Hallucination in World Models is Predictable and Preventable

TL;DR

Hallucinations in world models stem from data coverage gaps and can be predicted and prevented using the same signals that detect them.

Problem

Modern visual world models generate realistic futures but often hallucinate by drifting from ground-truth dynamics. Understanding where and why these hallucinations occur is a key challenge.

Approach

Results & Contribution

Reveals that hallucination is fundamentally a data coverage issue and demonstrates that the same detection signals can be used for mitigation. The method adapts pretrained models to unseen environments with as few as 50 real trajectories.