Directional Ablation

Intervention method that removes a direction from residual stream activations to disrupt corresponding behavior

Neighborhood — ranked by edge-count

paper

framework

Concept Cones
uses
The central framework this paper extends from refusal to propositional truth; identifies multi-dimensional subspaces that causally mediate target behaviors

concept

Surgical Ablation Property
associated_with
Property requiring that ablating a truth direction shifts model output from truthful to false without other side effects

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.

Zero Ablationmethod0.805
Intervention type that sets activations to zero, used for interpretability analysis
Grafting and Ablationmethod0.789
Classical techniques to interrogate regulative capacity of embryos and neural crest by tissue removal or transplantation.
Adversarial ablationmethod0.769
Technique used in VPD to enforce mechanistic faithfulness of parameter decompositions.
Term Importance Analysis via Ablationmethod0.749
An algorithm that determines the marginal effect of n-th order path terms by running the model multiple times with frozen attention patterns and progressively replacing activations
spatializationconcept0.719
The translation of semantic values into spatial coordinates and relations.
Direct Manipulationconcept0.712
Special case of immediate feedback loop where user interacts with artifacts in a lifelike manner, typically through cursor or finger-based dragging.
Goal-Directed Activityconcept0.706
Central property of agency: energy expended to reach specific states despite disturbances.
Boost Level Ablation Sweepmethod0.701
Systematic sweep of 10 boost levels from threshold-3σ to threshold+3σ to characterize ESR vs. steering strength