steering: lower side-effect cost on general capabilities