Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of Readout
A two-sided geometric diagnostic is introduced that measures an extracted subspace's overlap with the dominant right-singular directions of the unembedding matrix, evaluated against output-oriented positive controls.
Ao-Jie Yuan, Z. Su, Hai-Yu Zhang et al.
· 0 citations