Generative world models hallucinate in predictable ways tied to data coverage, and a single data-centric intervention can prevent all three failure modes simultaneously.
Characterizing Hallucination
We identify three distinct ways in which a generative world model may hallucinate, each tied to a different stage of the imagination pipeline.
Action-marginalized hallucination. Conditional on a context, the predicted next latent is largely insensitive to the input action. The rollout is visually plausible but collapses onto an action-marginalized future, so the model behaves more like a video generator than a controllable world model. Operationally, we expose this mode by intervening on the action stream at evaluation time, e.g. by randomly shuffling actions within a batch and measuring the resulting change in flow MSE; a model that hallucinates in this way is one whose flow MSE barely moves under the intervention.
Scene-diverging hallucination. It is well understood that autoregressive rollouts accumulate compounding error as the prediction horizon increases. However, scene-diverging hallucination is a very specific failure mode where physically implausible events (such as a ball teleporting back into play when scoring in Pong) are predicted. This type of hallucination is most frequent in states with poor data coverage.
These three types of failure modes probe disjoint pieces of the model: the tokenizer, the action-conditioning of the dynamics model, and the multi-step accumulation of dynamics error.
Mitigating Hallucination
The taxonomy and predictors above suggest a single data-centric lens on hallucination: each of the three failure modes are, mechanistically, a consequence of the model having seen too little of some region of the state-action space. A perceptual hallucination is a coverage gap in the tokenizer's reconstruction distribution, an action-marginalized hallucination is a coverage gap in action-conditional transitions, and a scene-diverging hallucination is a coverage gap along the trajectory the model is asked to imagine.
Two interventions follow naturally:
Coverage-aware training. We propose to resample the existing dataset to upweight under-represented regions of the state-action space, then ask whether closing those gaps at training time reduces all three failure modes simultaneously. Because the lens above identifies coverage as the underlying lever for every type of hallucination, a single reweighting recipe is expected to move all three signals in the right direction at once, rather than requiring a separate intervention per type. In practice, we rebalance training data by adjusting sampling to be uniform across tasks rather than frames. We also experiment with loss re-weighting but find interventions in sampling to be superior.
Frequently Asked Questions
What are the three types of world model hallucination? The three types are perceptual hallucination (reconstruction gap), action-marginalized hallucination (action-conditioning failure), and scene-diverging hallucination (accumulated error from poor trajectory coverage).
How can hallucination be predicted? Hallucination is predictable through measurable indicators tied to data coverage — checking how action-insensitive the latent dynamics are or detecting when physically implausible events appear in rollouts.
What is coverage-aware training? It is a data-centric intervention that reweights training samples to upweight under-represented regions of the state-action space, addressing all three hallucination types with a single recipe.
Is coverage-aware training more effective than loss re-weighting? Yes, experiments show that adjusting sampling to be uniform across tasks rather than frames is superior to loss re-weighting for preventing hallucination.