One predictive core, seven different worlds

Machine learning usually follows a path of specialization. One network predicts the next video frames, another predicts robotic arm movements, another forecasts the weather, and yet another tracks molecular trajectories. The team behind the paper "JEPA-Anything" asked the opposite question: is there one shared principle for building a predictive model that works across radically different systems?

What a world model is

In this framing, a world model is any network that builds an internal state from available information and uses that state to predict another state of the same system. That "other state" can be the future in time, a hidden part of space, a different point of view, or the effect of an intervention. The JEPA family (Joint-Embedding Predictive Architecture) predicts in representation space rather than in raw pixels. This directs the model's capacity toward structure that actually predicts something, instead of reproducing textures and noise.

The problem arises when different scales and rates of change land in one shared prediction target. Easier, stronger signals then consume most of the capacity, pushing out subtle but important changes. The authors address this with a mechanism called OPF (orthogonal predictive factorization). It splits the target state into complementary subspaces, each handled by a separate prediction branch, and then merges them back. A set of constraints, including orthogonality of the factors, their activity and the variance of the encoder, prevents the branches from learning the same thing and protects against representation collapse. Crucially, the factors are not assigned in advance to specific physical quantities. The structure of the data determines what they represent.

A test on seven systems

The authors did not stop at video or control. They tested the same predictive core in seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. Each domain keeps its own observation method, its own construction of the context-target pair and its own encoder. The prediction layer and a single hidden-state interface are shared across all of them.

In a controlled environment, an intervention version of Pong, the model was trained only on data with single changes and tested on unseen combinations of factors. On the training distribution, the prediction error fell by about 11.7 percent relative to standard JEPA. On unseen combinations, the mean squared error dropped by about 3.5 percent. The improvement appeared in all five random draw pairs.

In the "Matched Dynamics Benchmark", both architectures received identical data, encoder, state-transition backbone, compute budget and evaluation split. JEPA-Anything improved the result in nine of ten tasks. Gains included 39.7 percent on the Burgers equation, 39.3 percent on the shallow water equation and 10.5 percent on WeatherBench 2. Long step-by-step predictions, where the model's own results feed back into the input, were tested separately.

The most telling example concerns astronomy. The model was given simulated positions and velocities of planets without being told Kepler's laws. An analysis of internal prediction patterns revealed a dependence of orbital frequency on semi-major axis with a slope of -1.4991. The theoretical value for Kepler's third law is -1.5. The fit coefficient R² was 0.99999999.

In biology, the factors learned by the model were used to propose an intervention combining cytokines with antibody blockade. The idea passed further stages of verification: on cell lines, on patient-derived organoids and in mice. Molecular dynamics was checked on liquid water, alpha quartz, paracetamol and benzene. The authors' conclusion is cautious but clear: it is not about replacing domain knowledge, but about a shared layer of prediction learning on which that knowledge can be built.