Reproduce “Inside-JEPA: From Prediction to Relational Grounding in Latent 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
Hossein Resani, Javen Qinfeng Shi
Inside-JEPA adds soft channels for ordered pairs of entities to a predictive state and trains them to recover actions from changes in those channels, without relation labels. On Something-Else, with annotated boxes and a frozen ResNet-18, it reaches .658 matched relational AUROC, versus .593 for prediction alone. The relations are image-plane box proxies, not evidence of physical containment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.