Reproduce “Plausible Videos Can Lead to Poor Decisions: Diagnosing and CURE-ing World Models for Planning” 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
Raphael C. Kim, Michele Santacatterina, Jingsen Zhu, Ramin Zabih
CURE checks whether a video world model’s action rankings are trustworthy by estimating prediction error and training-data coverage. Its certificate guides whether to use a model, collect data, or run more rollouts. In Language-Table simulation, certified decisions had mean regret 0.06, versus 0.38 for withheld decisions; selecting the same share by PSNR performed worse. This result is limited to the tested setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.