Latent Structure Emergence in Diffusion Models via Confidence-Based Filtering

Wei Wei, Yizhou Zeng, Kuntian Chen, Sophie Langer, Mariia Seleznova, Hung-Hsu Chou · 2026

Read Latent Structure Emergence in Diffusion Models via Confidence-Based Filtering by Wei Wei, Yizhou Zeng, Kuntian Chen, Sophie Langer, Mariia Seleznova, Hung-Hsu Chou (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-2602-06155.txt.

Description

Picture-making diffusion models build an image step by step from a random seed, and Kuntian Chen and 5 co-authors ask when the final meaning becomes readable along the way. They call this semantic accessibility. With a fixed sampling method, each seed fixes one path and one final image.

The team trained simple probe classifiers at several points along the path to guess the final label or trait. On four image data sets, the label could be guessed above chance even from the starting noise. Paths ending in confident classifications carried much more early meaning than shaky ones. The measures give a concrete timeline for when sense emerges during generation.

Catalog record

CREATORWei Wei, Yizhou Zeng, Kuntian Chen, Sophie Langer, Mariia Seleznova, Hung-Hsu Chou
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06155

Browse the full library · About the Tower of Babel