研究者将 LeCun 的 JEPA 扩展为跨物理到生物的通用世界模型

Key Points

Researchers have expanded the JEPA architecture pioneered by Yann LeCun so it works across seven very different fields. The effort also produced a liver cancer treatment candidate that the team tested in the lab.

World models predict how a system will evolve, whether it's a robot, a molecule, or a patient's health. Until now, each domain has typically needed its own model. A team led by PhAI Labs, with collaborators from Stanford, Oxford, and Princeton, wants to show that a single shared principle is enough.

Their paper introduces JEPA-Anything, built on Joint-Embedding Predictive Architectures (JEPA). These models don't reconstruct raw data like pixels but instead predict an abstract summary of a missing or future state, filtering out irrelevant details. The authors see a weakness in the standard approach: everything gets funneled into a single prediction, so easy patterns end up drowning out harder ones.

Multiple specialized predictions instead of one

JEPA-Anything breaks the predicted state into several parts, each handled by its own prediction module. An added constraint pushes the modules to capture different aspects rather than learning the same thing, and the model then reassembles their partial predictions into a complete picture.

The researchers don't assign meanings to the parts, letting those roles emerge during training. Instead, they only change how they prepare the data for each field.

Big gains in dynamics tests, mixed results for robots

The team compared JEPA-Anything against a standard JEPA with the same architecture, trained on the same data under identical conditions. Dynamic systems showed the clearest gains: in a simplified Pong environment with targeted interventions, prediction error dropped by 35 percent. For combinations of interventions the model never saw during training, it fell by 13 percent.

JEPA-Anything consistently beat the baseline across ten test tasks spanning physics, robotics, and weather forecasting, according to the authors. On the Burgers equation, a common fluid dynamics benchmark, error fell by nearly half in a separate evaluation. Over 50 prediction steps the advantage held but shrank to about three percent. The method also scored best in simulations of water, quartz, acetaminophen, and benzene, even after 100 steps.

For single-cell data, the model assigned cell types more reliably, and on clinical data it predicted more than 1,000 possible disease events slightly better. The difference on image tasks was small. In locomotion planning for simulated walking robots, JEPA-Anything won in two of three environments, while the standard model came out ahead in the third.

The model flagged a liver cancer treatment, so the team tested it

The team's boldest claim comes from liver cancer research. The researchers analyzed the partial predictions a model had learned from biological data including gene activity, protein levels, and CRISPR screens. The top candidate paired IL-18, a signaling molecule that activates immune cells, with blockade of the enzyme CD73, which tumors use to suppress nearby immune responses.

The team tested the combination on liver cancer cells co-cultured with immune cells, on organoids and tumor tissue from three patients each, and in mice. In the organoids and tissue samples, the combination killed more tumor cells than either IL-18 or CD73 blockade alone, and T cells and natural killer cells showed stronger activation. The study doesn't establish whether this could become an actual therapy.

In a second case, the researchers trained the model on simulated orbits without giving it any physical quantities. The learned patterns turned out to be a near-exact match for 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. The team only evaluated one training run, and picked the one with the lowest error.

Clean separation of the learned parts doesn't mean they capture real cause-and-effect relationships, the authors caution. It also remains an open question when such a model becomes reliable enough to guide experiment design. That's the team's long-term goal: AI agents would use JEPA-Anything to propose and rank experiments, then feed results back into the model. Code and models are publicly available.

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 up with LeJEPA , a theoretical foundation designed to keep training stable without the usual workarounds. LeCun is now pursuing the approach through his startup AMI Labs , which raised over a billion dollars in March 2026 to build world models.

Google Deepmind had already tested an AI-generated cancer hypothesis in the lab in October 2025. Its Gemma-based model C2S-Scale 27B proposed the drug silmitasertib to make tumor cells more visible to the immune system, and experiments with human cell models confirmed the prediction. Google Deepmind's multi-agent system Co-Scientist now plans experiments and operates lab equipment, covering parts of the loop the JEPA-Anything team wants to build. Loading samples into the machines, though, still requires humans.