Weight matrix decomposition

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

• Rank-one matrix decompositionmethod0.805
  Constraint in VPD where each parameter subcomponent is constrained to be a rank-one matrix for simplicity.
• Transition Matrix (B matrix)concept0.738
  Prescribes transitions among hidden state factors under action; encodes policy-dependent dynamics
• Singular Value Decompositionmethod0.737
  Used to summarize principal patterns of internal functional states.
• Weight spaceconcept0.733
  The space of the model's parameter matrices, where VPD operations take place.
• Decomposition of all 24 weight matrices in a 67M-parameter LM yields ~10,000 parameter subcomponentsfinding0.732
  Quantitative result of VPD application; the network's 24 matrices decompose into approximately 10,000 rank-one subcomponents.
• Activation decompositionconcept0.726
  The conventional approach (e.g., SAEs, transcoders) of decomposing activations into interpretable features.
• Rank-one matrixconcept0.724
  The simple matrix form into which VPD constrains subcomponents to enforce mechanistic simplicity.
• Copying Matrixconcept0.720
  An OV circuit matrix that maps tokens to increasing the logit of those same tokens; detectable via positive eigenvalues