Reproduce “world2vec: Mechanistic world models through latent dynamics learning” 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
Tobias Schmidt, Rodrigo Gonzalez Laiz, Krisztián Szuppinger, Steffen Schneider
world2vec learns compact latent representations and action-conditioned dynamics from observations, using multi-step prediction matching to train them together without reward information. It advances latent states with a learned vector field and Euler integration. In pixel-based control, it succeeds on Pendulum and Reacher but not CartPole; its scientific equation-recovery demonstration uses ground-truth alignment and does not establish general recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.