Reproduce “SteerAct: Unlocking Action Models through Few-Shot Temporal Steering” 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
Kaichen Zhou, Zhiyan Li, Chongkai Gao, Yuhan Wang, Wenlong Huang, Maggie Wang, Fangneng Zhan, Manling Li and 6 more
SteerAct adapts a frozen robot action model using outcome-labeled successful and failed rollouts, without updating its weights during steering. A pretrained sparse autoencoder (SAE) exposes feature activity, whose temporal patterns select one feature offline and determine online when and how to intervene. On RoboTwin 2.0 with π₀.₅, mean success reaches 77.7%, versus 69.0% for Base, but gains are not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.