Multimodal LLM-guided Query Optimization for Visual-Language Retrieval

GuangHao Meng, Tao Dai, Letian Zhang, Jieming Zhu, Jinpeng Wang, Qing Li, Yong Jiang

Keywords: vision-language retrieval, cross-modal retrieval, prompt engineering, query rewriting, large language model

TL;DR: This paper introduces the multimodal LLM guided query rewriter (MGQRe), which uses MLLMs to guide the rewriter to generate high-quality queries that match the retriever's perference, thus improving the performance of vision-language retrieval.

Abstract: Vision-language retrieval (VLR), involving the use of text (or images) as queries to retrieve corresponding images (or text), has been widely used in multimedia and computer vision tasks. However, ambiguous or complex concepts contained in queries often confuse retrievers, making it difficult to effectively align these concepts with visual content, thereby limiting their performance. Existing query optimization methods neglect the feedback of retrievers' preferences, thus resulting in sub-optimal performance. Inspired by the powerful ability of Multimodal Large Language Models (MLLMs), we propose a Multimodal LLM-Guided Query Rewriter (MGQRe) for query optimization. Specifically, MGQRe first utilizes MLLM to explore the retriever's weakness and perform targeted iterative optimizations to capture the retriever's expressive preferences. Subsequently, we develop a trainable rewriter that learns this preference knowledge through a three-step tuning strategy: supervised fine-tuning, preference learning, and reinforcement learning. This ensures that the queries generated by the rewriter align with the retriever’s preferences, thereby enhancing the retriever's performance. Extensive VLR benchmark experiments have demonstrated the superiority of MGQRe, as well as its generalizability and transferability. This work showcases the potential of using advanced language models to overcome the inherent limitations in current VLR technology.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13945