Scientific search engines have a rare topic problem, and a new framework called REPAIR claims to fix it by making retrieval systems check their own facts.
Researchers built REPAIR, a self-evolving data augmentation framework for scientific dense retrievers. Dense retrievers are the systems that power semantic search over research papers, matching queries to relevant documents by meaning rather than keywords. The problem is long-tail concepts, obscure terms and fine distinctions that show up too rarely in training data for models to learn well. REPAIR cycles through three steps: diagnosing which long-tail concepts a retriever confuses, pulling in API-guided evidence to expand coverage of those concepts, and mining hard negatives (near-miss documents) to sharpen the model's ability to tell similar concepts apart. Across nine materials science and biomedical benchmarks, it beat 19 existing baselines.
Scientific literature keeps growing, and a search tool that muddles fine-grained distinctions can send a researcher down the wrong path or bury the right paper under irrelevant hits. Most retrieval improvements lean on bigger models or more raw data. REPAIR's contribution is narrower: it pinpoints exactly where a retriever is weak and manufactures training examples to patch that specific gap, instead of betting that scale alone will fix it.
That's a modest fix for a modest slice of the problem, but rare compounds and obscure gene variants are exactly where scientific search tends to fail, so aiming small here looks like the right call.