Research team releases JEPA-Anything, a shared approach to world models across seven domainsMachine translation

JEPA-Anything splits a prediction target into orthogonal factors, gives each factor its own predictor, and recombines them into a complete latent state. The team tested it across seven domains; reported results favored it over matched standard JEPA baselines on all 10 dynamics tasks, but planning results varied by environment, with standard JEPA doing better on Hopper.