Reproduce “SteerTrace: Tracing How Target Attacks Steer World Action Models” 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
Yukun Dai, Tianshi Wang, Fengling Li, Jingjing Li, Xin Li, Lei Zhu
SteerTrace targets actions in world action models with one fixed 4/255 image perturbation shared across frames, tasks, and camera views. It uses white-box training and interventions to trace the targeted effect. In FastWAM-Joint on LIBERO, the preserve attack’s imagined future was closer to clean execution in 93% of 200 states, despite a 13.3 cm end-effector displacement; this result is limited to that setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.