JEPA-Anything: one world-model core spans seven systems, from cancer to Kepler
量子位 · wechat · 2026-09-19
PhAI Labs with Oxford, Stanford, Princeton, and CUHK released JEPA-Anything, a cross-domain world-model framework sharing one predictive core across visual, biological, clinical, control, molecular, physics, and weather systems.
Key mechanism: Standard JEPA funnels all signals through one prediction path, letting strong signals crowd out subtle ones. OPF (Orthogonal Predictive Factorization) splits the target state into orthogonal subspaces handled by separate branches, with activity and encoder-variance constraints to prevent collapse. Learned factors form an analyzable internal interface.
Results: 11.7% in-distribution and 3.5% unseen-combination error reduction vs standard JEPA on controlled Pong; wins on 9/10 matched-dynamics tasks (Burgers MSE -39.7%, shallow water -39.3%, WeatherBench2 -10.5%); lowest error across four molecular systems on 100-step rollouts, though not uniformly better on control tasks.
Science cases: Internal factors yielded an IL-18 + NT5E/CD73 blockade candidate for liver cancer, validated in patient-derived organoids and immunocompetent mice; planetary orbital data alone produced a frequency-semi-major-axis slope of -1.4991 vs Kepler's theoretical -1.5 (R²≈1). Positioned within PhAI Labs' Discovery Foundation Models roadmap.
More from Research
- Training data-poisoning victim speaks out; ByteDance's Xinggang Wang says VAR路线皆 bitter lesson — zibuyu9 · 2026-10-25
- JevBench v1.7 adds a 10-second rule: late decisions count as wrong in agent benchmark — airesearch12 · 2026-10-25
- Open-source virtual human body: 700 muscle actuators, 206 joints, 90 segments — mtizard · 2026-10-25
- EMNLP paper: users trust model verdicts over reasoning, confident tone hides errors — MaartenSap · 2026-10-25
- Prime Intellect merges IcePop score centering with truncated sampling replay in prime-rl — samsja19 · 2026-10-25
- EMBER Targets Token Embeddings: New Method Blocks LLM Knowledge Relearning — megamor2 · 2026-10-25