Reproduce “WorldFoundry: Towards Unified Infrastructure for World Intelligence” 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
Juanxi Tian, Jiadong Hou, Baiang Li, Xin Lin, Zhengbo Jiao, Weijia Zhang, Yeying Jin, Truong Nguyen and 8 more
WorldFoundry defines a world model by a loop in which a system maintains state from interaction history, predicts action effects, and revises state from new observations, then provides contracts for running and evaluating systems while preserving native semantics. Its catalog lists 270 model entries, 74 benchmarks, and 62 metric entries; one OpenVLA task succeeds in 179 steps, while two reach 520-step limits.
Key insights
- Realistic-looking outputs alone do not establish world modeling; evidence should test whether actions have persistent consequences and whether later observations revise predictions.
- A shared orchestration and record-keeping layer can connect heterogeneous systems without forcing them into one representation or interface.
- Catalog inclusion and runner verification are bounded claims, not proof of universal availability, full benchmark reproduction, or world-model capability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.