World Modeling is an open, continually-updated encyclopedia of world models and the Joint-Embedding Predictive Architecture (JEPA) family. It collects the methods, theory, and resources of the field in one place, explains each at a consistent technical depth, and cross-links them so the lineage of ideas is easy to follow.
The organizing idea is simple: a world model predicts the future in representation space rather than raw observation space. A context encoder embeds what is visible, an exponential-moving-average target encoder embeds what is hidden, and a predictor — optionally conditioned on an action — maps one to the other. This single recipe now spans images, video, audio, 3D, graphs, molecules, time series, biosignals, language, and robotics.
How this encyclopedia is organized
- Encyclopedia — every method as a short, self-contained card, grouped by topic and filterable by modality.
- Journal — short, dated entries on new ideas, added as they appear.
- Resources — the full link tables: papers, code, datasets, benchmarks.
The biology thread ★
One theme runs throughout: whether the same latent-prediction recipe can yield a controllable, action-conditioned world model of the cell — one that predicts how biological state responds to interventions and helps prioritize decisions across the drug-discovery pipeline. Entries marked ★ are the building blocks of that program.
The collection grows continuously — coverage is broad but not exhaustive. Corrections and additions are welcome.
Contact
Remek Kinas
remigiusz.kinas@gmail.com
x.com/KinasRemek