method
active
method:activation-patching-causal-interventionsActivation patching / causal interventions
Substitute activations to observe effects.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (2)
framework
- Three-step approach to identify roles of network components.
- Five-step pipeline: task definition, decomposition, description, description-validation loop, final validation.
Methods (5)
method
- Method of shifting hidden state activations along probe directions to cause the model to treat false statements as true and vice versa; evaluated on OOD inputs
- Causal Intervention via Activation Shiftrelated_toIntervening in model forward pass by adding/subtracting probe direction to group (b) hidden states to flip truth judgments
- Mean-activation patching interventionrelated_toSubstitutes mean activation for count c to modify perceived counts.
- Attribution patching / AtP*extendsGradient-based approximation to activation patching.
- Path PatchingextendsMethod by Goldowsky-Dill et al. 2023 for localizing model behavior via targeted activation interventions
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Standard method in mechanistic interpretability that intervenes on activations; VPD flips this paradigm by patching parameters.
- The general experimental approach of intervening along geometrically-defined paths rather than single-point or linear-direction interventions
- The use of interventions (rather than correlations) to establish a causal link between representation geometry and behavioral geometry.
- Methods that intentionally introduce divergent representations to test sufficiency and completeness of circuits
- Confound where naming injected concepts reflects direct logit effects rather than metacognitive awareness, raised by Morris & Plunkett
- Activation patching of arithmetic intermediates is effective at layers matching J-lens emergence.finding0.769Causal accuracy.
- Pearson correlation of feature activations across 40M tokens used to measure feature similarity and universality across models
- Assertion that understanding causal emergence may lead to methods for manipulating agent representations to improve performance.