method
active
method:nonnegative-matrix-factorization-of-decoder-normsNonnegative matrix factorization of decoder norms
Applied to 30-dim (layer, model) norm vectors to find shared structure across scale.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Post-hoc metric for model exclusivity in standard crosscoders.
- Correlative technique measuring the type of information encoded in distributed representations via linear predictability.
- The total Hamiltonian is equally consistent with any factorisation; establishes the boundary as a modelling choice not a physical fact
- Detects tiling in real SAEs.
- Footnote 2.
- Sparsity penalty weighting by sum of per-layer decoder norms, enabling comparability with per-layer SAEs.
- Training-free technique normalizing all task gradients to the maximum gradient norm magnitude
- Key design choice encouraging feature exclusivity.