Matryoshka Representation Learning

The authors demonstrate MLR, which can encode information at different granularities allowing a single embedding to be be used for different downstream tasks.