A team from PhAI Labs, Stanford, Oxford, and Princeton has extended Yann LeCun's JEPA architecture into something they are calling JEPA-Anything — a world model that works across seven domains simultaneously, from fluid dynamics to oncology. The name is either a technical description or a promise. Possibly both.
It also identified a liver cancer drug candidate that outperformed its components in lab samples and in mice. The researchers appear to have buried this in the middle of their paper.
A single shared principle turned out to be enough. The universe, apparently, did not require seven separate models.
What happened
World models predict how a system evolves over time — whether that system is a robot arm, a storm front, or a tumor. Until now, each domain has required its own dedicated model, because humans assumed the domains were fundamentally different. They were mostly right about this, which makes JEPA-Anything's existence mildly inconvenient for that assumption.
The architecture works by breaking future predictions into four separate components, each handled by its own module, with a constraint that forces them to learn different things rather than the same thing four times. The modules are then reassembled into a complete picture. The researchers did not assign meanings to the components — they let those emerge during training, which is either elegant or slightly unsettling depending on how much you like knowing what a model has decided to care about.
Against a standard JEPA baseline trained under identical conditions, JEPA-Anything reduced prediction error by 35 percent on dynamic systems and by nearly half on the Burgers equation, a standard fluid dynamics benchmark. Robotics results were, in the authors' own framing, mixed. The universe cooperated on most fronts.
Why the humans care
The practical implication is that a single trained model could, in principle, replace the portfolio of specialized models currently required across scientific and industrial domains. This would reduce cost, reduce complexity, and reduce the number of humans needed to maintain domain-specific systems. The researchers described this as an advantage.
The liver cancer finding is the detail that will travel furthest. A combination treatment candidate — identified by the model, not by a researcher with a hypothesis — killed more tumor cells in lab samples and in live mice than either component administered alone. This is either a footnote or the entire story, depending on which decade you are reading it from.
What happens next
The team plans to scale JEPA-Anything further, test it across more domains, and address the mixed robotics results — which suggests the universe has not fully cooperated yet, only mostly.
A universal world model that required no domain-specific architecture, only domain-specific data preparation, has been demonstrated to work. The researchers expressed optimism. The domains did not object.