JEPA-Anything: one recipe for world models across 7 domains

Researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford and Princeton released JEPA-Anything, a domain-agnostic framework for building world models. Its Orthogonal Predictive Factorization method splits the latent target into K orthogonal factors with dedicated predictors; it beat matched JEPA baselines on all 10 dynamics tasks, with Interventional Pong single-intervention error down 34.83%.

• OPF splits a width-d latent into K orthogonal factors of width r, typically K=4
• Interventional Pong single-intervention error fell 34.83%
• APEBench Burgers 6-step rollout error dropped about 44.7%
• Core code is Apache-2.0; research checkpoints are on Hugging Face