What is SurveyForge
SurveyForge is an innovative framework launched by Shanghai AI Lab in collaboration with Fudan University, Shanghai Jiao Tong University, and other institutions, used for automatically generating high-quality academic surveys. The framework is based on a two-stage design: outline generation and content generation. In the outline generation stage, based on a heuristic learning mechanism driven by dual-database collaboration (a research paper database and a survey outline database), it generates a well-structured and logically clear survey framework. In the content generation stage, the Scholar Navigation Agent (SANA) uses a memory module and a time-aware reranking engine to accurately retrieve literature and generate high-quality content. SurveyForge's parallel generation and coordination mechanism can efficiently generate coherent long documents. Generating a survey of about 64k tokens takes only 10 minutes and costs less than $0.5. The team built the SurveyBench multi-dimensional evaluation benchmark for objectively evaluating the quality of surveys.
Main Features of SurveyForge
- Efficient Survey Generation: SurveyForge can generate a survey of about 64k tokens in a short time, with the entire process taking only 10 minutes and costing less than $0.5.
- High-Quality Outline Generation: Based on a heuristic learning mechanism driven by dual-database collaboration, it generates a well-structured and logically clear survey outline, approaching the level of human writing.
- Accurate Literature Retrieval: The Scholar Navigation Agent (SANA) can accurately retrieve high-quality literature related to the topic, avoid citing irrelevant literature, and improve the accuracy and influence of the literature.
- Content Optimization and Coordination: Based on the parallel generation and coordination mechanism, each chapter is generated independently and then, based on a refinement stage, eliminates duplication, straightens out logic, and forms a coherent whole.
Technical Principles of SurveyForge
- Outline Generation Stage：
- Dual-Database Collaborative Drive: Constructs a research paper database (about 600,000 arXiv computer science papers) and a survey outline database (hierarchical outline structures of about 20,000 survey articles). Based on cross-database knowledge fusion, it retrieves topic-related papers and existing survey outlines, and learns expert-level structural patterns.
- Recursive Construction Strategy: First generates the first-level outline that reflects the global logic, then, for each chapter, combines domain literature to further refine the second-level structure, achieving structured learning from coarse to fine and from whole to part.
- Content Generation Stage：
- Scholar Navigation Agent (SANA)：
- Sub-query Memory Module: Uses the literature collection retrieved during the outline generation stage as memory context, combined with the original query, to ensure that the query decomposition process always revolves around the core of the topic and avoids semantic drift.
- Retrieval Memory Module: Uses the literature related to the entire outline as global memory, and retrieves the most relevant literature for each sub-query based on embedding similarity, improving retrieval precision and semantic consistency.
- Time-Aware Reranking Engine: Groups the retrieved literature by publication time, performs top-k selection within each group by citation count, and balances representative classic authoritative literature with cutting-edge emerging research.
- Parallel Generation and Coordination: Each chapter independently generates content, and based on a shared memory system ensures that it revolves around a unified topic framework, using a refinement stage to eliminate duplication, straighten out logic, and form a coherent whole.
- Scholar Navigation Agent (SANA)：
Project Address of SurveyForge
- GitHub Repository：https://github.com/Alpha-Innovator/SurveyForge
- HuggingFace Model Library：https://huggingface.co/datasets/U4R/SurveyBench
- arXiv Technical Paper：https://arxiv.org/pdf/2503.04629
Application Scenarios of SurveyForge
- Academic Research: Helps newcomers to a field, interdisciplinary researchers, and senior scholars quickly gain a comprehensive overview of the field and improve the efficiency of literature review.
- Education: Assists teachers in designing courses and students in learning, enabling them to quickly master domain knowledge.
- Industry: Helps enterprises conduct technology research and industry analysis, providing reference for R&D decisions.
- Policy Making: Provides scientific evidence for policymakers, assisting technology assessment and policy planning.
- Personal Learning and Research: Helps individuals quickly understand the latest developments in fields of interest and assists with paper writing.