Learning from Failures: Retrieval-Centric CoT via Hard Negatives for Unified Multimodal Retrieval
UniME-R1 makes the retriever explain what it got wrong, then uses that feedback to search again.
The paper argues that query-only CoT misses the cues that caused similar candidates to be confused. Its adviser looks at the initially retrieved set, identifies discriminative failures, and either reranks the top-k or generates retrieval-centric CoT for a full-corpus re-retrieval. Training uses mined hard negatives, supervised learning, and retrieval-oriented reinforcement learning. The authors report consistent gains on MMEB-V2 and other multimodal retrieval benchmarks. HF Daily Papers' note
The paper argues that query-only CoT misses the cues that caused similar candidates to be confused. Its adviser looks at the initially retrieved set, identifies discriminative failures, and either reranks the top-k or generates retrieval-centric CoT for a full-corpus re-retrieval. Training uses mined hard negatives, supervised learning, and retrieval-oriented reinforcement learning. The authors report consistent gains on MMEB-V2 and other multimodal retrieval benchmarks. HF Daily Papers' note
score 4