CoGR co-trains query and item LLMs, gains 36.1% F1 on WANDS

CoGR, a generative-retrieval framework introduced in an arXiv preprint by Runpeng Dai, Kaili Huang, Changsung Kang and Ciya Liao, reports a 36.1% F1 gain over the strongest of ten baselines on the public WANDS shopping benchmark, and a 10.9% lift on an internal APP Marketplace dataset.

The twist is where the LLMs sit. "The generative component is typically used for query-side augmentation, while final matching is still delegated to a downstream retriever," the authors write. CoGR instead "trains LLMs to directly construct retrieval representations on both query and item sides." Each generator emits a compact keyword set that plugs into an inverted index, so the approach stays "compatible with existing keyword-based retrieval infrastructure."

Training runs in two stages. Supervised fine-tuning first aligns the keyword space; then co-evolving reinforcement learning alternates GRPO updates on the query- and item-side generators against the opposite side's frozen index. Both sides optimize the same query-to-item F1 objective, but asymmetrically: the query side "receives retrieval F1 directly, while the item side receives a counterfactual marginal reward measuring the change in query-side F1 caused by its generated keywords."

The abstract does not break out per-baseline numbers, inference cost, or how the item-side generator is refreshed as a catalog churns. The comparison covers ten sparse, dense and generative baselines; the paper claims CoGR is best on both datasets.