JEPA-Anything Applies One World-Model Recipe Across Domains

JEPA-Anything splits latent targets into four predictors and beats JEPA baselines across seven domains, sources say.

MarkTechPost and The Decoder both describe JEPA-Anything, a world-model recipe that splits a single JEPA latent target into four orthogonal factors, each with its own predictor, and then reassembles them. MarkTechPost says it beat matched JEPA baselines on all 10 dynamics tasks across seven domains and cut Interventional Pong intervention error by 34.8%; The Decoder rounds that Pong drop to 35% and adds that Burgers-equation error nearly halved, with smaller gains on robotics, weather, images, and clinical prediction. The Decoder attributes the work to a PhAI Labs team with Stanford, Oxford, and Princeton collaborators, calls it an extension of Yann LeCun’s Joint-Embedding Predictive Architecture, and says code and models are public. Only The Decoder reports that biological factors led the team to test IL-18 plus CD73 blockade, which killed more liver-cancer cells than either treatment alone in organoid, patient-tissue, and mouse tests, while stating this does not establish a therapy and that experiment-design reliability remains unproven. MarkTechPost does not mention the biology tests, Burgers results, affiliations, or a code release.

• JEPA-Anything splits one JEPA latent target into four orthogonal factors, each with its own predictor, then reassembles them.
• MarkTechPost: it beat matched JEPA baselines on all 10 dynamics tasks across seven domains.
• Interventional Pong intervention error fell 34.8% (MarkTechPost) versus about 35% (The Decoder).
• The Decoder: Burgers-equation error nearly halved, with smaller gains on images, robots, weather, and clinical prediction.
• The Decoder attributes the work to PhAI Labs with Stanford, Oxford, and Princeton collaborators, as an extension of Yann LeCun’s JEPA, and says code and models are public.
• The Decoder only: IL-18 plus CD73 blockade killed more liver-cancer cells than either alone in organoid, patient-tissue, and mouse tests; it does not establish a therapy, and experiment-design reliability is unproven.