JEPA-Anything: One Self-Supervised World Model Beats Domain Specialists Across Seven Sciences

A research team spanning Peking University, CUHK, and affiliated labs released a preprint on Thursday arguing that a single self-supervised training principle can do something the AI research community has long treated as structurally impossible: outperform domain-specific world models across seven radically different scientific fields — from molecular simulation to clinical trajectory forecasting to weather — without rewriting the underlying architecture for each one.

The paper, titled "JEPA-Anything: Learning Predictive Models across Different Worlds," introduces Orthogonal Predictive Factorization, or OPF, as an extension of Joint Embedding Predictive Architectures. In experiments conducted against matched domain-specific JEPA baselines, the framework improved reported metrics on all ten dynamics tasks tested and produced the lowest molecular simulation errors across four chemical systems. Most unusually, a factor extracted from the model's latent space nominated a biological intervention that subsequently received experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and live mice — a step from machine learning into wet-lab validation that does not commonly appear in AI architecture papers.

What JEPA-Anything Extends and Why That Matters

The Joint Embedding Predictive Architecture is Yann LeCun's answer to a long-standing critique of generative AI: that predicting pixels or tokens forces a model to spend capacity on perceptually irrelevant details. In a standard JEPA, a context encoder reads an observed portion of the world and a predictor estimates the representation of a hidden or future portion — operating entirely in latent space rather than in the space of raw observations. A frozen exponential-moving-average copy of the encoder generates target representations without gradients, preventing collapse to trivial solutions.

Meta's research team developed this principle through a series of progressively more capable systems. I-JEPA in 2023 applied it to image patches, predicting the embedding of masked image regions from visible context. V-JEPA in 2024 extended it to spatiotemporal video tubes. V-JEPA 2, released in June 2025, trained on over one million hours of video and produced a robot planning system requiring only 62 hours of domain-specific footage to enable zero-shot manipulation on unseen robot arms. Each of these advances operated within the visual and physical-interaction domain.

JEPA-Anything's claim is that the latent prediction interface — the core of what makes JEPA work — is not inherently visual. If domain-specific encoders handle the conversion from raw observations to tokens, the same factorized predictive learning objective should be able to organize the predictive capacity of molecular graphs, clinical records, weather fields, and cell transcriptomic profiles using identical training machinery.

The Mechanism: Splitting Prediction Into Orthogonal Subspaces

The key technical addition is Orthogonal Predictive Factorization. Where a standard JEPA uses one predictor to predict a monolithic target embedding, OPF introduces K learned projection matrices that partition the target latent space into non-overlapping subspaces. Each subspace has its own dedicated predictor. After prediction, the factor predictions are recombined through a pseudoinverse synthesis step that maps them back into the full-dimensional latent space.

The orthogonality is enforced during training through two penalty terms. Within each factor block, the projection matrix is penalized toward being orthonormal. Across different factor blocks, a Frobenius-norm penalty drives the subspaces apart, discouraging different predictors from targeting the same latent directions. The design prevents a predictive-capacity waste problem the authors identify in monolithic JEPA: when a single predictor must simultaneously represent location, identity, transformation, and dynamics, high-variance or easily predicted structure dominates the gradients and weakly predictable structure is deprioritized.

Two regularizers handle collapse. A factor activity term penalizes projectors when any coordinate's empirical standard deviation falls below a threshold, keeping every factor dimension active across training batches. A separate online encoder variance term directly pushes the context encoder to maintain variation across its output coordinates. These are additive to each domain's existing base loss — the OPF terms are a drop-in addition to whatever training objective already exists for a domain-specific baseline.

Mathematically, if the K projection matrices are exactly orthogonal and their subspace dimensions sum to the full latent dimension d, the factors form a lossless partition: a synthesized state computed from predicted factor coordinates has the same norm as the original and can be exactly reconstructed by summing the outer products of each factor's projection with its predicted coordinate vector. This property holds approximately during training under the orthogonality penalty and exactly in the zero-penalty limit, giving a well-conditioned synthesis path that the authors prove reduces synthesis error amplification compared to unconstrained multi-head designs.

Read more: Programmable World Model Hits 98% State Accuracy by Separating Engine from Renderer

Seven Domains, Three Experiment Groups

The evaluation framework is organized into three groups that test different modes of using the world state.

The first group — terminal readout — tests whether OPF pretraining improves frozen encoder representations on visual binding tasks in rendered MuJoCo scenes and on single-cell gene-expression state representation across kidney, peripheral blood, and perturbation datasets using the scGPT backbone. It also tests whether a one-step synthesized future state can improve longitudinal disease event forecasting across more than one thousand clinical event risks in a multimodal patient cohort using a GPT-2 clinical language model. Against matched JEPA baselines — same encoder, same optimization budget, same data, same downstream readout — the OPF pretraining improves across these tasks.

