QUESTER: Query Specification for Generative Retrieval
This work addresses efficiency and generalization challenges in retrieval systems for information retrieval applications, representing an incremental improvement.
The paper tackles the generalization and scalability issues of Generative Retrieval by introducing QUESTER, which reframes it as query specification generation using a small LLM trained with reinforcement learning. The model outperforms BM25 and is competitive with neural IR models in both in- and out-of-domain evaluations.
Generative Retrieval (GR) differs from the traditional index-then-retrieve pipeline by storing relevance in model parameters and directly generating document identifiers. However, GR often struggles to generalize and is costly to scale. We introduce QUESTER (QUEry SpecificaTion gEnerative Retrieval), which reframes GR as query specification generation - in this work, a simple keyword query handled by BM25 - using a (small) LLM. The policy is trained using reinforcement learning techniques (GRPO). Across in- and out-of-domain evaluations, we show that our model is more effective than BM25, and competitive with neural IR models, while maintaining a good efficiency