T-GEMs: Text-Guided Exit Modules for decreasing clip image encoder

Alberto Presta, Grzegorz Stefanski, Michal Byra, Krzysztof Arendt · 2026

Read T-GEMs: Text-Guided Exit Modules for decreasing clip image encoder by Alberto Presta, Grzegorz Stefanski, Michal Byra, Krzysztof Arendt (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-2605-17499.txt.

Description

Presta and colleagues introduce text-guided exit modules for shrinking CLIP image encoders, letting easy inputs leave the encoder early. They study how semantic content spreads across intermediate layers and whether text descriptions can reveal it. The resulting modules plus a rate-based regularizer control encoder usage cost while preserving cross-modal understanding. The approach cuts compute for image-text pairs without surrendering matching performance.

Catalog record

CREATORAlberto Presta, Grzegorz Stefanski, Michal Byra, Krzysztof Arendt
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21045
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.17499

Browse the full library · About the Tower of Babel