method
active
method:attribution-patching-atpAttribution patching / AtP*
Gradient-based approximation to activation patching.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (2)
method
- Attribution patchingrelated_toGradient-based method to estimate the effect of zeroing a feature on a specific logit difference.
- Substitute activations to observe effects.
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.
- Used by prior work to identify consistency heads.
- Measure causal importance of upstream variables.
- Standard method in mechanistic interpretability that intervenes on activations; VPD flips this paradigm by patching parameters.
- The task of attributing model behaviors to specific training datapoints.
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality
- Method by Goldowsky-Dill et al. 2023 for localizing model behavior via targeted activation interventions