Pairwise matrices for SAEs demonstrate that single-feature inspection mislabels causal axes, with joint suppression and matched-geometry controls revealing distinct output regimes not captured by single-feature or random perturbations.
TIGHTER TAN (TIGHTER TAN)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Pairwise matrices for sparse autoencoders: single-feature inspection mislabels causal axes
Pairwise matrices for SAEs demonstrate that single-feature inspection mislabels causal axes, with joint suppression and matched-geometry controls revealing distinct output regimes not captured by single-feature or random perturbations.