ReCQR: Incorporating conversational query rewriting to improve Multimodal Image Retrieval

A team of researchers has published a paper on ReCQR, a novel method that addresses a core weakness in modern image search: vague, conversational user queries. Existing multimodal retrieval systems often struggle when users ask long, ambiguous questions or reference previous parts of a dialogue. ReCQR tackles this by introducing a conversational query rewriting (CQR) task, where an AI agent takes a user's final, messy request within a full chat history and rewrites it into a clear, standalone prompt optimized for retrieval.

The researchers first built a dedicated dataset to train and test this approach. They leveraged Large Language Models (LLMs) to generate candidate rewritten queries at scale, then used an 'LLM-as-Judge' mechanism combined with manual review to curate approximately 7,000 high-quality, multi-turn multimodal dialogues. This dataset, named the ReCQR dataset, provides a new benchmark for the field.

Experiments benchmarking state-of-the-art multimodal models on this new dataset show that the CQR technique delivers significant gains. By clarifying user intent and distilling complex dialogue into concise semantics, it enhances the accuracy of traditional image retrieval models. The work provides a new framework for modeling real-world user interactions in AI systems, moving beyond single, perfect queries to handle the messy reality of human conversation.