Tessera embeddings remain markedly more label-efficient than from-scratch models even when classes are temporally stable

On both datasets, Tessera embeddings remain markedly more label-efficient.

http://arxiv.org/abs/2608.27175v1