RADD: Retrieval-Anchored Discrete Diffusion for Multi-Modal Knowledge Graph Completion
RADD proposes a framework to improve multi-modal knowledge graph completion by separating retrieval and reranking. It uses a relation-aware multimodal retriever to find relevant candidates and a conditional discrete denoiser to refine predictions locally within that shortlist. This decoupling achieves state-of-the-art results by optimizing global recall during retrieval and local precision during denoising.