🛰️ Daily AI Frontier
‹ back to 2026-09-18

JEPA-Anything: Learning Predictive Models across Different Worlds

Research LLMs & Foundation Models

Ranking

Overall 86
Content 95
Popularity 66

Observed public metrics from 1 member.

Representative image for JEPA-Anything: Learning Predictive Models across Different Worlds

Merged summary

TL;DR - JEPA-Anything is a domain-agnostic world-modeling framework that factorizes latent prediction targets into complementary components. It improves predictive performance across seven heterogeneous domains and links learned representations to interventions and experimentally supported scientific findings.

  • Orthogonal predictive factorization extends JEPA by learning target factors through dedicated pathways and recombining them in a shared predictive architecture.
  • The framework outperforms matched JEPA baselines on all 10 dynamics tasks and reduces single-intervention prediction error in Interventional Pong by 34.8%.
  • It achieves the lowest reported one-step and 100-step molecular-dynamics errors among compared methods across four systems.
  • Learned factors produced a biologically nominated intervention supported in multiple experimental models and recovered Keplerian scaling with a fitted exponent of -1.4991.

Sources (1)

JEPA-Anything: Learning Predictive Models across Different Worlds

arXiv cs.CL Taoyong Cui, Zhongyao Wang, Xinyue Xu, Weiyang Liu, Zhaochen Yu, Yuying Zhang, Qiang Gao, Mengyue Yang, Wanli Ouyang, Pheng Ann Heng, Yingcheng Wu, Zhenfei Yin, Ling Yang 2026-09-17 arXiv:2609.20800
Public signals Hugging Face upvotes 70
Providers: Hugging Face · Upvotes 70 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-25 14:18:35.348720 UTC

TL;DR - JEPA-Anything is a domain-agnostic world-modeling framework that factorizes latent prediction targets into complementary components. It improves predictive performance across seven heterogeneous domains and links learned representations to interventions and experimentally supported scientific findings.

  • Orthogonal predictive factorization extends JEPA by learning target factors through dedicated pathways and recombining them in a shared predictive architecture.
  • The framework outperforms matched JEPA baselines on all 10 dynamics tasks and reduces single-intervention prediction error in Interventional Pong by 34.8%.
  • It achieves the lowest reported one-step and 100-step molecular-dynamics errors among compared methods across four systems.
  • Learned factors produced a biologically nominated intervention supported in multiple experimental models and recovered Keplerian scaling with a fitted exponent of -1.4991.
item →