Tag
This paper demonstrates that hallucinations in world models are a data coverage issue, and proposes data-centric signals to detect them and coverage-aware sampling techniques to mitigate them, using the new MMBench2 dataset.