SMA: Submodular Modality Aligner For Data Efficient Multimodal Learning

Truong Pham, Anay Majee, Rishabh Iyer · 2026

Read SMA: Submodular Modality Aligner For Data Efficient Multimodal Learning by Truong Pham, Anay Majee, Rishabh Iyer (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-2605-12872.txt.

Description

Pham, Majee and Iyer propose a set-based approach to aligning images and text in multimodal models, where scarce paired data limits progress. Their Submodular Modality Aligner treats several augmentations and descriptions of one entity as a set, drawing on richer cross-modal structure. Built on submodular mutual information, it boosts shared information across modalities while cutting cross-modal divergence. On 14 benchmark tasks it gains steadily in low-data settings, using only tens of thousands of samples.

Catalog record

CREATORTruong Pham, Anay Majee, Rishabh Iyer
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39905
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.12872

Browse the full library · About the Tower of Babel