Zero-Shot Quantization via Weight-Space Arithmetic

Daniele Solombrino, Antonio Andrea Gargiulo, Adrian Robert Minut, Luca Zhou, Alessandro Zirilli, Emanuele Rodolà · 2026

Read Zero-Shot Quantization via Weight-Space Arithmetic by Daniele Solombrino, Antonio Andrea Gargiulo, Adrian Robert Minut, Luca Zhou, Alessandro Zirilli, Emanuele Rodolà (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-2604-03420.txt.

Description

Solombrino and colleagues show that toughness to post-training quantization is a shared path in weight space. A quantization vector pulled from a donor task by simple weight math can patch a target model, lifting post-quantization scores up to 60 points in a 3-bit setting without target-side training. Across four ViT scales and 22 image tasks, donor vectors often help even when tasks differ a lot.

Catalog record

CREATORDaniele Solombrino, Antonio Andrea Gargiulo, Adrian Robert Minut, Luca Zhou, Alessandro Zirilli, Emanuele Rodolà
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77755
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.03420

Browse the full library · About the Tower of Babel