JEPA-Anything: One Predictive Framework Modeling Worlds Across Seven Domains

Gen-Verse (PhAI Labs with Oxford, Stanford, Princeton and CUHK) publishes JEPA-Anything (arXiv 2609.20800), a domain-agnostic world-modeling framework where a single predictive core spans seven heterogeneous domains, with code open-sourced and a liver-cancer intervention candidate generator built on top.

Confirmed

• The core mechanism is orthogonal predictive factorization (OPF): it splits the single latent target of standard JEPA into complementary, individually inspectable and testable factors learned via separate pathways, avoiding collapse or poor diagnosability of a single JEPA objective. @Taoyong Cui frames it as an extension of the joint-embedding predictive architecture; @burnytech calls it a generalization of LeCun's JEPA路线.
• Evaluation covers vision, biology, clinical, control, molecular, physical fields and weather, sharing one predictive core.
• Per @akhaliq, the paper page is on Hugging Face, the code lives at Gen-Verse/JEPA-Anything, and it was featured by HuggingPapers.

Why it matters

• JEPA has mostly been validated within a single modality; this work applies one factorized predictive framework across heterogeneous worlds from pixels to physical fields, offering a decomposable, testable recipe toward general world models.
• OPF improves interpretability and engineering control over cross-domain transfer, and the liver-cancer application hints at scientific-discovery use cases.

2026-09-18 ~ 2026-09-20 · 6 related posts

• [source] JEPA-Anything: One Predictive Framework Spans Vision, Biology, Weather and More — Taoyong Cui · 2026-09-18
• [source] JEPA-Anything paper: learning predictive models across different worlds — _akhaliq · 2026-09-19
• JEPA-Anything: one predictive framework spanning vision, biology, weather and more — rbhar90 · 2026-09-19
• [source] JEPA-Anything: one world-model core spans seven systems, from cancer to Kepler — 量子位 · 2026-09-19

2 near-duplicate retellings: burny_tech · burny_tech