causal importance network

Auxiliary model trained alongside VPD to predict which subcomponents are causally important for each prompt, enabling mechanistic isolation of components.

Neighborhood — ranked by edge-count

Papers (1)

paper

Interpreting Language Model Parameters
introduces

Methods (1)

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.

Concepts (1)

concept

Causal importance
implementsrelated_to
A measure of whether a subcomponent is necessary to reproduce model behavior on a specific prompt, predicted by the causal importance network.

Related by similarity (8)

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.

Causal Mechanismconcept0.753
Function determining the value of a variable based on its causal parents in an acyclic causal model.
Causal powerconcept0.752
The ability of an agent to be a driver of subsequent events; a hallmark of cognition that causal emergence quantifies.
Causal Influence Diagramsframework0.742
Framework informing path-specific objectives by identifying causal chains leading to risky behaviors
Causal abstractionconcept0.742
A framework the paper uses alongside feature geometry to deepen mechanistic understanding of LMs
Causal Invarianceconcept0.740
Property that causal mechanisms remain stable across environments; desirable for OOD.
Causal Mediationconcept0.732
Whether an internal direction causally controls a target behavior, verified by intervention success
Gene Regulatory Networksframework0.727
Causal Tracingconcept0.724
Mechanistic interpretability technique for locating factual associations, mentioned as future work direction.