Reproduce “DWM: Separating World Effects from Actions in Latent 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

Yi-Ge Zhang, Tianqi Du, Qi Zhang, Yisen Wang

DWM separates passive world motion from action-driven change during training for latent world models. It adds a world head and defines the action-driven component as the residual needed to reconstruct the original prediction. On PushT-W, CEM planning success rises from 32.0% to 44.0%. Across three altered simulated tasks, DWM reports a mean 13.1 percentage-point improvement, limited to these controlled settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.