Blink of an eye: a simple theory for feature localization in generative models

Marvin Li, Aayush Karan, Sitan Chen · 2025

Read Blink of an eye: a simple theory for feature localization in generative models by Marvin Li, Aayush Karan, Sitan Chen (2025) 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-2502-00921.txt.

Description

Marvin Li and co-authors offer a simple theory for why AI generators suddenly settle on key features, in this 2025 paper. Language models can veer off course in a blink, and image models decide core features within narrow windows of their process. Past explanations leaned on heavy assumptions about the data.

The new theory uses a general sampling frame that fits both text and image models. It shows the effect arises naturally as generation narrows onto a sub-group of the modeled distribution. The math needs only basic tools and sharpens older bounds even for image models. Tests back the prediction, and the authors note a link to sudden shifts seen in statistics.

Catalog record

CREATORMarvin Li, Aayush Karan, Sitan Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE133 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-00921

Browse the full library · About the Tower of Babel