JEPA-Anything is an independent model-method study of latent world models and orthogonal predictive factorisation.
It explores how to learn latent representations that predict changes in world state, and how to offer one interface for intervention prediction, state synthesis and scientific diagnosis. The seven domains are vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather.
In the research vision, JEPA-Anything can be read as a unified predictive model or science world simulator over many scientific environments: when a researcher proposes an action or intervention, the model predicts the likely state changes, outcomes and risks in advance, helping screen out ineffective or costly plans. That is a direction to explore: today it does not simulate every scientific environment, and it is not connected to ScienceBuddy or ScienceIDE.
In the wider research picture it asks how to represent the many predictable factors of a complex system, how to predict the state changes different interventions bring, how to ground experiment selection and scientific diagnosis in prediction, and how to reuse a latent world-state interface across domains.
The paper introduces orthogonal predictive factorisation (OPF): on top of joint-embedding predictive architectures it decomposes latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design. The evaluation spans representation learning, intervention prediction, out-of-distribution generalisation and long-horizon dynamics, including 10 matched dynamics tasks, forecasting of over 1,000 clinical events and 100-step molecular rollouts across four systems.
Against matched JEPA baselines, JEPA-Anything improves reported metrics on all 10 dynamics tasks and reduces single-intervention prediction error on Interventional Pong by 34.8%; among the compared methods, its one-step and 100-step molecular errors are the lowest in all four systems. A factor-nominated biological intervention receives experimental support in cell co-cultures, patient-derived organoids, tumour fragments and mice, and latent orbital modes recover the Keplerian scaling exponent with a fitted slope of -1.4991. These are observations on selected domains and tasks.
The technical report and the Hugging Face model collection are public. This release includes no public demo.