HyBIRD uses hyperbolic geometry to explain method inspiration retrieval

Methodology Inspiration Retrieval (MIR) requires systems to find prior papers whose methods can inspire new research proposals—a fundamentally different challenge from traditional topical search. Existing dense retrievers provide strong rankings but fail to expose how a proposal's needs are bridged by retrieved methods or where evidence is weak. To solve this, researchers introduce HyBIRD, a frozen-anchor framework that treats MIR as hyperbolic bridge retrieval and post-hoc diagnosis. HyBIRD keeps a strong MIR dense retriever fixed, then learns three lightweight hyperbolic variants: point, cone, and factorized bridges. It then uses LLM-assisted method blocks to generate explanations and select complementary evidence snippets.

The factorized bridge variant achieves 59.034 mAP on the MIR benchmark while preserving the dense anchor's retrieval behavior. More importantly, HyBIRD converts ranked papers into inspectable query need profiles, factor coverage and maturity views, and complementary evidence bundles. The results suggest hyperbolic geometry is most useful as calibrated structure atop a dense anchor, rather than a standalone replacement. This opens the door to more transparent research discovery tools that show not just what papers are relevant, but why their methods can be applied to new problems—a key step toward AI-assisted scientific creativity.