JEPA-Anything

"JEPA-Anything: Learning Predictive Models across Different Worlds", arXiv 2609.20800, submitted 2026-09-17, read on alphaXiv on 2026-09-22. Eighteen authors, listed on the page as Taoyong Cui, Zhongyao Wang, Xinyue Xu, Weiyang Liu, Zhaochen Yu, Yuying Zhang, Qiang Gao, Mengyue Yang, Wanli Ouyang, Pheng Ann Heng, Yingcheng Wu, Zhenfei Yin, Ling Yang and five more; the affiliation shown is Phi AI Labs. Anselm's pick: "a perfect example of the kinds of topics I would like to capture".

What it proposes, per the abstract: predictive models are domain-specific, so the paper extends joint-embedding predictive architectures with what it calls orthogonal predictive factorization, decomposing latent targets into complementary, non-overlapping factors learned through dedicated pathways and reconstructed by a shared predictor. It is tested across seven domains, including ten matched dynamics tasks, forecasting of more than a thousand clinical events, and hundred-step molecular rollouts on four systems.

Results the abstract reports: single-intervention prediction error on Interventional Pong down 34.8% against a matched JEPA baseline; gains on all ten dynamics tasks; the lowest one-step and hundred-step molecular errors on all four systems; a biological intervention nominated by factor analysis with experimental support in organoids, tumour fragments and mice; and latent orbital modes recovering Kepler's third law with a fitted slope of −1.4991 against the exact −1.5.

Why it is here: the desk's models are hand-built, one per question. A learned world model that carries the same principle from weather to molecules is the other end of that spectrum, and the Kepler result is the kind of check a reader can hold in their head. Not read: the full paper; the summary above is the abstract's own claims, not an assessment of them.