The authors present CoGR, a retrieval framework in which large language models generate keyword representations for both queries and items, matched through an inverted index rather than dense vector similarity. A two-stage training pipeline combines supervised fine-tuning with co-evolving reinforcement learning, where the query and item generators alternately train against each other’s frozen index. On an internal marketplace dataset and the public WANDS benchmark, CoGR improved F1 over the strongest baseline by 10.9% and 36.1% respectively.
