Hugging Face Daily PapersTaoyong Cui, Zhongyao Wang, Xinyue Xu1 min readpaperadvanced
JEPA-Anything: Learning Predictive Models across Different Worlds
Summary
JEPA-Anything extends joint‑embedding predictive architectures with orthogonal predictive factorization, letting a single model learn complementary latent factors that can be recombined for prediction across disparate domains. The paper shows consistent performance gains on ten dynamics tasks, molecular simulations, and clinical event forecasting, plus experimental validation of a biologically‑de…
- OPF splits latent targets into orthogonal factors learned via dedicated pathways, enabling a shared predictive core across vision, biology, control, and physics domains.
- On ten benchmark dynamics tasks JEPA-Anything beats domain‑specific JEPA baselines, cutting single‑intervention error on Interventional Pong by 34.8%.
- It attains the lowest one‑step and 100‑step molecular dynamics errors across four chemical systems, outperforming prior methods.
- A factor‑selected biological intervention predicted by the model was confirmed experimentally in cell co‑cultures, organoids, tumor fragments, and mice.
Researchers building general world models and engineers needing cross‑domain predictive systems should see a concrete, experimentally‑validated approach that unifies representation learning and intervention prediction.
8/10
