Alignment or Integration? Rethinking Multimodal Fusion in DNA-language Foundation Models

Yanan Li, Christina Yi Jin, Yuan Jin, Manli Luo, Tie Xu, Shuai Jiao, Wei He, Qing Zhang · 2026

Read Alignment or Integration? Rethinking Multimodal Fusion in DNA-language Foundation Models by Yanan Li, Christina Yi Jin, Yuan Jin, Manli Luo, Tie Xu, Shuai Jiao, Wei He, Qing Zhang (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-2602-12286.txt.

Description

Combining DNA data with plain language in one model sounds simple, yet most designs bolt separate encoders together and then struggle to reconcile them. Yanan Li led this work with seven co-authors, tracing that struggle to a geometric gap between modalities, which forces the language model to spend capacity on alignment rather than reasoning.

The team shows that a native design with one shared vocabulary keeps that gap at zero across hidden layers. Their One Tokenizer architecture maps all inputs straight into shared token space, and tests on a DNA-text testbed beat encoder-based modular rivals. For deep biological reasoning, the paper argues, integration should be built into the model, not bolted on after.

Catalog record

CREATORYanan Li, Christina Yi Jin, Yuan Jin, Manli Luo, Tie Xu, Shuai Jiao, Wei He, Qing Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-12286

Browse the full library · About the Tower of Babel