What is a world model?

Eighty years after Craik proposed that the brain carries "small-scale models" of external reality, the term world model remains formally undefined. The field has been building prediction engines and calling them world models.

A world model is state that persists — Distinctions, Entities, Relations — on an axis of Time, moved by Transition Logic and bounded by Constraints, and a geometry of evolution: Observation, Uncertainty, Purpose, and Memory. Ten ingredients. Prediction reaches three; the other seven are what make a model a world.

Ten requirements for a world to be modelable.

StateWhat persists
TimeAxis of evolution
ChangeRules of motion
ConstraintLimits of motion
Geometry of evolutionFrom whom, with what knowing, toward what end

Six leading definitions, evaluated against the same ten ingredients.

In our published assessment of six leading definitions of a world model, none carries purpose as a structural part of the model.

IngredientCraik1943 · Mental modelHa / Schmidhuber2018 · V + M + CLeCun2022 · JEPADreamZero2026 · World action modelMeta VLWM2025 · Vision-languageConsensus2025–26 · Represent + predict + act
01Distinctionsimpl.Implicit — the model distinguishes aspects of reality●VAE learns to distinguish features in latent space●JEPA learns to separate relevant from irrelevant●Visual scene decomposition●Multimodal scene understandingimpl.Subsumed under "represent"
02Entitiesimpl.Implicit — external reality contains things◐Latent codes, but no persistent identity across episodes◐Representations without persistent named identity◐Object reps, but identity is frame-level◐Referenced in language, no persistent formal identityimpl.Subsumed under "represent"
03Relationsimpl.Not formalized○Latent space is flat; no relational structure○Embedding space — flat, not relational◐Spatial relations implicit in visual scene◐Language captures some relational structureimpl.Subsumed under "represent"
04Time●Future situations before they arise●RNN models temporal sequence●Hierarchical JEPA predicts across timescales●Temporal prediction through video sequences●Multi-step planning across temporal horizonimpl.Subsumed under "predict"
05Transition Logic●Core — try out various alternatives●M predicts state transitions in latent space●Core state-transition prediction●Physics learned from video — strong transition prediction●Four-component prediction at each step●Core of "predict future states"
06Constraints○No concept of what is prohibited○No concept of prohibited states or invariants◐Cost Module encodes preferences, not constitutive limits○Physics is statistical, not invariant◐Task-specific penalties or guard-rails — acknowledged but unspecified○
07Observation◐Acknowledges limited senses in surrounding text●V is explicitly a partial observer●Explicitly partial; JEPA designed around partiality●Camera-perspective partial observation●Vision-language input is inherently partial○Model assumed to observe correctly
08Uncertainty○Model treated as veridical◐MDN outputs mixture-density distributions●Latent variable z represents what cannot be predicted from context○Generates visual futures without confidence○Not formally represented○
09Purpose◐Conclude which is the best — goal presupposed, not formalized○Reward is external environment signal◐Cost Module — designer-specified, not constitutive○Intent arrives as flat imperative sentences◐Goal description present; task-level, not constitutive○"Enable action" — action toward what?
10Memory○Present-tense; no accumulated history◐RNN hidden state decays; not queryable or persistent◐Short-Term Memory Module — working memory only○No persistent history; each episode independent○No history beyond current planning episode○Prediction is forward-looking; history not required
Completeness4/105/107/105/106/102/10
● Explicit◐ Partialimpl. Implicit○ Absent

“Model” is not a synonym for “trained artifact.”

Models come into existence through three mechanisms. Each produces properties the others lack. A complete world model requires all three — the field has built one leg of a three-legged stool.

M1Episodic

Training

Discover statistical structure from data.

Expose a parameterized function to data, optimize against a loss function, update parameters. The model discovers structure from the statistical properties of the data.

Produces
Statistical capability — distributional patterns, correlations, stylistic conventions, factual associations, reasoning patterns present in the training data.
Cannot
Structure not present in the data. What is prohibited as distinct from what is unlikely. Persistent identity across episodes.
Analogy
Learning a language by immersion.
Ingredient coverage
01 Distinctions
● Yes
02 Entities
◐ Partial
03 Relations
◐ Partial
04 Time
● Yes
05 Transition Logic
● Yes
06 Constraints
○ No
07 Observation
◐ Partial
08 Uncertainty
○ No
09 Purpose
○ No
10 Memory
○ No
M2Architectural

Constitution

Specify the grammar directly.

Specify the primitives, invariants, rules of composition. The model's architecture is given based on domain expertise — declared, not discovered.

Produces
Structural guarantees — authority chains, constitutive constraints, epistemic status tracking, constitutive purpose. Structure that holds because it was specified to hold.
Cannot
Content. Constitution builds the container; it does not populate it.
Analogy
Writing the grammar book.
Ingredient coverage
01 Distinctions
● Yes
02 Entities
● Yes
03 Relations
● Yes
04 Time
● Yes
05 Transition Logic
● Yes
06 Constraints
● Yes
07 Observation
● Yes
08 Uncertainty
● Yes
09 Purpose
● Yes
10 Memory
◐ Partial
M3Continuous

Accretion

Fill through governed operational capture.

Every decision, every state transformation, every boundary crossing deposits another layer of organizational truth. The model gets richer through accumulation of governed records.

Produces
Organizational truth — the accumulated record of what actually happened, who decided, under what authority, with what evidence, against what constraints.
Cannot
Structure. Accretion fills the container but does not build it.
Analogy
Accumulating a library in that language.
Ingredient coverage
01 Distinctions
◐ Partial
02 Entities
● Yes
03 Relations
● Yes
04 Time
● Yes
05 Transition Logic
◐ Partial
06 Constraints
◐ Partial
07 Observation
● Yes
08 Uncertainty
● Yes
09 Purpose
◐ Partial
10 Memory
● Yes
—TrainingConstitutionAccretion
TimescaleEpisodic (train, deploy)Rare (architectural events)Continuous (operational tempo)
ProducesStatistical capabilityStructural guaranteeOperational content
Cannot produceStructural guaranteesOperational contentStatistical capability
Evolves byRetraining on new dataAmending the specificationAccumulating through use
Error modeLearns the wrong patternsSpecifies the wrong structureCaptures garbage if ungoverned
The gap is not in the definitions. It is in the mechanism.

The field has built one leg of a three-legged stool. Training has produced extraordinary prediction engines. The systematic underinvestment in constitution and accretion explains why those engines cannot govern action, track epistemic status, or accumulate organizational memory. These are not capability gaps that more training will close.

Explore the WMI SeriesThe Conversation: who has asked what a world model is