InfoTok: Regulating Information Flow for Capacity-Constrained Shared Visual Tokenization in Unified MLLMs

Lv Tang, Tianyi Zheng, Bo Li, Xingyu Li · 2026

Read InfoTok: Regulating Information Flow for Capacity-Constrained Shared Visual Tokenization in Unified MLLMs by Lv Tang, Tianyi Zheng, Bo Li, Xingyu Li (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-2602-01554.txt.

Description

Lv Tang and co-authors propose InfoTok, information-theoretic regularization for size-limited tokenization in unified models. A shared visual tokenizer maps high-dimensional images onto a limited token budget, yet designs lack rules for what to preserve for abstract ideas and detail. Viewing the tokenizer as a bounded learner, InfoTok imposes mutual-information constraints enforcing a compression-relevance trade-off and cross-modal consistency. Dependence estimators, including variational bottleneck and Hilbert-Schmidt criterion variants, instantiate it. In three unified models without extra training data, it improved understanding and generation.

Catalog record

CREATORLv Tang, Tianyi Zheng, Bo Li, Xingyu Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41734
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.01554

Browse the full library · About the Tower of Babel