STRUCTSURVEY uses multi-agent graphs to auto-generate better research surveys
A new framework improves survey paper ROUGE scores by +2.9 without sacrificing precision.
Paolo Pedinotti and Enrico Santus have released STRUCTSURVEY, a novel hierarchical multi-agent framework designed to automate the generation of scientific survey papers. The system addresses a key limitation of current LLM-based approaches: rather than forcing models to infer conceptual, methodological, and taxonomic relationships from raw text during generation, STRUCTSURVEY pre-constructs explicit graph-based representations of entities, relations, and topical hierarchies at retrieval time. This structural shift allows the framework to produce surveys that better mirror human-written organization and reasoning.
Evaluated on a new reference-grounded benchmark of ACL survey papers, STRUCTSURVEY outperforms embedding-only retrieval baselines by +2.9 points in ROUGE-1 recall and +1.0 points in ROUGE-2 recall, all without sacrificing precision. LLM-as-a-Judge evaluations also rated the generated surveys higher for logical structure, depth, and synthesis. The framework is submitted to arXiv as a 8-page paper (arXiv:2607.01243) and is positioned as a step toward reproducible, long-form scientific summarization.
- STRUCTSURVEY uses a multi-agent architecture to build dynamic graph representations of entities, relations, and taxonomies during retrieval.
- Outperforms embedding-only baselines: +2.9 ROUGE-1 and +1.0 ROUGE-2 recall without precision loss.
- LLM-as-a-Judge ratings show improved logical structure, depth, and synthesis compared to prior automated survey methods.
Why It Matters
Automated survey generation that mimics human organization could save researchers hours of literature review time.