Loading
Life sciences · Preprint
arXiv · September 9, 2026
Early or partial results. Treat as a signal, not a conclusion.
SG-JEPA is a novel machine learning architecture designed to improve latent dynamics prediction in world models by incorporating physics parameters via action-conditioning. In controlled simulation tasks with varying gravitational fields, it demonstrates up to 2.5× improvement in robotic control success over a baseline model, though the work remains unreviewed and does not address clinical or biomedical applications.
Controlled computational experiment comparing neural architecture performance on simulated physics prediction tasks. Simulated dynamical systems and robotic agents in controlled virtual environments with variable gravitational parameters.. Intervention: SG-JEPA architecture: JEPA world model extended with physics-parameter conditioning and action-conditioning to temporal model.. Compared with: DINO-WM baseline model.
SG-JEPA reduces open-loop prediction error by up to 2 times on two-dimensional datasets compared to DINO-WM SG-JEPA increases control success rate up to 2.5 times for three-dimensional robotic datasets Feature analysis shows most performance gain derives from encoder learning better representations rather than predictor learning better dynamics
Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.
The source did not state who this applies to in practice.
This is an unreviewed preprint describing an initial machine learning architecture for physics prediction with controlled experimental validation but no peer review and no clinical or translational evidence.
As stated by the source record.
Quoted from the source exactly as published.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
Joint-Embedding Predictive Architecture (JEPA) world models learn a compact latent representation of the world that supports prediction and planning, but their capability to learn physics and generate physically realistic dynamics remains hitherto untested. In this work, we introduce SemiGroup-JEPA (SG-JEPA), which extends the LeWorldModel framework by supplying the parameter governing the physics to the temporal model via action-conditioning and jointly training an encoder and predictor through an autoregressive latent rollout. To evaluate the model's ability to generalize out of distribution, we design dynamical tasks under different gravitational fields that, despite obeying the same physical law, exhibit qualitatively different dynamics, ranging from floating motion in weak gravitational fields to rapid bouncing in strong ones. In contrast to DINO-WM, SG-JEPA reduces open-loop prediction error by up to 2 times on two-dimensional datasets, and increases control success rate up to 2.5 times for three-dimensional robotic datasets, for which we train independent diffusion policies. To explain this advantage, we develop a linear feature model that separates local law-conditioned error from its recursive amplification under rollout. Guided by this model, we find that back-propagating the multi-step rollout loss into the representation trains the encoder to keep the features that the predictor can carry forward, and that those are the features the dynamics depend on, so most of the gain comes from the encoder learning better features rather than from the predictor learning better dynamics. See project page at https://sg-jepa.github.io.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.