Recipes for training embeddings and for distillation landed the same day, pushing retrieval quality down into everyday practice.
Hugging Face published a guide to training and finetuning multi-vector embedding models with Sentence Transformers, alongside Apple's PROOF-Gen work on data-optimized distillation. Of five driving items, two come from labs and three from arXiv.
Effort is going into sharpening existing representations rather than rebuilding models. The idea that retrieval and RAG quality hinges on embedding training is now circulating as official how-to material. The three academic items sit further afield, so the cluster's cohesion is only moderate.
Whether multi-vector embeddings show up in production search, and whether the distillation recipe draws reproduction reports. Adoption as a library default would be the durable signal.