Adversarial ablation

Technique used in VPD to enforce mechanistic faithfulness of parameter decompositions.

Neighborhood — ranked by edge-count

concept

VPD (adVersarial Parameter Decomposition)
implements
Core methodological framework introduced in this paper; decomposes weight matrices into rank-one interpretable subcomponents using adversarial ablations.

claim

Adversarial ablations enforce mechanistic faithfulness.
cites
Methodological claim that the adversarial ablation approach ensures decomposed components causally correspond to computation.

method

Adversarial Parameter Decomposition (VPD)
implements
Core technique introduced in this paper for decomposing neural network weight matrices into mechanistically simple, interpretable rank-one subcomponents.

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.

adversarial interactionconcept0.799
Competitive multi-agent setting with conflicting incentives and direct opposition via bidding and bluffing.
Zero Ablationmethod0.781
Intervention type that sets activations to zero, used for interpretability analysis
Grafting and Ablationmethod0.777
Classical techniques to interrogate regulative capacity of embryos and neural crest by tissue removal or transplantation.
Contrastive Learning Ablation Studymethod0.773
Three-setting ablation (Before Training, Without CL, With CL) to isolate the contribution of contrastive learning
Coefficient Ablationmethod0.773
Robustness check ablating over five steering coefficient values to test whether SP/AS inversion is an artifact of steering strength
Adversarial Suffix Attackconcept0.771
Optimization-based jailbreak method appending strings to prompts to elicit harmful outputs.
Directional Ablationmethod0.769
Intervention method that removes a direction from residual stream activations to disrupt corresponding behavior
Adversarial Manipulation of Truthfulnessconcept0.767
Risk that multiple truth directions enable attacks that shift outputs without triggering the primary truth direction