Glonce

JEPA-Anything extends one predictive architecture across seven domains

researchThe Glonce editorsHugging Face🔥56Glonce rating56Impact65Novelty70Relevance100Surprise50Players20Credibility75
JEPA-Anything extends one predictive architecture across seven domains

World models let a system anticipate what happens next and what an intervention would change, but predictive models are usually built separately for each domain. A new paper proposes JEPA-Anything, a single framework meant to work across radically different kinds of data. It builds on joint-embedding predictive architectures (JEPA) and adds what the authors call orthogonal predictive factorization (OPF): the latent target is split into complementary factors, each factor is learned through its own pathway, and the factors are recombined inside one shared predictive design. The researchers evaluate JEPA-Anything across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. The tests cover representation learning, predicting the effect of interventions, generalization to out-of-distribution cases and long-horizon dynamics, including 10 matched dynamics tasks, forecasting of more than 1,000 clinical events, and 100-step molecular dynamics rollouts across four separate molecular systems. Against matched JEPA baselines, JEPA-Anything improves the reported metrics on all 10 dynamics tasks and cuts single-intervention prediction error on the Interventional Pong benchmark by 34.8%. On the molecular side, it produces the lowest one-step and 100-step errors among the compared methods in all four systems tested. Beyond prediction accuracy, the authors report that a biological intervention nominated by one of the model's factors held up experimentally in cell co-cultures, patient-derived organoids, tumor fragments and mice. Separately, latent orbital modes learned by the model recovered the Keplerian scaling exponent from orbital mechanics, with a fitted slope of -1.4991. The paper frames these results as evidence for a common factorized predictive principle that can span very different kinds of systems, tying world modeling to both intervention prediction and experimentally grounded scientific discovery. Code for JEPA-Anything is released on GitHub. The abstract does not name individual authors, an institution, a publication venue or which specific JEPA variants make up the 'matched baselines' it is compared against.

Key facts

  • JEPA-Anything adds orthogonal predictive factorization (OPF) to joint-embedding predictive architectures, splitting latent targets into factors learned via dedicated pathways.
  • It is evaluated across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather.
  • Against matched JEPA baselines it improves metrics on all 10 tested dynamics tasks and cuts single-intervention error on Interventional Pong by 34.8%.
  • It reaches the lowest one-step and 100-step molecular dynamics errors among compared methods across four molecular systems, in rollouts of 100 steps and forecasts of over 1,000 clinical events.
  • A factor-nominated biological intervention held up in cell co-cultures, patient-derived organoids, tumor fragments and mice; code is public on GitHub.

Why it matters

Most predictive world models are built one domain at a time: a vision model, a separate model for robotics control, another for molecular simulation. JEPA-Anything argues a single factorized architecture can serve vision, biology, clinical data, control, molecular dynamics, physical fields and weather at once, which is a claim about the generality of one learning principle rather than about any single benchmark.

Who it affects

The direct audience is researchers working on world models, representation learning and simulation across the domains tested. The biological intervention result, which held up in cell co-cultures, patient-derived organoids, tumor fragments and mice, also speaks to researchers looking for computational methods to nominate wet-lab experiments.

How to use it

The paper releases code for JEPA-Anything on GitHub, alongside the abstract. No license, pricing or usage terms for the repository are given in the source, and no specific matched-baseline models are named, so reproducing the exact comparisons requires reading the released code and paper rather than the abstract alone.

How solid is it

The reported numbers are specific and consistent: gains on all 10 tested dynamics tasks, a 34.8% error reduction on Interventional Pong, lowest errors in all four tested molecular systems over 100-step rollouts, and a fitted orbital-mode slope of -1.4991 against the known Keplerian exponent. These are the authors' own reported results against baselines they selected as 'matched'; the abstract does not identify those baseline models, name the authors or institution, or state a publication venue, so independent verification depends on the linked code and the full paper rather than this summary alone.

Risks and caveats

The claims come from the paper's own abstract, without independent replication described here. Improvements are measured against baselines the authors call 'matched' but do not name, which limits how far the comparison can be checked from the abstract alone. The framework spans seven quite different domains in one paper, from clinical forecasting to molecular dynamics to biological experiments, which is a broad claim to substantiate in a single piece of work.