Reproduce “World Modeling through Spectral Alignment” 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
Holger Molin, William Peng, Marco Bagatella, Randall Balestriero
SpecWM trains visual world models by making pairwise observation similarities an explicit target for latent representations, including a temporal target based only on distance between observations in a trajectory. Its action-conditioned predictor supports goal-image planning with CEM. On OGBench Scene, temporal SpecWM achieves 77% success versus 49% for LeWorldModel; gains are benchmark-specific, not evidence of faster or cheaper planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.