Researchers have developed LACONIC, a family of learned sparse retrievers that can efficiently search large datasets using commodity CPU hardware. Unlike dense retrieval models, which require high memory and GPU acceleration, LACONIC uses inverted indices to enable fast search. The team proposes a two-phase training curriculum, first adapting causal language models for bidirectional contextualization and then fine-tuning with curated hard negatives. They demonstrate that LACONIC can achieve state-of-the-art performance on the MTEB Retrieval benchmark while using significantly less memory than dense models. The code and trained checkpoints are open-sourced to facilitate reproducibility. --- Why it matters: This matters because it provides a scalable and efficient solution for real-world search applications, which is particularly important as datasets continue to grow in size and complexity. Source: https://arxiv.org/abs/2601.01684