Influence Malleability in Linearized Attention: Dual Implications of Non-Convergent NTK Dynamics

Jose Marie Antonio Miñoza, Paulo Mario P. Medina, Sebastian C. Ibañez · 2026

Read Influence Malleability in Linearized Attention: Dual Implications of Non-Convergent NTK Dynamics by Jose Marie Antonio Miñoza, Paulo Mario P. Medina, Sebastian C. Ibañez (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13085.txt.

Description

Jose Marie Antonio Miñoza and co-authors show that linearized attention cannot reach the kernel regime at any practical width. Softmax attention resists exact kernel study, so linearized forms serve as the tractable stand-in. Even that stand-in fails to settle to its NTK limit at sane widths. The attention step cubes the condition number of the data Gram matrix, they show. Needed width then grows past any feasible size for image data. They name the effect influence malleability, which also raises data attack risk.

Catalog record

CREATORJose Marie Antonio Miñoza, Paulo Mario P. Medina, Sebastian C. Ibañez
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE87615
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.13085

Browse the full library · About the Tower of Babel