Research teams from Zhejiang University and Shanghai Jiao Tong University, in collaboration with their partners, have successfully developed the Deep Academic Survey (DAS) system. This innovative system is capable of automatically generating an academic survey that is ready for publication within a mere one hour. DAS diverges from the conventional 'retrieval-writing' approach and embraces a Stateful Agentic Framework, underpinned by the DAS-2M literature metadata repository.

Initially, DAS conducts a thorough analysis of approximately 2 million arXiv papers, extracting structured metadata. Following this, it undertakes a series of steps, including retrieval, classification system construction, paper allocation, hierarchical planning and composition, as well as semantic review and deterministic verification. These steps collectively form a closed-loop process of generation-review-revision-reassessment, culminating in the production of a comprehensive PDF survey complete with charts and references.

In the DAS-Bench evaluation, DAS attained an average score of 4.34, surpassing the highest baseline score of 4.03. It also outperformed systems like Naive RAG and AutoSurvey in expert anonymous evaluations. Presently, the project website has made available 220 automatically generated surveys on trending topics, accompanied by comparison results. Nevertheless, the generated surveys still necessitate manual verification. Future endeavors could delve deeper into dynamic literature tracking and human-machine collaborative revisions to further enhance the system's capabilities.