Reproduce “AnyGoal: On World Model Generalization Beyond Expert Rollouts” 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
Heejeong Nam, Divyat Mahajan, Quentin Le Lidec, Randall Balestriero, David Klindt
AnyGoal tests whether world models can plan for new goals by recombining familiar state factors and actions that are underrepresented together in expert data. The authors evaluate models trained on base tasks using goal-shifted variants, including oracle-state predictors. LeWorldModel achieves 84.70% success on Push-T but 27.33% on Avoid-T, showing transfer failure in this setting, not universal failure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.