World models must capture structure that matters for prediction and intervention across very different physical and biological systems. JEPA-Anything argues that a common learning principle—orthogonal predictive factorization—lets a single predictive design allocate capacity to complementary latent factors rather than compressing all predictable content into a monolithic target.
Key Findings
- Orthogonal predictive factorization (OPF) decomposes latent targets via learned basis matrices and predicts each component with a dedicated branch; predicted components are synthesized back into a usable latent state.
- Evaluated across seven domains (vision, single-cell biology, clinical trajectories, control, molecular dynamics, physical fields, weather), the method improves reported metrics on 10 matched dynamics tasks and reduces single-intervention prediction error on Interventional Pong by 34.8% compared to JEPA baselines.
- Achieves the lowest one-step and 100-step molecular rollout errors among compared methods across four molecular systems, and a JEPA-nominated biological intervention received experimental support in cell co-cultures, organoids, tumor fragments, and mice.
- Learned latent orbital modes recovered Keplerian scaling with a fitted slope of -1.4991, illustrating that factorized latents can reveal physically meaningful structure.
Method overview
- Core mechanism: learn basis matrices that analyze each target latent into multiple components; assign each component a prediction branch; enforce within-factor and cross-factor orthogonality, factor-activity regularization, and online variance regularization to avoid collapse and redundancy.
- The predictive architecture is agnostic to observation geometry: adapters define context/target construction per application, while orthogonal factorization, branch prediction, and state synthesis remain shared.
Who it's for and tradeoffs
- Great fit if you research representation learning, generalizable world models, cross-domain dynamics, or biologically grounded interventions and need a single predictive principle applicable across modalities.
- Look elsewhere if you need minimal-compute, out-of-the-box solutions for a single narrow domain (the method adds modeling complexity and training regularizers) or if you require production-ready deployment pipelines rather than research code and reproducible experiments.
Implementation note: code is provided alongside the paper for reproducibility and adapting the adapter/context design to new domains.