What is a world model?
A world model is a learned internal model of how an environment evolves — one that an agent (or a generator) can use to represent the present state of the world and predict future states, optionally conditioned on actions. The essential purpose, in the framing of the canonical ACM Computing Surveys treatment (Ding et al.), is “to understand the dynamics of the external world and compute the next state.”
The idea is much older than the current wave. Kenneth Craik proposed in 1943 that organisms carry a “small-scale model” of external reality in their heads; Schmidhuber built neural world models in 1990; Sutton’s Dyna architecture (1991) interleaved learning a model with planning inside it. But the modern usage of the term was revived and popularized by Ha & Schmidhuber’s 2018 “World Models” (arXiv:1803.10122, published at NeurIPS 2018 as Recurrent World Models Facilitate Policy Evolution) — an RNN-based model that learned compressed latent representations of its environment, inside which a tiny policy could be trained entirely in the model’s own “dream.”
From there, the survey literature treats a handful of systems as canonical milestones:
- PlaNet (2019) — latent dynamics for planning from pixels
- The Dreamer family (2020–2025) — actor-critic learning inside imagined rollouts
- MuZero (2020) — learned dynamics for planning without given rules
- Sora (2024) — video generation marketed as a “world simulator”
- Cosmos (2025–2026) — NVIDIA’s open world-model family for physical AI
- Genie (2024–2025) — generative interactive environments from unlabeled video
The interesting thing about this list is that it spans at least three research communities (model-based RL, video generation, and robotics) that mostly did not talk to each other until “world model” became the term that unified them — and, as the taxonomy section shows, they still disagree about what the term means.