The rapid expansion of scientific papers makes survey writing increasingly difficult, motivating automatic generation. Existing approaches search an open space of papers and concepts without the community’s navigation trace, resulting in incomplete scope, weak taxonomies, and unreliable citations. The key insight is that author‑written Related Work sections contain human judgments: authors curate relevant citations and describe how they relate, providing paths for scope reconstruction and relational evidence. InsightRAG follows these author‑supported citation links to retrieve a verified research scope and to structure the relational evidence. It then aggregates the judgments under a minimum‑conflict objective into a taxonomy that captures recurring author perspectives, guides generation, and bounds the evidence for each section. Experiments on ten topics show improvements in scope reconstruction, taxonomy quality, citation support, and reference reliability.
Review