Hey everyone, Matryoshka Representation Learning (MRL) has gained a lot of traction for its ability to maintain strong downstream performance even under aggressive embedding compression. That said, I’m curious about its limitations. While I’ve come across some recent work highlighting degraded perfo
