Matryoshka Representation Learning

Nested, variable-dimension embeddings that pack coarse-to-fine information into prefixes of a single vector, so one model serves many embedding sizes without retraining. Led by Aditya Kusupati in UW's RAIVN lab (with Google). Though authored before the tracking window, MRL became the industry standard for adaptive embeddings squarely in-window — OpenAI's text-embedding-3 and most 2024–2026 embedding providers adopted Matryoshka-style truncatable dimensions.