ReFeed: Retrieval Feedback-Guided Dataset Construction for Style-Aware Query Rewriting

A research team from KAIST and Seoul National University has introduced ReFeed, a novel framework for constructing datasets to train style-aware query rewriting models. The core problem they address is the frequent failure of retrieval systems when user queries differ stylistically or semantically from the language used in domain documents. While query rewriting has been proposed as a solution, most existing approaches overlook the crucial stylistic characteristics of target documents—their domain-specific phrasing, tone, and structure. ReFeed offers a data-centric solution by creating a feedback loop that automatically identifies retrieval failures, leverages large language models to rewrite queries to match document style, and verifies improvements through re-retrieval.

The resulting corpus of (original, rewritten) query pairs enables the training of specialized rewriter models that are explicitly aware of both document style and retrieval feedback. This work, accepted at the AAAI 2026 Workshop on New Frontiers in Information Retrieval, represents a significant shift toward feedback-driven, data-centric approaches in IR. By focusing on aligning query language with real-world document distributions, ReFeed enhances the reasoning and adaptability of RAG (retrieval-augmented generation) systems, particularly in specialized domains like legal, medical, or technical documentation where stylistic alignment is critical for accurate information retrieval.