Attribution patching

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Attribution Graphsmethod0.815
  Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
• Attribution Similaritymethod0.803
  Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality
• Data Attributionconcept0.802
  The task of attributing model behaviors to specific training datapoints.
• Attribution graph constructionmethod0.770
  Method to trace how parameter subcomponents interact from input to output for a given next-token prediction, producing a subnetwork graph.
• Activation patchingmethod0.767
  Standard method in mechanistic interpretability that intervenes on activations; VPD flips this paradigm by patching parameters.
• Probe-based data attribution for alignmentconcept0.749
• Path Patchingmethod0.747
  Method by Goldowsky-Dill et al. 2023 for localizing model behavior via targeted activation interventions
• Gradient-based data attributionmethod0.745
  Baseline method against which probe-based ranking is compared; more computationally expensive.