GLInt is a late-interaction retrieval model that achieves a mean nDCG@10 of 57.43 on the BEIR benchmark by training on geometry-matched hard negatives. The research finds that multi-vector mining produces harder negatives than dense retrieval but also introduces more false negatives, requiring careful filtering and candidate pool management. It further shows that the breadth of training data across multiple sources matters more for knowledge distillation quality than the choice of teacher model.
