Reproduce “STAR: Steering Action Representations during Inference” 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
Guangyu Zhao, Yongchang Zhang, Haoxuan Ru, Kaichen He, Haoran Wen, Kun Zhan, Zeyu Zhang, Yin Li and 1 more
STAR smooths late-layer residual updates in vision-language-action models, adjusting the carry using recent robot motion without changing model weights. The authors measure object awareness by comparing action hidden states with the task object visible and masked. On π₀.₅’s LIBERO-PRO benchmark, STAR raises average success from 48.2% to 50.1%, though several evaluation splits decline; the probe indicates sensitivity, not its cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.