JEPA-Anything brings world models to every domain

JEPA-Anything is an open-source, domain-agnostic framework that helps you build predictive world models across biology, physics, and clinical domains.

Whether you are tracking single-cell dynamics, molecular interactions, physical fields, or patient trajectories, the fundamental problem is identical: inferring unobserved future states from present context. JEPA-Anything provides a shared predictive architecture to solve this across diverse disciplines.

How it works

Standard Joint-Embedding Predictive Architectures (JEPAs) typically treat a target latent representation as a monolithic block. When dealing with complex scientific data, guessing the entire state at once often fails.

JEPA-Anything introduces Orthogonal Predictive Factorization (OPF). An encoder first maps raw observations into a latent world state. OPF then factors this state into distinct, complementary predictive groups, routing each through a dedicated pathway before recombining them into a complete representation.

Downstream tasks can take this recombined state for intervention prediction, planning, scientific analysis, or multi-step rollouts. Crucially, each scientific domain keeps its own observation adapters and encoders, while sharing the underlying predictive engine and latent state interface.

Setup and limitations

You can clone the project directly from GitHub under Gen-Verse/JEPA-Anything. It requires Python 3.10 or newer. The core library runs on PyTorch, while the task-design tooling relies entirely on the Python standard library.

After setting up a virtual environment and installing the core package, you can run an included synthetic linear dynamics recipe to verify interface consistency. The repo also includes automated validation and scaffold generation scripts that turn JSON task specifications into starter code and contract tests.

Keep in mind that this release is a foundational toolkit rather than a ready-to-run scientific model. The repository contains the core logic and synthetic sanity checks, but does not ship with pre-trained weights for the visual, single-cell, or clinical benchmarks shown in its research paper. Translating factors into wet-lab biological insights still requires training your own models on your own datasets.

Why it matters

We think the smart move here is treating domain commonality as an architectural interface rather than forcing a single massive model onto every scientific field.

For computational biologists and researchers building AI world models, JEPA-Anything offers an elegant scaffold that avoids reinventing state-space decomposition from scratch. But without pre-trained checkpoints, the barrier to entry remains high. It is an engineering blueprint for dedicated research labs rather than an off-the-shelf tool for weekend hobbyists.