JEPA-Anything: One World Model From Physics to Biology
A research team has taken the JEPA architecture first proposed by Yann LeCun and turned it into a world model that runs across seven very different fields. Along the way, the model pointed to a liver cancer treatment candidate, and the team tested that candidate in the lab.
The project is led by PhAI Labs, with collaborators from Stanford, Oxford and Princeton. The model is called JEPA-Anything.
One principle instead of many models
World models try to predict how a system will change over time. That system might be a robot, a molecule or a patient's health. So far, each domain has usually needed its own purpose-built model. The PhAI Labs team wants to show that one shared principle can cover all of them.
JEPA stands for Joint-Embedding Predictive Architecture. These models do not rebuild raw data such as pixels. They predict an abstract summary of a missing or future state and ignore details that don't matter. The authors argue that the standard version has a flaw: it pushes everything into a single prediction, so simple patterns can drown out harder ones.
Splitting the prediction
JEPA-Anything breaks the predicted state into several parts. Each part gets its own prediction module. An extra constraint pushes these modules to learn different aspects of the data instead of all learning the same thing. The model then combines their partial predictions into one full picture.
The researchers do not tell the model what each part should mean. Those roles emerge during training. The only thing they adjust from field to field is how the data is prepared.
Where it helped, and where it didn't
The team compared JEPA-Anything with a standard JEPA that used the same architecture, the same data and identical training conditions.
- Dynamic systems: In a simplified Pong environment with targeted interventions, prediction error fell by 35 percent. For intervention combinations never seen in training, it fell by 13 percent.
- Physics and weather: The authors report that the model beat the baseline across ten test tasks spanning physics, robotics and weather forecasting. On the Burgers equation, a common fluid dynamics benchmark, error dropped by nearly half in a separate test. Over 50 prediction steps the lead held but shrank to about three percent.
- Molecular simulation: It scored best on water, quartz, acetaminophen and benzene, even after 100 steps.
- Biology and medicine: It assigned cell types in single-cell data more reliably and predicted more than 1,000 possible disease events in clinical data slightly better.
- Images and robots: The gap on image tasks was small. In locomotion planning for simulated walking robots, it won in two of three environments and lost in the third.
From partial predictions to a lab test
The most ambitious result comes from liver cancer research. The team examined the partial predictions a model had learned from gene activity, protein levels and CRISPR screens. The top candidate combined IL-18, a signaling molecule that activates immune cells, with blocking CD73, an enzyme tumors use to suppress nearby immune responses.
They tested the pairing on liver cancer cells co-cultured with immune cells, on organoids and tumor tissue from three patients each, and in mice. In organoids and tissue samples, the combination killed more tumor cells than either component alone, and T cells and natural killer cells showed stronger activation. The study does not show whether this could become a real therapy.
In a second case, the model trained on simulated orbits without any physical quantities. Its learned patterns closely matched Kepler's third law: the law sets the exponent at minus 1.5, and the model landed on minus 1.4991. The team evaluated only one training run, choosing the one with the lowest error.
The authors add a caveat. A clean split between learned parts does not prove they capture real cause and effect. Code and models are publicly available.
Our Take
This paper fits into a crowded moment for world models. Meta released the V-JEPA 2 video model in June 2025, LeCun and Randall Balestriero published LeJEPA in November 2025, and LeCun's startup AMI Labs raised over a billion dollars in March 2026. Investors are betting heavily too, from General Intuition's world model raise to AMD's purchase of World Labs.
What stands out here is breadth, not one big score. Most gains are moderate, some are slim, and the robotics results are mixed. That suggests the decomposition idea is a useful general tweak rather than a breakthrough in any single domain.
The cancer result deserves cautious attention. Google Deepmind's C2S-Scale 27B already saw a lab-confirmed cancer hypothesis in October 2025, and its Co-Scientist system now plans experiments and runs lab equipment. The JEPA-Anything team wants agents that propose and rank experiments, then feed results back in, much like efforts to turn language models into lab tools. It is worth watching whether independent groups reproduce these results with the released code, and whether the Kepler finding holds beyond a single hand-picked run.