Approximate Causal Abstraction

Graded notion of causal abstraction measured by IIA; when IIA is alpha < 100%, the model is alpha-on-average approximately abstract.

Neighborhood — ranked by edge-count

paper

method

Interchange Intervention Accuracy
implements
Proportion of aligned interchange interventions with equivalent high-level and low-level effects; graded measure of causal abstraction.

concept

Causal abstraction
related_to
A framework the paper uses alongside feature geometry to deepen mechanistic understanding of LMs
Constructive Causal Abstraction
extends
Formal definition: H is a constructive abstraction of L under alignment Π when interchange interventions have equivalent effects at both levels.

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 abstraction analysismethod0.849
The formal method used to establish that the identified circuit causally mediates the model's cyclic reasoning behavior
Causal Abstraction over Cyclic Tasksconcept0.791
What is the connection between information encoding assumptions and causal abstraction?question0.791
Identified as exciting future work direction
An interplay between causal abstraction and feature geometry deepens mechanistic understanding of language modelsclaim0.778
Methodological claim about the scientific value of combining causal abstraction with representational geometry analysis
Causal abstraction is not enough for mechanistic interpretability because it becomes vacuous without assumptions about how models encode informationclaim0.777
Central thesis of the paper
Causal importanceconcept0.770
A measure of whether a subcomponent is necessary to reproduce model behavior on a specific prompt, predicted by the causal importance network.
interpretive abstraction (method)method0.769
Programming technique to restructure a fine-grained Linda program for efficiency by replacing live data structures with passive ones and coarser-grain processes.
Causal abstraction implicitly relies on strong assumptions about feature encoding in DNNs, and becomes trivial without such assumptionsclaim0.762
Authors' interpretation connecting their proof to practical interpretability methodology