JEPA-Anything tests one predictive core across seven systems, from cancer cells to planetary orbits

The paper defines a latent world model broadly: a model that builds an internal state from available information, or context, and uses it to predict another state in the same system. That other state may be a future time point, a hidden spatial region, a different view, or the result of an intervention. JEPA, or Joint-Embedding Predictive Architecture, is a representative route in this area. Instead of reconstructing a target state in raw space, JEPA encodes observations into latent representations and predicts in representation space, leaving more capacity for structures with predictive value and reducing interference from textures and noise.

Standard JEPA usually sends information into a single target embedding and a single prediction path. The report argues that real-world changes are often not single-threaded: molecular translation can accompany atomic vibration, slow weather evolution can contain violent local disturbances, and biological systems can involve multi-scale, multi-path changes at once. When signals of different scales, entities, and difficulty are forced together, easier or more prominent signals may take more predictive capacity, while weaker but important information can be canceled in gradient conflict or masked by stronger signals.

JEPA-Anything addresses this through Orthogonal Predictive Factorization, or OPF. The mechanism splits a target state into complementary subspaces handled by separate prediction branches, then reassembles them into the full world state. Orthogonal constraints reduce redundant learning across branches, while factor activity constraints and encoder variance constraints prevent some factors from becoming inactive and lower the risk of representation collapse. The factors are not preassigned to represent speed, temperature, or a biological pathway; what they learn is determined by predictable structure in the data. After training, they leave an internal interface for later analysis.

The first experiment used CITRIS Interventional Pong, a controlled dynamic environment. During training, the model received only single-factor changes; during testing, it faced multi-factor combinations it had never seen. JEPA-Anything reduced prediction error by about 11.7 percent in in-distribution single interventions compared with standard JEPA, and lowered mean squared error by about 3.5 percent on unseen multi-factor combinations, improving in all five paired training seeds.

The report then built a matched dynamics benchmark with ten tasks, ensuring that standard JEPA and JEPA-Anything used identical training data, encoders, state-transition backbones, compute budgets, and evaluation splits. JEPA-Anything improved performance on nine of the ten prediction tasks. Mean squared error fell by 39.7 percent on the Burgers equation from PDEBench, by 39.3 percent on the shallow-water equation, and by 10.5 percent on WeatherBench 2. The tests also included multi-step rollouts, where each prediction became the next input. In molecular systems—liquid water, alpha-quartz, acetaminophen, and benzene—JEPA-Anything achieved the lowest error for both one-step MAE and final position RMSD after 100-step free rollout.

The results were not uniformly better. In continuous control, JEPA-Anything scored higher than standard JEPA on Walker2d and HalfCheetah, but standard JEPA performed better on Hopper. In additional Burgers tests, when prediction length rose from 20 to 50 steps, JEPA-Anything remained ahead but its advantage narrowed. The report says it validates the cross-domain applicability of shared predictive principles, not that it has achieved a single world model covering all domains.

A third set of experiments asked whether factors formed in the model's latent space could support scientific exploration. In liver cancer research, the team analyzed the model's orthogonal factor coordinates and selected a candidate combination: IL-18 plus NT5E/CD73 blockade. The combination was tested in a Huh7-PBMC co-culture system, then in patient-derived hepatocellular carcinoma organoids and tumor tissue fragments, and later in immunocompetent mice. The report says that in three patient organoids and three tumor tissue fragments, the IL-18 and CD73 blockade combination showed the strongest tumor cell killing and was accompanied by significantly enhanced immune cell activation.

The planetary case worked in the opposite direction, using a known physical law to check what the model had learned. Researchers fed the model simulated planetary positions and velocity trajectories without providing Kepler's laws. After training, they performed spectral analysis on the model's latent modes and paired extracted frequencies with orbital semi-major axes. The fitted slope was -1.4991, compared with the theoretical value of -1.5 for Kepler's third law, with an R-squared of 0.99999999.

The report positions JEPA-Anything at the world-model layer of PhAI Labs' Discovery Foundation Models strategy, which aims for AI to participate in identifying unknowns, forming hypotheses, designing interventions, and revising them based on external evidence. Before longer and more expensive real experiments, the framework seeks to compress predictable parts of scientific trajectories and real-world changes into faster predictive capability. The distance to AI understanding everything remains substantial, but the report presents a possibility: structures learned inside a world model can serve prediction and also be exposed for researchers to analyze and test against experiments and known laws.