Hallucination in World Models is Predictable and Preventable

Nicklas Hansen and Xiaolong Wang at UC San Diego show that hallucination in modern world models is not mainly an architectural flaw but a data-coverage problem: it concentrates in low-coverage regions of state-action space and can be predicted from internal signals. Using their new MMBench2 benchmark, a 427-hour, 210-task corpus with ground-truth actions, rewards, and live simulators, they train a 350M-parameter Dreamer 4 world model and separate hallucination into three failure modes: perceptual errors in the tokenizer, action-marginalized transitions in the dynamics model, and scene-diverging rollouts. They then introduce three label-free predictors—the tokenizer round-trip residual, flow instability, and inter-seed variance—that correlate strongly with open-loop rollout error, with Spearman ρ around −0.80 across all three. For mitigation, they reweight sampling to be uniform across tasks rather than frames, which improves rollout fidelity by 0.88 dB and lowers all three normalized hallucination signals; they also use the predictors as curiosity rewards for online data collection, where only 50 real trajectories per task adapt the pretrained model to entirely unseen environments. On 10 unseen tasks, closed-loop MPC performance rises from 0.276 to 0.325 with curiosity-driven data, compared with 0.118 for a random policy and 0.362 for expert or human data, showing that predictive hallucination signals can be turned directly into data-efficient control improvements.