Precise scientific retrieval is hampered by two intertwined issues: concepts follow a long‑tail distribution and the corpora are highly fact‑sensitive. These factors weaken dense retrievers and make large language model augmentation prone to hallucination.\ \ We introduce REPAIR, a self‑evolving data‑augmentation framework tailored for scientific dense retrieval. The system repeatedly executes three core operations to close knowledge gaps:\ \
- Long‑tail concept diagnosis – analyze retrieval failures and pinpoint missing rare concepts.\
- API‑guided evidence expansion – query external knowledge bases or tools to synthesize reliable text supporting the identified concepts.\
- Hard negative mining – select from the expanded evidence the most confusing yet incorrect samples to form contrastive pairs.\ \ These steps generate new training pairs that are fed back into the dense retriever, enabling iterative self‑refinement. By grounding the model in factual evidence, it learns fine‑grained distinctions and dramatically reduces hallucination.\ \ Extensive experiments on nine material‑science and biomedical benchmarks show REPAIR outperforms 19 strong baselines, highlighting the importance of diagnosing and fact‑augmenting long‑tail deficits for robust scientific retrieval.\ \ Review