QUESTER: Query Specification for Generative Retrieval

2025-11-07

AI Summary

Generative retrieval (GR) shows promise but suffers from poor generalization and high scalability costs. This paper proposes QUESTER, the first framework to reformulate GR as a query semantic generation task: it employs a lightweight language model to generate BM25-compatible keyword queries and introduces GRPO, a novel reinforcement learning algorithm, for end-to-end joint optimization. By generating retrievable keywords instead of document IDs, QUESTER circumvents the generalization bottleneck inherent in direct ID generation, thereby enhancing cross-domain robustness and computational efficiency. Experiments demonstrate that QUESTER outperforms BM25 on both in-domain and zero-shot cross-domain retrieval tasks, matching the effectiveness of state-of-the-art neural retrievers while offering low latency and straightforward deployment. Its core innovation lies in bridging generative paradigms with traditional retrieval infrastructure through semantically grounded, BM25-compatible query generation.

Technology Category

• Natural Language Processing: Generation
• Search and Optimization: Learning to Search
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Querying, indexing, and retrieval in Web-related graphs

Abstract

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

Problem

• Generative Retrieval struggles with generalization and scaling costs
• It reframes retrieval as query specification using small LLMs
• Trains policy with reinforcement learning for competitive neural IR performance

Innovation

• Generative retrieval using reinforcement learning training
• Query specification generation with small LLM
• Competitive effectiveness with neural IR models