Beyond Domain-Specific World Models: JEPA-Anything Uses 1 Recipe for 7 Fields
One recipe, 7 domains, 10/10 wins. JEPA-Anything factors latent space to beat domain-specific baselines across physics, vision, and control tasks.
Why it matters
A factorized JEPA architecture generalizes across 7 domains without retraining, beating matched baselines on dynamics prediction and intervention tasks. Relevant to practitioners building world models and foundation model researchers exploring composable latent representations.
The key facts
10 to knowTested across 7 domains (dynamics prediction benchmark)
Beat all 10 matched JEPA baselines on dynamics tasks
34.8% reduction in Interventional Pong intervention error
4 orthogonal factors in latent space, each with dedicated predictor
Single recipe (no domain-specific tuning)
JEPA-Anything splits a single latent target into 4 orthogonal factors, each with its own predictor
Tested across 7 domains
Beat matched JEPA baselines on all 10 dynamics tasks
Cut Interventional Pong intervention error by 34.8%
Source: MarkTechPost (research summary, not independent validation)
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: JEPA-Anything splits a JEPA's single latent target into 4 orthogonal factors, each with its own predictor. Tested across 7 domains, it beat matched JEPA baselines on all 10 dynamics tasks and cut Interventional Pong intervention error by 34.8%.