MIMO retrieval model beats baselines in multilingual search tasks

Multilingual Information Retrieval (MLIR) remains challenging because queries and documents may appear in different languages within the same corpus. Existing embedding models are optimized for mono-lingual or multi-monolingual retrieval and often degrade in MLIR settings. Directly applying contrastive learning can exacerbate language clustering and hurt cross-lingual alignment. To address this, researchers Youngjoon Jang, Seongtae Hong, and Heuiseok Lim present MIMO (Multilingual Information Retrieval via Monolingual Objectives).

MIMO is a two-stage framework that first initializes a student model's cross-lingual alignment by distilling knowledge from a high-performing English-only teacher model, creating a stable English semantic space as an anchor. In the second stage, it jointly optimizes both the distillation loss and a cross-lingual contrastive learning objective to improve retrieval discrimination while preserving alignment. Experiments across diverse MLIR and multi-monolingual benchmarks show that MIMO consistently outperforms existing cross-lingual training baselines. Moreover, it remains competitive with off-the-shelf models of similar or larger size. The paper also provides an Alignment-Uniformity analysis, clarifying the distinct roles of the two loss components and demonstrating that their combination yields a favorable trade-off.