The second group — latent world dynamics — tests whether predicted states can be recursively reused, which is the operationally challenging regime for any world model. The evaluated settings include intervention-conditioned prediction on Interventional Pong from the CITRIS benchmark suite, where JEPA-Anything reduces single-intervention prediction error by a reported 34.8 percent; out-of-distribution and long-horizon dynamics on the PDEBench physical field benchmark, WeatherBench 2 weather forecasting, and CausalWorld robot manipulation; continuous-control planning on Hopper, Walker2d, and HalfCheetah; and force-free molecular simulation on four chemical systems using a TrajCast-style O(3)-equivariant backbone. The molecular results — lowest one-step and 100-step errors across all four systems, as reported by the authors — are notable because molecular rollouts compound errors over 100 autoregressive steps, where even small per-step improvements in latent organization can prevent trajectory drift.

The third group moves from predictive performance to scientific interpretability: whether the factor coordinates learned by OPF encode physically or biologically meaningful structure even without explicit supervision toward that structure.

When Latent Factors Go to the Wet Lab

The most striking experiment in the paper involves the biological factor analysis. In Group III, the researchers use the OPF factor projections — the same matrices learned during pretraining — as a diagnostic interface on single-cell data. One factor was nominated by this analysis as a candidate for a biological intervention. The paper reports that this factor-nominated intervention subsequently received experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and mice.

This is not a computational benchmark. The model did not observe the wet-lab outcome during training. The claim, if it withstands scrutiny, is that the latent factorization captured structure in the cellular transcriptomic state space that corresponds to a causal or mechanistic relationship with sufficient specificity to generate a valid testable hypothesis. Whether this holds at the level of peer review — where the experimental protocol, the nature of the "experimental support," and the statistical design will be examined — remains to be established. The preprint has not yet been peer-reviewed.

A second scientific analysis recovers orbital mechanics from latent factor modes without any supervision toward physical laws. Fitting a log-log relationship between the frequency modes extracted from learned factors and orbital period structure, the model's latent representation produces a slope of -1.4991, compared to the theoretical Keplerian value of -1.5 from Kepler's third law. The near-exact recovery of a known physical scaling relationship — from a world model trained only to predict future states — is consistent with the paper's central argument that predictive factorization can surface physically meaningful structure as a side effect of learning to predict.

Where the Argument Remains Open

The experimental design addresses a clear comparison: OPF versus matched JEPA with the same encoder, data, and budget. What it does not yet provide — based on the available preprint — is a clean ablation separating the contribution of orthogonal subspace partitioning from the contribution of the additional regularizers and the additional predictor capacity that K separate prediction heads introduce compared to one monolithic predictor. A reviewer of the August 2026 predecessor paper, Orthogonal JEPA, which established the OPF mechanism on a narrower set of visual tasks, noted specifically that the reported gains are not yet attributable to factorization because the comparison adds capacity and regularizers without ablations. JEPA-Anything extends the scope to seven domains and three evaluation groups but does not resolve this attribution question by itself.

A second limitation is that the GitHub repository, released under Apache 2.0, provides the reusable OPF core library, task design tools, and a synthetic structural example — but does not include the trained model weights for the domain experiments described in the paper. Researchers who want to reproduce the molecular dynamics or clinical forecasting results will need to re-run training using their own data and compute. The repository is explicitly designed as a starting point for new instantiations, not as a model hub.

The paper is also filed under the cs.CL category on arXiv — Computation and Language — which seems like a mismatch for a framework that encompasses molecular simulation, PDE fields, and wet-lab biology. This may affect how it reaches its most relevant audiences in the computational biology and physical-simulation communities.

The Significance: One Training Principle, Many Scientific Worlds

Domain-specific world models exist today for almost every scientific field that JEPA-Anything covers. Weather forecasting has GraphCast and WeatherBench-trained models. Molecular simulation has force-field-based and machine-learning interatomic potential approaches. Clinical AI has longitudinal patient models. Single-cell biology has scGPT and related transformers. Each of these is engineered around the structure of its domain's data and typically cannot be repurposed across domains without substantial reengineering.

JEPA-Anything's claim — that a single predictive-factorization training objective, applied additively to each domain's existing architecture, consistently improves over that domain's own JEPA baseline — is a meaningful step toward the theoretical goal LeCun articulated in 2022: that intelligence, whether biological or artificial, is organized around a domain-agnostic world-modeling principle rather than a collection of domain-specific modules. The OPF mechanism is more modest than that vision — domains still require their own encoders and adapters — but it is one of the clearest empirical arguments yet that the learning principle generalizes across heterogeneous physical and biological systems.

Whether this remains true as domain complexity and dataset scale increase, whether the wet-lab validation survives peer review, and whether the factor-level diagnostics become a reliable hypothesis-generation tool in drug discovery or systems biology are the milestones worth watching. A framework that can nominate a valid biological intervention from latent space alone — before any targeted experiment — would change the economics of hypothesis generation in life sciences. That is what JEPA-Anything's Group III experiments are positioned to demonstrate, and the case for that claim is currently resting on one set of biological experiments described in an unreviewed preprint.