World models can hallucinate the future, and until now nobody could catch it while the system was running.
Researchers studied frozen, self-supervised world models, the systems that predict what a scene will look like next given a current state and an action. They found these models sometimes predict a next frame that looks statistically normal but never actually happens, a hallucination that the model itself cannot distinguish from a correct guess. Because world models feed their own predictions back in as the basis for the next prediction, one bad guess compounds silently over time instead of being caught and discarded. The team built MEND, a single neural network trained by denoising score matching, that flags a hallucination, localizes it to the exact image patches responsible, and nudges the latent prediction back toward something real - all without ever training on labeled examples of errors.
World models are the technology underpinning robotics planning, self-driving simulation, and video-generation systems, and they are increasingly pitched as a foundation layer the way language models became for text. If a model can silently drift into a fabricated scene while reporting normal confidence, that is a problem wherever its output feeds a real decision, like a robot planning a grasp. On two navigation test environments MEND caught hallucinations with an accuracy (AUROC) up to 0.80 and localized them to specific patches with an AUPRC up to 0.87, respectable numbers but short of reliable.
Call it a smoke detector, not a fire extinguisher: it is better at telling you something is wrong than at putting it out, and the researchers themselves note some of the error sits too close to normal data to fully correct.