VPD subcomponents are sparse, interpretable, and avoid feature splitting.

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

• VPD subcomponents avoid feature splitting, improving interpretability over SAE approachclaim0.885
  Core interpretative claim that VPD's parameter-based decomposition prevents the feature fragmentation seen in activation-based methods.
• VPD is a meaningful step toward bottom-up interpretabilityclaim0.803
  Positioning of VPD as advancing the paradigm of explaining computation in the model's terms.
• VPD achieves a better sparsity-reconstruction tradeoff than transcoders.claim0.790
  Quantitative advantage claimed for VPD over a prior activation-decomposition method.
• VPD decomposes parameters, not activations, flipping the standard SAE / activation-patching paradigm.claim0.769
  Core proposition of the paper: a substrate-level critique of existing interpretability methods.
• The ability to make precise edits demonstrates that VPD identifies real computational machineryclaim0.764
  Claim that editing success validates VPD's decomposition.
• are the resulting parameter subcomponents actually interpretable objects?question0.763
  First question posed after applying VPD, investigating whether the subcomponents make sense.
• VPD achieves better sparsity-reconstruction tradeoff than transcoders on 67M modelfinding0.762
  Empirical result demonstrating VPD's efficiency advantage in parameter decomposition.
• Future interpretability techniques will fundamentally resemble VPDclaim0.757
  Prediction/hypothesis about the direction of the field.