Skip to content

Three programs

World models

By 2026, “world model” names at least three communities with different artifacts and different failure modes. Most overclaims come from sliding between columns mid-argument.

Latent dynamics Generative simulators Language / digital
Predicts Next latent state Next pixels / steppable world Next observation in text tools - or “state inside an LLM”
Built for Plans that transfer Plausible, consistent rollouts Agent loops and code/env simulators
Exemplars Dreamer, V-JEPA 2-AC, LeWM Genie 3, NVIDIA Cosmos, Marble Meta CWM, Qwen-AgentWorld, Othello-GPT probes
Characteristic failure Compounding latent error over horizon Physics as regularity, not law; short consistency Locally perfect, globally incoherent (taxi map)

Program one - latent dynamics for control. Predict in representation space. Skip the leaf flutter. JEPA is the loudest modern thesis in this column; Dreamer is the RL instantiation that trains the policy inside the model’s imagination. This track walks JEPA first, then DreamerV3 / Dreamer 4.

Program two - generative simulators. Keep the pixels. The world model is the environment: Genie 3 playable worlds, Cosmos open weights used mainly as synthetic-data engines. Horizon is measured in minutes of consistency, not in “laws of physics.”

Program three - language world models. Either probe whether an LLM already contains a coherent world (partial, task-shaped - Othello yes, taxi map no), or train an explicit one: code execution traces (CWM), learned digital environments for agents. Later wave; it is the bridge back to this site’s language stack.

It is tempting to treat Genie demos and JEPA robot papers as the same announcement. They are not. One ships a steppable video environment; the other ships a latent transition model used inside model-predictive control. Both can be valuable. Only the second is graded on whether a plan survives on a real Franka arm.

The same table, built column by column - then the one misconception it exists to prevent.

Worth having first: What is a world model

Unlocks: JEPA

Taxonomy adapted from the 2026 “three research programs” framing in the world-models literature; exemplar years and roles verified against the primary papers named on later pages.