REPAIR is a self-evolving augmentation framework designed to resolve long-tail concept confusion in scientific retrieval. It uses API-guided evidence expansion and hard negative mining to synthesize training data, outperforming 19 baselines on materials science and biomedical benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can improve RAG performance in specialized domains by iteratively augmenting data for long-tail concepts.
●
researcherThis method demonstrates the efficacy of fact-verified iterative refinement for dense retrievers.