Reproduce “Spoofing the Latent State: Imperceptible Goal-Image Attacks Derail Planning in LeWorldModel” 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
Kunal Bhosikar, Nicole Meng, Charu Sharma, Yingjie Lao
LeWorldModel steers its unchanged planner by altering a goal image’s latent representation, causing it to pursue a different block pose. It encodes camera and goal images, predicts action effects, and uses the Cross-Entropy Method to select actions. At ε=4/255, goal-image attacks reduce Push-T success from 87.0% to 32.3% in paired episodes. The study is limited to white-box digital attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.