Preprint2026

Two new papers extending latent safety filters! In Anysafe we explore how to parameterized latent safety filters in order to flexibly specify constraints at deployment time. In our second work we study temperature-based constraints to investigate how partial observability can degrade the ability of latent safety filters to prevent safety violations. We quantify partial observability in the latent space of self-supervised world models using mutual information and provide a multimodal training strategy for effective filtering even under partial observability!