It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning
CoGR trains LLMs to generate the keyword representations used for retrieval on both the query and item sides.
The framework keeps the output compatible with inverted-index retrieval by matching compact keyword sets directly. It first uses supervised fine-tuning to align the keyword space, then alternates reinforcement learning between query and item generators. The paper reports best results against 10 sparse, dense, and generative baselines, with F1 gains of 10.9% on an internal APP Marketplace dataset and 36.1% on WANDS. HF Daily Papers' note
The framework keeps the output compatible with inverted-index retrieval by matching compact keyword sets directly. It first uses supervised fine-tuning to align the keyword space, then alternates reinforcement learning between query and item generators. The paper reports best results against 10 sparse, dense, and generative baselines, with F1 gains of 10.9% on an internal APP Marketplace dataset and 36.1% on WANDS. HF Daily Papers' note
score 4