The AI world talks a lot about world models, yet most predictive stacks remain siloed: one architecture for vision, another for molecular dynamics, yet another for clinical trajectories. JEPA-Anything (arXiv 2609.20800, published Sep 17) asks a more fundamental question — can cells, molecules, physical fields, and weather share a single predictive learning principle?

The core idea: factor the latent state orthogonally

The proposed answer is Orthogonal Predictive Factorization (OPF). Built on joint-embedding predictive architectures (JEPA), it replaces a monolithic latent target with a set of complementary predictive factors: each factor is learned through a dedicated pathway, then recombined within a shared predictive design into a complete state for downstream readout, intervention prediction, planning, and multi-step rollout.

The README puts it plainly: different domains keep their own observations, context–target construction, and encoders, while sharing the predictive core and the latent world-state interface. The commonality lives at the principle level — domains are not required to share a single encoder or one set of weights. That design choice is what makes the framework domain-agnostic: porting to a new domain means writing an observation adapter, not rebuilding the predictor.

The seven-domain scorecard

Evaluation spans vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather, across representation learning, intervention prediction, out-of-distribution generalization, and long-horizon dynamics — including 10 matched dynamics tasks, forecasting of over 1,000 clinical events, and 100-step molecular rollouts across four systems. The key numbers:

  • Against matched JEPA baselines, reported metrics improve on all 10 dynamics tasks.
  • On Interventional Pong, single-intervention prediction error drops by 34.8%.
  • One-step and 100-step molecular errors are the lowest among compared methods in all four systems.
  • Latent orbital modes recover the Keplerian scaling exponent with a fitted slope of -1.4991.

The last two matter most. The Keplerian fit suggests the learned factors align with physical law rather than merely fitting numbers. And a factor-nominated biological intervention received experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and mice — a full loop from predictive model back to bench experiment, which is what separates this paper from ordinary benchmark-chasing work.

Cold water: no trained weights in the repo

The GitHub repository's "Repository scope" section is explicit: the open-source release contains the reusable core, task-design tools, and a synthetic structural example — not the domain datasets or trained model weights from the paper. The checkpoint manifest records an untrained example with no performance claims. In other words, reproducing that 34.8% means assembling datasets and training pipelines yourself. As of writing, the repo sits at 62 stars, 6 commits, Apache-2.0, with 40 upvotes on the Hugging Face paper page — early-stage community traction.

So what

For AI-for-Science practitioners, the value is not "another world model" but a testable blueprint for cross-domain predictive learning: the factorized OPF interface is modular, and the cost of adding a domain is pushed down to the observation-adapter layer. For everyone else, it points at a direction — while the industry competes on per-domain generation quality, the universality claim of "one principle, many worlds" is ultimately arbitrated by wet experiments and physical laws, not leaderboards.

Paper: https://arxiv.org/abs/2609.20800
Code: https://github.com/Gen-Verse/JEPA-Anything