For in-context linear regression, merged key/query linear attention learns via one abrupt loss drop, while separate key/query attention learns via multiple drops, with each stage adding one principal component of the input distribution.
A convergence analysis of gradient descent for deep linear neural networks
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Training Dynamics of In-Context Learning in Linear Attention
For in-context linear regression, merged key/query linear attention learns via one abrupt loss drop, while separate key/query attention learns via multiple drops, with each stage adding one principal component of the input distribution.