Beyond Domain-Specific World Models: JEPA-Anything Uses One Recipe Across Seven Fields
By Asif Razzaq โ October 5, 2026
Researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford, and Princeton have released JEPA-Anything, a domain-agnostic framework for building world models. Rather than designing a separate predictive model for each field, the framework applies a single shared learning recipe to radically different systems. It extends joint-embedding predictive architectures (JEPAs) with a method called Orthogonal Predictive Factorization (OPF).
The research team evaluated the approach across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather.
What Problem Does JEPA-Anything Solve?
A standard JEPA, such as I-JEPA or V-JEPA 2, relies on a context encoder, an EMA target encoder, and a single predictor. That predictor produces one monolithic target embedding.
The researchers describe this as a capacity-allocation problem: high-variance structure dominates the learned representation, while weaker modes receive conflicting gradients and are effectively crowded out.
The Core Innovation: Orthogonal Predictive Factorization
OPF addresses the capacity-allocation problem by decomposing the prediction target into orthogonal components. Instead of forcing one predictor to capture every mode of variation at once, JEPA-Anything assigns dedicated predictive capacity to each factor. This keeps dominant structure from overwhelming subtler signals and allows the same learning recipe to transfer across domains without architectural redesign.
A Single Recipe, Seven Domains
The framework was tested across:
- Vision
- Biology
- Clinical trajectories
- Control
- Molecular dynamics
- Physical fields
- Weather
This breadth is notable. Domain-specific world models typically require bespoke architectures, loss functions, and tuning. JEPA-Anything suggests that a shared predictive factorization may generalize across domains that differ widely in data type, scale, and dynamics.
Why It Matters for 2026 AI Infrastructure
As world models move from research demos into production pipelines for robotics, simulation, and scientific discovery, the cost of maintaining separate architectures per domain becomes a major bottleneck. A domain-agnostic recipe lowers engineering overhead, simplifies transfer learning, and supports unified infrastructure for multi-domain training. JEPA-Anything is an early signal that predictive world models may follow the same consolidation path seen in foundation models for language and vision.
via MarkTechPost