Researchers have released JEPA-Anything, a latent world model that applies a technique called orthogonal predictive factorisation (OPF) across seven scientific domains. Published on 18 September 2026, the model improves reported metrics on all 10 dynamics tasks against matched JEPA baselines and reduces single-intervention prediction error on Interventional Pong by 34.8%.
The system splits latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design. The evaluation spans vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. It includes forecasting of over 1,000 clinical events and 100-step molecular rollouts across four systems.
How orthogonal predictive factorisation works
OPF builds on joint-embedding predictive architectures. Rather than treating a latent representation as one monolithic target, the method decomposes it into complementary factors. Each factor gets its own learning pathway, and the factors are recombined for prediction. The result is a shared interface for intervention prediction, state synthesis and scientific diagnosis.
Among the compared methods, JEPA-Anything's one-step and 100-step molecular errors are the lowest in all four systems tested. Latent orbital modes recover the Keplerian scaling exponent with a fitted slope of -1.4991. A factor-nominated biological intervention received experimental support in cell co-cultures, patient-derived organoids, tumour fragments and mice.
A unified predictive model across domains
The research vision positions JEPA-Anything as a unified predictive model or science world simulator. When a researcher proposes an action or intervention, the model predicts likely state changes, outcomes and risks in advance. The goal is to help screen out ineffective or costly experimental plans before they reach the lab.
Today, the system does not simulate every scientific environment and is not connected to ScienceBuddy or ScienceIDE. The paper frames these as directions to explore rather than current capabilities. The technical report and Hugging Face model collection are public, though this release includes no public demo.
For professionals building predictive models in scientific settings, the AI Scientific Research Courses cover representation learning approaches relevant to this type of work.
What the evaluation covers
The paper tests representation learning, intervention prediction, out-of-distribution generalisation and long-horizon dynamics. Ten matched dynamics tasks provide direct comparison against JEPA baselines. The clinical component involves forecasting over 1,000 clinical events, while the molecular dynamics tests run 100-step rollouts across four different systems.
The core question the research asks is how to represent the many predictable factors of a complex system, predict the state changes different interventions bring, and ground experiment selection in prediction. The goal is reusing a latent world-state interface across domains rather than building bespoke models for each scientific field.
Why this matters for science and research professionals
If you run experiments in drug discovery, clinical trial design, or physical simulation, a model that predicts intervention outcomes across domains could reduce the number of wet-lab iterations you need. The 34.8% error reduction on Interventional Pong is a controlled benchmark, but the biological validation - spanning organoids, tumour fragments and mouse models - shows the approach translating to real experimental settings. The clinical event forecasting component, covering over 1,000 events, suggests potential for patient trajectory modelling. These are selected domains and tasks, not a universal claim, but the consistent improvement across all 10 dynamics tasks indicates the factorisation method generalises better than standard JEPA architectures on the problems tested.
Your membership also unlocks:
