Paper applies one predictive framework across seven domains, from molecules to mice

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to act on. The code is public, but every number is the authors' own, measured against their own baselines, so there is no API, no price and no model to migrate to.

JEPA-Anything, posted to arXiv on 16 September, runs one predictive framework across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. It extends joint-embedding predictive architectures — models trained to predict the representation of future data rather than raw pixels — by splitting each prediction target into factors learned on separate pathways. Against matched JEPA baselines the authors report gains on all 10 dynamics tasks, a 34.8% cut in intervention-prediction error and the lowest molecular errors in four systems. One factor-nominated biological intervention drew support in cell co-cultures, organoids, tumour fragments and mice. The paper is a preprint, not peer-reviewed.

Key facts

  • ·JEPA-Anything was posted to arXiv on 16 September 2026 and evaluated across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. source
  • ·The authors report improved metrics on all 10 matched dynamics tasks against JEPA baselines, and a 34.8% reduction in single-intervention prediction error on Interventional Pong. source
  • ·The method reports the lowest one-step and 100-step molecular errors among compared methods in all four systems tested. source
  • ·A factor-nominated biological intervention received experimental support in cell co-cultures, patient-derived organoids, tumour fragments and mice. source
  • ·Latent orbital modes recovered the Keplerian scaling exponent with a fitted slope of -1.4991. source
  • ·Code is published at github.com/Gen-Verse/JEPA-Anything. source

What the sources say

  • Hugging Face Daily Papers — The authors' own abstract, setting out the factorised architecture, seven evaluation domains and cross-domain results.

Sources

The original reporting. Follow these — they did the work.

← the wire