World Models: Understanding Dynamic Systems

Three properties distinguish it from a predictive text system. First, it maintains a state: a compressed summary of everything about the situation that matters for what happens next. Second, it contains dynamics: a function that maps the current state and an action to a distribution over next states. Third, it supports rollout: the same function applied repeatedly, cheaply, to explore futures.

This is precisely the structure set out in LeCun's position paper on autonomous machine intelligence, and it is the structure implemented empirically by Ha and Schmidhuber: an encoder compresses observations, a recurrent model predicts forward, a decoder reconstructs, and a small controller acts on the latent state rather than on raw input.

Raw observations are too large and too noisy to plan over directly. An encoder maps them into a latent vector that is meant to preserve what is relevant for predicting the future. This compression is lossy by design: the model keeps what its training objective rewards and discards the rest.

• Fully observable systems (chess, a closed ledger) permit a state that is essentially the position itself.
• Partially observable systems require the state to include a belief: what the system thinks is true but cannot see.
• The choice of state fixes the horizon over which the model can be useful. Add the wrong variable and the error is structural, not statistical.

In a deterministic environment the transition is a function. In most real environments it is a distribution, because the system is noisy, because some variables are unobserved, or because other agents act. A world model therefore predicts a distribution — and the planner has to reason over it rather than over a single forecast.

• Stationary dynamics can be learned from historical data and trusted for a while.
• Non-stationary dynamics require explicit regime detection, because the training distribution stops describing the present.
• Adversarial dynamics (other agents reacting to us) break the assumption that the model is external to the system it describes.

• Latency: an action taken now may only register in the state many steps later, so credit must be assigned across a delay.
• Multiple timescales: settlement, liquidity and policy operate on different clocks, and a single step size cannot represent all of them well.
• Irreversibility: some transitions cannot be undone, which changes the planning objective from optimisation to risk-constrained optimisation.

Temporal reasoning is also where language models are structurally weak: they are trained to produce a locally coherent sequence, not to maintain a consistent causal ledger across thousands of steps. A world model addresses this by holding an explicit state instead of relying on the context window.

A useful simulation must be scored three ways: calibration (are the predicted distributions honest about their own certainty), fidelity at horizon (does the trajectory resemble reality at step N), and decision relevance (does acting on the simulation improve real outcomes). The third is the only one that matters commercially, and it is the one vendors rarely measure.

A counterfactual requires intervening on the model rather than observing it: hold everything else fixed, change one action, and roll forward. This is exactly what a learned transition function supports and what a language model cannot do — a language model can describe a counterfactual fluently, but it cannot compute one under controlled assumptions.

• Useful for: pre-mortems, stress design, policy comparison, capital allocation scenarios.
• Not useful for: attributing causation without an identification strategy. The model's counterfactual is only as good as its dynamics.

• Aleatoric uncertainty: irreducible randomness in the environment. More data does not remove it.
• Epistemic uncertainty: ignorance about the model or its parameters. More or better data can reduce it.
• Distribution shift: the present is outside the training distribution. This is the failure mode that kills deployed models quietly.

Practically: require the system to output a distribution or an interval, not a point. Then test whether the stated interval contains the outcome at roughly the stated rate. That test — calibration — is cheap, and almost never performed in vendor demonstrations.

This is where a planner can actively exploit model error: if the model is over-optimistic about an action, the planner will prefer it. The pathology is well known in model-based reinforcement learning, and the mitigation is either conservatism under uncertainty or periodic correction against real outcomes.

• Design rule: keep a human approval gate wherever an action is irreversible.
• Design rule: log the simulated branch that justified each real decision, so it can be audited later.

A generative video model can be visually convincing while being causally wrong — it optimises appearance. A chatbot can be articulately wrong — it optimises plausible text. A financial world model has to be right about what happens next, under intervention, on data it has never seen. That is a different and much harder objective.

• Supported: a compact learned world model can substitute for an environment during policy learning (Ha & Schmidhuber, 2018).
• Supported: a single world-model architecture with fixed hyperparameters has solved control tasks across many domains (Hafner et al., Nature, 2025).
• Supported in a different domain: language-model agents can produce believable individual and emergent social behaviour in a simulated sandbox (Park et al., 2023).
• Not supported: that any of these results implies reliable market-scale simulation.
• Not supported: that rollout error stays bounded over long horizons in open, adversarial environments.