Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference

Yudong Liu, Yuan Li, Zijia Tang, Yuxi Zheng, Yueqian Lin, Qinsi Wang, Shuangjun Liu, Shuai Zhang, Taotao Jing, Dashan Gao, Ning Bi, Jingwei Sun, Yiran Chen, [0.5em] Duke University, [0.3em] {yudong.liu, zijia.tang, yuxi.zheng, yueqian.lin, qinsi.wang, yiran.chen, hai.li}@duke.edu · 2026

Read Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference by Yudong Liu, Yuan Li, Zijia Tang, Yuxi Zheng, Yueqian Lin, Qinsi Wang, Shuangjun Liu, Shuai Zhang, Taotao Jing, Dashan Gao, Ning Bi, Jingwei Sun, Yiran Chen, [0.5em] Duke University, [0.3em] {yudong.liu, zijia.tang, yuxi.zheng, yueqian.lin, qinsi.wang, yiran.chen, hai.li}@duke.edu (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-2605-02739.txt.

Description

Yudong Liu and fourteen co-authors present Latent Bridge, a small model that makes dual-system robot controllers run faster. In such systems, a large vision-language backbone must run at every control step, yet its features change little between steps. That redundancy is the cost Latent Bridge removes, by predicting how the backbone's output will change.

The action head then runs on predicted outputs, and the costly backbone is called only from time to time. The bridge is shown on two distinct robot systems, one in feature space and one in cache space, so the idea is not tied to a single design. A task-agnostic training pipeline moves across benchmarks unchanged. Tests kept nearly all task performance while cutting backbone calls by half or more, for a clear net speedup.

Catalog record

CREATORYudong Liu, Yuan Li, Zijia Tang, Yuxi Zheng, Yueqian Lin, Qinsi Wang, Shuangjun Liu, Shuai Zhang, Taotao Jing, Dashan Gao, Ning Bi, Jingwei Sun, Yiran Chen, [0.5em] Duke University, [0.3em] {yudong.liu, zijia.tang, yuxi.zheng, yueqian.lin, qinsi.wang, yiran.chen, hai.li}@duke.edu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.02739

Browse the full library · About the Tower of Babel