Reproduce “World Models Favor Learned Priors over Visual Evidence” 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

Han Xiao, Zhiyong Bao, Xiaotong Liu, Yanhao Chen, Qingqiang Wu

RIFT predicts trajectories under physical rules inferred from calibration interactions, transferring the selected rule to a query with a different initial state. It combines observed motion with Rule Atlas references and an adjustable training-frequency preference. On TailWorld’s 2D physical-rule-transfer test, its ADE is 0.0895 m versus MaNGO’s 0.1767 m, an overall benchmark result rather than a rare-rule-only score.

Summary by alphaXiv.

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