We have reached an era where AI can generate stunningly realistic video simulations of the future, yet these models are fundamentally broken. They produce visually fluent rollouts that drift dangerously far from reality, creating a deceptive illusion of competence. This is not a glitch in the rendering engine but a systemic failure in how these models perceive the world.

Researchers have identified that these hallucinations are not random noise but predictable failures concentrated in low-coverage regions of the state-action space. In simpler terms, when the AI encounters situations it has rarely seen during training, it starts guessing wildly. To prove this, the team introduced MMBench2, a massive dataset comprising 427 hours of visual world modeling data with ground-truth actions, rewards, and live simulators. They trained a 350-million-parameter model on this data to dissect exactly where and why these errors occur.

The study reveals three distinct modes of failure: perceptual errors, action-marginalized drifts, and scene-diverging hallucinations. Each mode is anchored to a different stage of the AI’s processing pipeline. Crucially, the researchers developed specific signals that can accurately predict where the model will fail before it happens. This transforms hallucination from an unpredictable bug into a manageable data coverage issue.

The most surprising consequence is that the same signals used to detect these errors can also fix them. By using hallucination predictors as curiosity rewards, the team created a targeted data collection method. This allows for a data-efficient fine-tuning recipe that adapts the pretrained model to entirely unseen environments using as few as 50 real environment trajectories. The takeaway is clear: if you want to stop AI from hallucinating, you do not need a bigger model. You need better data coverage and the right signals to find the gaps.