Same Geometry, Opposite Noise: Transformer Magnitude Representations Lack Scalar Variability
Jon-Paul Cacioli · 2026
Read Same Geometry, Opposite Noise: Transformer Magnitude Representations Lack Scalar Variability by Jon-Paul Cacioli (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-2604-04469.txt.
Description
Cacioli tests whether transformer models show scalar variability, where noise grows with magnitude. Gauging hidden-state spread across carriers for 26 values in three 7-8B models, he finds the reverse: variability falls as magnitude rises. The scaling exponent sits near negative 0.19, with no primary layer showing a positive exponent. The anti-scalar pattern runs three to five times stronger along the size axis than on other axes. The findings show distributional learning alone cannot produce scalar spread: they copy log-squeezed size shape but not the noise mark of biology.
Catalog record
| CREATOR | Jon-Paul Cacioli |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 12249 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.04469