The paper teaches small AI models to make high‑quality text embeddings by first copying a big expert model (distillation) and then practicing four jobs with special mini‑modules (LoRA adapters): retrieval, similarity, clustering, and classification.
The paper asks a simple question: if a language model becomes better at step-by-step reasoning (using RLVR), do its text embeddings also get better? The short answer is no.
This paper shows how to get strong text embeddings from decoder-only language models without any training.
Co2S is a new way to train segmentation models with very few labels by letting two different students (CLIP and DINOv3) learn together and correct each other.