Reproduce “Controlling Action-Head Co-Adaptation to VLM Representation Anisotropy” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ruijie Zhang, Jiayin Zheng, Yequan Zhao, Liyan Tan, Ziyue Liu, Zhengyang Wang, Kunpeng Song, Nathan Zhao and 3 more

AdaIE reweights vision-language model features passed to a robot action model according to the action model’s changing channel sensitivity, without adding learnable parameters. In controlled tests, equalization improved results on two tasks designated mismatched, while aligned tasks showed small differences. On RoboCasa-GR1, it raised average success from 42.36% to 47.64% across 24 tasks, but benefits were not uniform.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.