Reproduce “Structural-Test Geometry as an Inductive Bias for Partially Observable World Models” 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
Todd Y. Zhou
QP trains a partially observed world model to preserve distinctions revealed by a supplied family of simulator tests. It forks simulations from posterior hidden states, measures outcome differences between histories, and aligns their test-derived distances with native latent distances. Across six fixed tasks, QP’s held-out structural-shift planning success exceeds MD by 26.1 percentage points and BD-S by 6.4, within this task suite and known probe families.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.