Spectral outliers in attention weight matrices, identified by the Marchenko-Pastur threshold, contain a dominant share of functionally important learned structure across 11 transformers.
A mathematical framework for transformer circuits,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Spectral Outliers Reveal Dominant Learned Structure in Transformer Attention
Spectral outliers in attention weight matrices, identified by the Marchenko-Pastur threshold, contain a dominant share of functionally important learned structure across 11 transformers.