A new AI model called JEPA-Anything aims to make world models less specialized. Instead of building a separate model for every domain, the team behind it tested whether one shared approach could work across very different systems, from simulated physics to biological data.
The work builds on Joint-Embedding Predictive Architectures, or JEPA, an approach associated with Yann LeCun. The central idea is not to recreate every raw detail of data, but to predict useful abstract representations of what is missing or what may happen next.
Why JEPA-Anything Matters
World models are designed to estimate how a system changes over time. That system could be a robot moving through an environment, a molecule evolving in a simulation, a weather pattern, or a patient’s future health state.
Until now, the source article says each field has usually needed its own model. JEPA-Anything, introduced by a team led by PhAI Labs with collaborators from Stanford, Oxford, and Princeton, tests a broader claim: the same predictive principle may be useful across many domains.
The researchers evaluated the model across seven very different fields. Those tests included physics, robotics, weather forecasting, single-cell data, clinical data, images, and biological discovery. The results were not uniformly dramatic, but they suggest that the architecture can transfer across problems that normally look unrelated.
How the Model Changes JEPA
Standard JEPA models predict an abstract summary of a target state. That makes them different from systems that try to reconstruct raw pixels or other low-level details. The benefit is focus: the model can ignore noise that is not useful for prediction.
The researchers argue that the standard setup has a weakness. When everything is compressed into a single prediction, easy patterns can dominate, while subtler patterns receive less attention.
JEPA-Anything addresses that by splitting the predicted state into several parts. Each part is handled by a separate prediction module. A constraint encourages the modules to learn different aspects of the target rather than duplicating one another, and the system then combines those partial predictions into a fuller view.
The team does not manually label what each part should represent. Instead, those roles emerge during training. The researchers adjust the way data is prepared for each field, but they do not assign fixed meanings to the separate learned components.
Where the Gains Appeared
The clearest improvements came in dynamic systems. In a simplified Pong environment with targeted interventions, prediction error dropped by 35 percent compared with a standard JEPA using the same architecture, data, and training conditions. For unseen combinations of interventions, the error dropped by 13 percent.
According to the authors, JEPA-Anything beat the baseline across ten test tasks spanning physics, robotics, and weather forecasting. On the Burgers equation, described in the source as a common fluid dynamics benchmark, error fell by nearly half in a separate evaluation. Across 50 prediction steps, the advantage remained but narrowed to about three percent.
The model also performed best in simulations involving water, quartz, acetaminophen, and benzene, including after 100 steps. These results point to one of the model’s most important claims: splitting prediction into multiple learned pieces may help when a system has several interacting factors.
The results were more mixed elsewhere. For single-cell data, the model assigned cell types more reliably. On clinical data, it predicted more than 1,000 possible disease events slightly better. Image tasks showed only a small difference. In simulated walking robot planning, JEPA-Anything won in two of three environments, while the standard model performed better in the third.
The Liver Cancer Experiment
The most concrete biological claim involves liver cancer research. The team analyzed partial predictions learned from biological data, including gene activity, protein levels, and CRISPR screens. From that analysis, the leading candidate combined IL-18 with blockade of CD73.
In the source article, IL-18 is described as a signaling molecule that activates immune cells. CD73 is described as an enzyme tumors use to suppress immune responses nearby.
The researchers tested the combination in several settings: liver cancer cells co-cultured with immune cells, organoids and tumor tissue from three patients each, and mice. In the organoids and tissue samples, the combination killed more tumor cells than either IL-18 or CD73 blockade alone. T cells and natural killer cells also showed stronger activation.
That does not establish that the approach can become a therapy. The study shows a treatment candidate and lab results, not a proven clinical outcome.
Promise, Limits, and the Bigger Race
The model also produced an intriguing physics result. Researchers trained it on simulated orbits without giving it physical quantities. The learned pattern nearly matched Kepler’s third law, which says bodies on larger orbits move much more slowly. The law sets orbital frequency at orbit size to the power of minus 1.5; the model landed on minus 1.4991.
There is an important caveat. The team evaluated one training run and selected the one with the lowest error. More broadly, the authors caution that cleanly separated learned parts do not prove that the model has found real cause-and-effect relationships.
That distinction matters because the long-term ambition is larger than prediction. The team wants AI agents to use JEPA-Anything to propose and rank experiments, then feed results back into the model. Whether the system is reliable enough to guide experiment design remains open.
The work also sits inside a broader push around JEPA and AI-driven scientific discovery. LeCun proposed JEPA in 2022 as an alternative to generative models. In June 2025, Meta released V-JEPA 2, a JEPA video model with 1.2 billion parameters that controlled robotic arms in unfamiliar environments without additional training. In November 2025, LeCun and Randall Balestriero followed with LeJEPA, described as a theoretical foundation for stable training without usual workarounds.
LeCun is now pursuing the approach through AMI Labs, which raised over a billion dollars in March 2026 to build world models. The source also notes related work from Google Deepmind: in October 2025, its Gemma-based C2S-Scale 27B proposed silmitasertib to make tumor cells more visible to the immune system, and experiments with human cell models confirmed the prediction. Google Deepmind’s Co-Scientist now plans experiments and operates lab equipment, though humans still load samples into the machines.
For JEPA-Anything, the immediate takeaway is narrower but still significant. A model built around multiple partial predictions appears to improve performance across several demanding prediction tasks, and it generated a biological hypothesis that researchers could test. The harder question is how far that pattern can be trusted when the model moves from recognizing structure to helping design experiments.