Zhejiang & SJTU's DAS Generates Publication-Ready Academic Surveys in One Hour

As academic literature grows rapidly, high-quality surveys have become essential for researchers to enter new fields, map research landscapes, and track progress. However, writing a rigorous survey demands reading hundreds of papers, continuous tracking, and extensive expert time.

Large language models have enabled Deep Research and automatic survey generation systems that can retrieve, analyze, and generate long reports. Yet these systems typically follow a simple "retrieve–analyze–write" pipeline and produce research reports rather than structurally sound, publication-ready academic surveys. Key challenges remain: how to organize the survey, design a taxonomy, assign papers to sections, ensure each claim is backed by appropriate citations, and guarantee the final manuscript's references, cross-references, figures, tables, and PDF compilation are reliable.

To address these gaps, researchers from Zhejiang University and Shanghai Jiao Tong University propose Deep Academic Survey (DAS), a system that aims to generate a publication-ready academic survey in roughly one hour. Unlike traditional pipelines, DAS employs a Stateful Agentic Framework that redefines the process as a Stateful Agentic Closed-Loop Manuscript Construction.

DAS-2M Literature Metadata Lake

The foundation of DAS is the DAS-2M Literature Metadata Lake, built by parsing nearly 2 million arXiv papers published since January 2020. Using LLMs, the team extracts structured metadata oriented toward survey writing — including methods, datasets, experimental results, innovations, and limitations — for each paper. This metadata lake is fully open-sourced on Hugging Face. Compared to relying only on titles and abstracts, DAS-2M provides a fine-grained "pre-understanding" of papers at scale: each paper is parsed once and can be reused across different survey topics. Moreover, DAS-2M is a continuously updated lake; as new arXiv papers appear, they are processed through the same pipeline, allowing DAS to stay current.

On top of DAS-2M, DAS builds lexical and semantic indexes. Given a user-provided topic, the system retrieves a global candidate set from DAS-2M, constructs a Candidate-Grounded Taxonomy, and uses Reverse Paper-to-Section Routing to decide which sections each paper should support. Consequently, taxonomy design, paper assignment, paragraph planning, and citation planning all share the same reusable literature representation and candidate pool.

Stateful Agentic Closed Loop

DAS maintains a continuously updated Manuscript State: candidate papers determine the taxonomy; taxonomy and routing constrain writing; and issues found during review reactivate the corresponding writing state. This makes survey construction a persistent, revisable process rather than a one-shot pipeline.

Before drafting, DAS performs hierarchical planning: Taxonomy → Paragraph Plan → Writing Points → Claim → Citation Group. Citations are decided at the claim-planning stage — determining "which papers support which claim" — before paragraph drafting begins.

After drafting, a Semantic Reviewer checks section goals, logical flow, and citation support. If problems are detected, the system executes Direct Edit, Paragraph Replan, or Section Replan based on issue scope, then re-enters review. In parallel, Deterministic Validation verifies citation identifiers, cross-references, LaTeX syntax, and compilation. This creates a Generation → Review → Repair → Re-evaluation loop.

Only content that passes review proceeds to Manuscript Finalization, where the system assembles figures, tables, bibliography, cross-references, and LaTeX, then compiles a complete survey PDF. This end-to-end reliability is what DAS means by "publication-ready": ensuring consistency from literature organization through claim/citation planning to final manuscript artifacts.

Evaluation Results

On DAS-Bench's 30 survey topics, DAS is evaluated across four core dimensions — Citation, Taxonomy, Discourse, and Manuscript Reliability — achieving an average score of 4.34, while the strongest baseline that completed all 30 topics scored 4.03. In anonymous expert evaluations, DAS was preferred over Naive RAG on 27 of 30 topics and over AutoSurvey on 19 of 21 shared computer science topics.

The project website publicly releases 220 complete surveys on popular research directions, alongside outputs from Codex, GPT Deep Research, AutoSurvey, and InteractiveSurvey for direct side-by-side comparison. Qualitative analysis in the paper highlights differences in citation stacking, coarse-grained attribution, repetitive structure, and figure-text coupling.

Limitations and Future Work

The authors emphasize that "publication-ready" does not mean DAS can replace researchers for final submission; generated surveys still require human verification. Academic survey evaluation itself remains an open problem, with cross-judge discrepancies observed in the paper's cross-judge evaluation. As DAS-2M continues to update, future directions include dynamic literature tracking, human-AI collaborative editing, and maintainable, continuously updatable agentic academic survey construction. The ultimate question DAS explores is not whether AI can write a long article in one shot, but whether a traceable, correctable stateful agentic closed loop can help researchers enter unfamiliar fields faster and continuously organize evolving academic knowledge.

Resources: Paper (arXiv:2608.18034), Project (https://zhikaixu24.github.io/projects/DAS/), Code (https://github.com/ZhikaiXu24/DAS), Data (https://huggingface.co/datasets/ZhikaiXu24/DAS-2M).