Hallucination in World Models is Predictable and Preventable

Hansen and Wang reframe hallucination in visual world models as a data coverage problem rather than a model capacity problem. Three failure modes are identified: perceptual, action-marginalized, and scene-diverging. Three model-internal signals are derived that predict hallucination with approximately -0.80 Spearman correlation. They introduce MMBench2, a 427-hour 210-task dataset with ground-truth actions and rewards. Coverage-aware training and curiosity-reward fine-tuning enable adaptation to new environments with as few as 50 trajectories. 41 upvotes on HF Daily Papers.