MotionJEPA: Oxford-led team fixes JEPA world models' bias toward slow, simple features

randall_balestr · x · 2026-09-23

Researchers from Oxford, vivo, and others released MotionJEPA, addressing JEPA world models' inductive bias toward slow features that suppresses dynamic information. In Pong experiments, baselines reconstruct scoreboards and bats but miss the fast-moving ball, while MotionJEPA learns balanced static/dynamic representations. The DISReg regularizer predicts temporal difference image embeddings without pixel reconstruction, combining a static term and an unconstrained dynamic term to prevent collapse. Paper and code are open.

Original post →

More from Research

Research channel →