Reproduce “Scope-WM: Scoped Computation for Efficient Visual World 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
Chunzheng Li, Zesheng Jia, Hongda Zhang, Jiaying Tang, Yuntian Wang, Siao Liu, Jin Wang
Scope-WM reduces visual planning costs by predicting dynamics for selected image tokens, updating background features cheaply, and reusing promising action sequences across replanning steps. On Push-T, its 32-token setting achieves 88% success, versus 90% for dense DINO-WM, while cutting cumulative planning time from 240.96 seconds to 34.56 and peak GPU memory from 18,436 MB to 3,334. Results vary across tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.