📝 SummarXiv

Abstract

Generative cross-modal retrieval, which treats retrieval as a generation task, has emerged as a promising direction with the rise of Multimodal Large Language Models (MLLMs). In this setting, the model responds to a text query by generating an identifier corresponding to the target image. However, existing methods typically rely on manually crafted string IDs, clustering-based labels, or atomic identifiers requiring vocabulary expansion, all of which face challenges in semantic alignment or scalability.To address these limitations, we propose a vocabulary-efficient identifier generation framework that prompts MLLMs to generate Structured Semantic Identifiers from image-caption pairs. These identifiers are composed of concept-level tokens such as objects and actions, naturally aligning with the model's generation space without modifying the tokenizer. Additionally, we introduce a Rationale-Guided Supervision Strategy, prompting the model to produce a one-sentence explanation alongside each identifier serves as an auxiliary supervision signal that improves semantic grounding and reduces hallucinations during training.