Learning from Failures: Retrieval-Centric CoT via Hard Negatives for Unified Multimodal Retrieval
UniME-R1 is an embedder-adviser framework for unified multimodal retrieval that conditions reasoning on initial retrieval results rather than on the query alone. The adviser examines candidates to identify fine-grained cues missed by the embedder. It reranks directly when the target is in the initial top-k; otherwise, it generates retrieval-centric chain-of-thought guidance and runs full-corpus retrieval again with a dual-mode embedder. Training uses mined hard negatives, supervised learning, and retrieval-oriented reinforcement learning. The supplied abstract reports consistent gains on MMEB-V2 and other benchmarks but provides no numerical results.
Why it's worth reading
It reframes chain-of-thought as diagnosis of retrieval failures, offering a concrete design direction for multimodal search and RAG systems that struggle with semantically similar candidates.