Layer updates in language models consist of a dominant tokenwise reparameterization plus a geometrically separate residual that carries most of the functional significance.
naacl-main.401/
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
-
On the Geometric Structure of Layer Updates in Deep Language Models
Layer updates in language models consist of a dominant tokenwise reparameterization plus a geometrically separate residual that carries most of the functional significance.