Researchers propose DrIG for universal AI retrieval
DrIG uses dual-role identifiers to handle text, images, and mixed media 2x faster
DrIG is a new generative framework for universal multimodal retrieval that uses dual-role identifiers to handle text, image, and mixed image-text items across diverse tasks and domains. Each candidate gets a single residual-quantized identifier that works two ways: sequential decoding models modality first, then finer semantics, while a set-based reinterpretation provides a prefix-independent relevance prior to guide beam search and avoid local-optimum errors. Tested on M-BEIR and text-to-image datasets, DrIG consistently beats state-of-the-art generative multimodal baselines, and hybrid reranking delivers a favorable efficiency-effectiveness trade-off against strong dense retrievers. The paper also shows how base LMM, beam size, reranking depth, and fusion strategy impact performance, offering practical guidance for system design.
- DrIG uses dual-role identifiers to handle text, images, and mixed media with a single residual-quantized token
- Outperforms state-of-the-art generative multimodal baselines on M-BEIR and text-to-image datasets
- Hybrid reranking achieves better efficiency than dense retrievers while maintaining accuracy
Why It Matters
DrIG could revolutionize how AI systems retrieve and process multimodal data, enabling faster and more accurate search across text, images, and mixed formats.