NoiseGate: Learning Per-Latent Timestep Schedules as Information Gating in World Action Models

Wen Huang, Haoran Sun, Yongjian Guo, Yunxuan Ma, Haoran Li, Jing Long, Zhouying Mo, Zhong Guan, Yucheng Guo, Shuai Di, Junwu Xiong · 2026

Read NoiseGate: Learning Per-Latent Timestep Schedules as Information Gating in World Action Models by Wen Huang, Haoran Sun, Yongjian Guo, Yunxuan Ma, Haoran Li, Jing Long, Zhouying Mo, Zhong Guan, Yucheng Guo, Shuai Di, Junwu Xiong (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-2605-07794.txt.

Description

Wen Huang and ten co-authors improve world action models, robot policies that generate actions together with imagined future video. In current systems every predicted latent frame shares one noise schedule during generation, which implicitly assumes all imagined frames are equally reliable guides for action. The authors reject that prior and treat each frame's schedule as a learnable gate on information.

Their NoiseGate samples per-frame timesteps independently in training and adds a small gating policy network that adjusts each frame's noise level during denoising, trained by task reward rather than hand-designed shapes. Raising a frame's noise mutes its contribution to the action tokens. Built on a joint video-action backbone, NoiseGate delivered consistent gains across diverse random-scene manipulation tasks in RoboTwin.

Catalog record

CREATORWen Huang, Haoran Sun, Yongjian Guo, Yunxuan Ma, Haoran Li, Jing Long, Zhouying Mo, Zhong Guan, Yucheng Guo, Shuai Di, Junwu Xiong
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.07794

Browse the full library · About the Tower of Babel