Automating cognitive distillation for expert-level scientific literature synthesis

Jingyi Chai, Xiangrui Liu, Yanbo Han, Yuwen Du, Tian Jin, Yuzhi Zhang, Linfeng Zhang, Siheng Chen

Track: Track 1: Original Research/Position/Education/Attention Track

Abstract: Scientific progress depends on synthesizing a body of literature that is now growing faster than experts can read it. Recent large-language-model systems can write survey-like text in hours, but the result still falls far short of an expert-written review. Here we present SurveyMaster, an artificial intelligence system that generates expert-level literature reviews across scientific disciplines. Given a research description, SurveyMaster (i) calibrates generation with a discipline-specific writing skill, (ii) builds a comprehensive paper pool through seed-survey-driven hierarchical retrieval across a 160-million-paper scientific database, and (iii) anchors the manuscript on a small set of core papers that form its conceptual backbone. To evaluate SurveyMaster, we introduce SurveyMasterBench, a multidisciplinary benchmark of 100 expert-curated synthesis tasks across ten natural- and social-science disciplines. SurveyMaster achieves the highest overall score in all ten disciplines. It matches or exceeds expert-authored reviews on topical relevance, coverage, coherence and critical analysis, and improves citation grounding from 3.08 to 3.95 (fact correctness) and from 2.93 to 3.55 (citation precision) over the strongest automatic baseline. Controlled ablations confirm that these gains come from the three design choices, not from prompt-level tuning. By recasting survey generation as evidence-grounded cognitive distillation, SurveyMaster offers the scientific community a scalable way to keep pace with the rapidly growing literature.

Keywords: large language models, literature review, automated survey generation

Submission Number: 343