Optimizing Retrieval-Augmented Generation through Adaptive Rewrite Selection

ACL ARR 2025 May Submission6501 Authors

Abstract:

With the advancement of Retrieval-Augmented Generation (RAG) in open-domain question answering (ODQA), query rewriting has gained increasing attention as a means to better handle complex queries. By generating alternative formulations of a question, query rewrites can help bridge the gap between user intent and the structure of retrieved knowledge, thereby enhancing multi-hop reasoning.
However, existing approaches often produce static rewrites that lack adaptability and fail to capture the evolving intent behind complex queries. To overcome this challenge, we propose ARS-RAG, an adaptive rewrite selection approach to dynamically determine the optimal number of rewrites for each query. ARS-RAG generates multiple rewrites for a given query and dynamically selects the effective ones. We train a self-supervised ranker to assess the relevance of each rewrite, as well as a contextual bandit selector that dynamically selects the optimal top-$K$ rewrites. This enables query-specific adaptation and efficient retrieval. Experimental results on four ODQA datasets confirm the effectiveness of ARS-RAG. Importantly, our adaptive selection strategy introduces negligible overhead and requires no additional fine-tuning of the rewriter.

Paper Type: Long

Research Area: Generation

Research Area Keywords: retrieval-augmented generation; open-domain QA; large language models

Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 6501