JEPA-Anything Uses One Recipe to Model Seven Different Worlds
Researchers build a domain-agnostic world model that works across vision, biology, medicine, physics, and weather without redesigning for each field.

One Model, Seven Worlds
A team of researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford and Princeton has released JEPA-Anything, a framework that learns predictive models across seven wildly different fields using a single shared recipe.
The idea behind world models — AIs that predict how the world changes over time — has traditionally been domain-specific. You might train one model for video, another for robotics, and yet another for molecular simulations. JEPA-Anything skips that. It applies the same core method to vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather.
The team credits this flexibility to a technique called Orthogonal Predictive Factorization, or OPF, which they say solves a key weakness in standard JEPA systems.
The Factorization Fix
A standard JEPA, such as I-JEPA or V-JEPA 2, predicts a single unified target. The researchers call this a capacity-allocation problem: high-variance features dominate while weaker modes receive conflicting training signals. In practice, this means the model becomes good at predicting some things but poor at others.
OPF solves this by splitting the latent target into K separate subspaces — most experiments use four — each with its own dedicated predictor. After prediction, the factors are recombined mathematically to form a complete latent state. Three built-in rules keep each factor useful and prevent them from overlapping.
On a test called CITRIS Interventional Pong, the orthogonal version of the model achieved a condition number near 1.0, compared to 438.52 for a capacity-matched non-orthogonal model. That means the factors stayed cleanly separate during training.
Results Across Domains
JEPA-Anything improved reported metrics on all ten matched dynamics tasks tested. On Interventional Pong, single-intervention prediction error fell by about 35 percent. On the Burgers equation benchmark, six-step rollout error dropped roughly 45 percent [source 1]. In molecular dynamics, the model achieved the lowest one-step and 100-step errors across four substances including water and benzene.
Beyond pure prediction, the system has produced unexpected scientific insights. When analyzing single-cell biological data, a factor-nominated combination of IL-18 and CD73 blockade was later confirmed through wet-lab experiments in cell cultures, patient-derived organoids, tumor fragments and mice. The model also recovered Kepler's law of planetary motion from latent orbital patterns, with a fitted exponent of -1.4991 against the theoretical value of -1.5 [source 1].
Planning performance was mixed. The system improved results on the Walker2d and HalfCheetah robotics benchmarks but the standard JEPA performed better on Hopper.
Open for the Community
The core code is licensed Apache-2.0, and research checkpoints are available on Hugging Face. The researchers argue that their results show a common factorized predictive principle can connect world modeling with real-world intervention and scientific discovery.
The paper is available on arXiv as JEPA-Anything: Learning Predictive Models across Different Worlds, along with the full codebase and model checkpoints.
Key facts
- JEPA-Anything uses one shared recipe across seven domains instead of custom models per field
- Orthogonal Predictive Factorization splits predictions into separate factors with dedicated predictors
- Single-intervention error on Interventional Pong fell by 34.83 percent versus matched baselines
- Factor analysis nominated a cancer intervention later confirmed in wet-lab experiments
- Core code is Apache-2.0 with checkpoints on Hugging Face
Sources
- Beyond Domain-Specific World Models: JEPA-Anything Uses 1 Recipe for 7 Fields — marktechpost.com
- JEPA-Anything: Learning Predictive Models across Different Worlds
- JEPA-Anything: Learning Predictive Models across Different Worlds
- GitHub - Gen-Verse/JEPA-Anything: JEPA-Anything: Learning Predictive ...
- JEPA-Anything: Single Model Trains Across 7 Domains — aidailypost.com
This story was written before each paragraph carried a source number: every fact in it comes from one of the sources above.
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
How it worksNew method lets robot brains learn from human videos without touching expensive robot dataHerald of AIShixart1985, CC BY 2.0, via Wikimedia Commons
New method lets robot brains learn from human videos without touching expensive robot data
A research team showed how to use everyday human videos to teach robots new tasks, cutting the cost of collecting rare robot experience.
Researchers propose Avatar Intelligence to make AI agents better at handling your daily tasks
A new research framework suggests measuring AI progress by how well personal agents help people pursue their goals with less supervision.
Apps and productsllama.cpp v1.0.0 brings stability to the growing world of local AIHerald of AI
llama.cpp v1.0.0 brings stability to the growing world of local AI
The popular open-source tool that lets anyone run AI models on their own computer has reached version 1.0, making it more reliable for developers building private AI apps.