UniIR: Training and Benchmarking Universal Multimodal Information Retrievers

Instruction-guided retrieval across modalities: score-level versus feature-level fusion, how the M-BEIR benchmark was assembled, and what the experiments say about instructions and the two fusion designs.