Beyond Contrastive: Generative Reranking for Multilingual Text-Based Person Retrieval

Abstract

Current text-based person retrieval (TBPR) systems are typically built on English-pretrained vision-language backbones and evaluated on English-only, single-domain benchmarks. This setting limits their applicability to multilingual queries and encourages reliance on dataset-specific shortcuts, resulting in weak cross-domain generalization. To address these issues, we propose GRAM (Generative Reranking and Alignment for Multilingual text-based person retrieval), a compact MLLM-based TBPR framework built around a generative paradigm. By jointly optimizing contrastive retrieval and caption generation, the model is encouraged to encode fine-grained visual attributes rather than holistic image-text alignment, yielding more transferable representations across languages. We introduce cross-lingual knowledge distillation to transfer retrieval capability from English to eight target languages. This distillation stage improves English retrieval, suggesting a multilingual regularization effect. During inference, the decoder is reused to rerank top-K candidates with conditional generation probability, providing fine-grained verification with low additional latency. We also construct M-PEDES, a pioneering multilingual TBPR benchmark with eight target languages and stratified human auditing of translation quality. Extensive experiments demonstrate zero-shot multilingual transfer, further gains from distillation, and improved cross-dataset generalization over strong English and multilingual baselines.