AnalysisAI ModelsSeptember 18, 2026

JEPA-Anything applies one world-model recipe across 7 domains

Read original source →arxiv.org

Researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford and Princeton released JEPA-Anything, which extends joint-embedding predictive architectures with Orthogonal Predictive Factorization. OPF splits the latent target of width d into K learned subspaces (K=4 in most experiments), each with its own predictor.

2 sources

More stories today

Open the live feed