World Model
An internal model of how things change, learned well enough to imagine what happens next — and the leading candidate for what current AI is missing.
Reviewed July 16, 2026Stable
When not to use it
- As an explanation for anything a language model does. Whether LLMs have world models is the open question, not the answer, and using the phrase settles it by assertion.
- Where you have a real simulator. If physics is known, write the physics — a learned approximation of a thing you can compute exactly is strictly worse.
- Anything requiring accuracy over long horizons. Learned dynamics compound error, and a long imagined rollout is fiction.
Reach for something else instead
- Model-free RL skips the simulator and learns the policy directly — simpler, more sample-hungry, and no dream to drift.
- An explicit simulator beats a learned one wherever the rules are known and codeable.
- Retrieval solves the version of this problem people usually mean in LLM contexts: the model didn't lack a world model, it lacked the fact.
Sources & further reading
- Ha & Schmidhuber (2018), World Models — the canonical formulation; a controller trained entirely inside a learned simulation.
- Hafner et al. (2023), Mastering Diverse Domains through World Models — DreamerV3; learning in imagination across 150+ tasks with fixed hyperparameters.
- LeCun (2022), A Path Towards Autonomous Machine Intelligence — the architectural case for predicting in representation space, and the argument that current LLMs lack this.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Using "world model" to mean "the model knows things". It's a specific claim about learned dynamics you can roll forward, not a synonym for knowledge.
- Trusting long imagined rollouts. Learned dynamics compound error, so a policy trained deep in a dream is optimised against a world that has drifted from the real one.
- Treating the LLM-world-model debate as settled in either direction. It isn't, by anyone, and the confidence on display substantially exceeds the evidence available.
At a glance
FieldFoundations
Core idealearned dynamics you can roll forward
Working today inmodel-based RL
The debateis this what LLMs lack?
DifficultyAdvanced