VPD scales to a 4-layer 67M-parameter model trained on The Pile.

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

• VPD enables manual model editing through direct parameter manipulation.claim0.792
  Applied capability claim: VPD enables surgical changes to model behaviour at the parameter level.
• VPD achieves better sparsity-reconstruction tradeoff than transcoders on 67M modelfinding0.789
  Empirical result demonstrating VPD's efficiency advantage in parameter decomposition.
• VPD decomposes parameters, not activations, flipping the standard SAE / activation-patching paradigm.claim0.772
  Core proposition of the paper: a substrate-level critique of existing interpretability methods.
• Does VPD mechanistic faithfulness and interpretability survive at frontier model scale?question0.769
  Open research question about whether VPD generalizes beyond the tested 67M-parameter regime.
• The ability to make precise edits demonstrates that VPD identifies real computational machineryclaim0.750
  Claim that editing success validates VPD's decomposition.
• VPD subcomponents are sparse, interpretable, and avoid feature splitting.claim0.749
  Assertion about the qualitative advantages of VPD's rank-one decomposition.
• VPD identifies real, computational structure in neural network parametersclaim0.746
  Central claim that VPD successfully uncovers genuine mechanisms.
• Phi-4 shows U-shaped cohesion with falling mismatch; peak depth varies by modelfinding0.741
  E3 backbone-specific finding showing three-stage trajectory generalizes across architectures