← survey

A taxonomy of world models

Two comprehensive surveys give ready-made organizations of the field:

Ding et al. (arXiv:2411.14499, ACM Computing Surveys) split the field by function:

  1. Implicit/internal representation for understanding world dynamics — LeCun’s JEPA line, and the Dreamer/MuZero decision-making tradition;
  2. Future-state prediction via generative simulation — Sora-style video generation, embodied and spatial prediction.

The survey is explicitly motivated by the tension between those two camps (see the debates).

Zidan et al. (arXiv:2606.00133, May 2026 preprint) instead use a four-axis scheme: architecture; methodological family (state-space/recurrent, transformer-based, diffusion-based, physics-informed, language-augmented multimodal); reasoning strategy (imagination-based planning, latent policy learning, counterfactual reasoning, planning under uncertainty); and application domain.

The map this site uses

For a practitioner, the most useful cut is what the model predicts, and for whom. We organize the survey into six branches:

Branch Predicts Exemplars
Latent prediction abstract representations, not pixels I-JEPA, V-JEPA, V-JEPA 2
Video-generative (interactive) video Sora, Veo, Genie 1–3
3D / spatial persistent 3D space World Labs Marble/RTFM, 3DGS generation
Agents & robotics consequences of actions, for control Dreamer, MuZero, Cosmos, V-JEPA 2-AC
Game / simulation playable environments Oasis, GameNGen, Muse/WHAM
Driving traffic futures, for autonomy GAIA-1/2, Vista

The branches are converging. NVIDIA’s Cosmos 3 (May 2026) is deliberately omnimodal in both input and output — text, images, video, audio, and action trajectories in; text, images, video, audio, and actions out — making it simultaneously a video-generative world model and an action-conditioned one, collapsing two branches into a single system. That is exactly the “unified multimodal world models” future the surveys forecast.

A framing to avoid: an “explicit vs. implicit world models” two-way taxonomy sometimes attributed to Oh’s ACM Computing Surveys tutorial (arXiv:2606.12783) does not survive verification against the tutorial itself — we do not use it. The tutorial remains a good source on V-JEPA 2 and Genie mechanics.


← What is a world model?Latent prediction: the JEPA family →