arxivcs.CV2026-07-08
MMAgent-R$^2$: Learning to Rerank and Reject for Agentic mRAG
Tao Zhang, Ziqi Zhang, Zongyang Ma, Yuxin Yang, Bing Li, Chunfeng Yuan, et al.
Knowledge-based Visual Question Answering (KB-VQA) requires models to retrieve visual entities matching the query image from large-scale encyclopedic knowledge bases and answer related questions. Existing multimodal Retrieval Augmented Generation (mRAG) methods rely on global vis…