finding
active
finding:spd-finds-102-w-in-components-and-a-rank-51-w-out-component-spanning-all-three-layers-in-the-3-layer-modelSPD finds 102 W_in components and a rank-51 W_out component spanning all three layers in the 3-layer model
APD struggled beyond two layers.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- SPD is more scalable and robust to hyperparameters than APDassociated_withsupportsCentral claim.
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.
- SPD finds 100 W_in components and a single rank-50 W_out component spanning both layers in the 2-layer modelfinding0.945Multi-layer mechanisms.
- SPD splits W_out into 50 subcomponents forming a single rank-50 component in compressed computation modelfinding0.858Contradicts APD's 100 components for W_out.
- SPD decomposes W_in of compressed computation model into 100 subcomponents, one per input featurefinding0.800Neuron contributions match target.
- Ground truth mechanisms identified.
- Figure 1.
- Neuron contributions of SPD subcomponents closely match target model across all 100 featuresfinding0.765Figure 7.
- Empirical demonstration of VPD on a mid-scale transformer, establishing feasibility.
- Author interpretation.