Parameter subcomponent

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

• Rank-one subcomponentsconcept0.799
  The atomic pieces into which weight matrices are decomposed by VPD; each rank-one component is interpretable.
• are the resulting parameter subcomponents actually interpretable objects?question0.788
  First question posed after applying VPD, investigating whether the subcomponents make sense.
• A good parameter subcomponent is causally important only for specific roles and can be removed from the model without hurting performance on irrelevant promptsclaim0.784
  Definitional principle guiding VPD: subcomponents should encode narrow, targeted computational roles rather than distributed, multi-purpose functionality.
• Parameter subcomponents cleanly isolate true mechanisms of the modelclaim0.780
  Interpretive claim that the subcomponents correspond to real functional units.
• can we use these parameter subcomponents to perform clean, targeted changes?question0.772
  Implicit question driving the editing experiment.
• Attention computations distribute across heads via parameter subcomponents with interpretable rolesfinding0.737
  Mechanistic discovery about how attention mechanisms decompose into interpretable parameter components.
• VPD subcomponents are sparse, interpretable, and avoid feature splitting.claim0.732
  Assertion about the qualitative advantages of VPD's rank-one decomposition.
• Emoticon eye subcomponentconcept0.731
  The part of the emoticon subcomponent responsible for recognizing the 'eyes' of emoticons like ';', ':' or '=', which was edited in the demo.