Tag
HappyWorld-Bench is a comprehensive benchmark that evaluates the reliability of world models under interaction and modification across video, spatial, and embodied tracks.
This paper introduces a method using Wasserstein barycenters to reconstruct language model embedding fields for predicting peer-misalignment penalties in spatial factor models, outperforming conventional weighting schemes.