ScholarCatalyst: A benchmark for retrieving papers that inspire new research

ScholarCatalyst evaluates whether AI systems can identify prior papers whose ideas could meaningfully advance an early-stage research project. Unlike conventional scientific retrieval benchmarks, ScholarCatalyst focuses on literature inspiration retrieval: identifying papers that may not be topically similar to a query, but offer the insight needed to formulate or advance a new research direction.

• 207 computer science research projects
• 184 lead authors
• 894 author-validated research questions
• 191K-paper temporally filtered computer science corpus

Dataset

• ScholarCatalyst/ScholarCatalyst: benchmark instances and annotations