Stanford University has open sourced its newly developed AI writing system STORM&Co-STORM. This system can integrate multi-source information to generate high-quality long articles based on simple topic input, greatly improving the efficiency of scientific research writing. STORM uses multi-angle Q&A by "LLM experts" and "LLM moderators" to iteratively generate articles, while Co-STORM generates dynamic mind maps through multi-agent dialogue to ensure comprehensive information. The system allows users to freely select modes and generate structured long articles within 3 minutes, and provides the function of viewing the brainstorming process and article examples.
The core technology of the STORM&Co-STORM system includes support for Bing search and GPT-4o mini. Its automated writing process is divided into three stages: multi-perspective question generation, outline generation and improvement, and full text generation. Although the system integrates information from multiple sources, the information sources may be biased towards the mainstream and may contain promotional content. Co-STORM aims to solve the problem of information omission and improve learning efficiency. User testing shows that it significantly reduces cognitive load. Currently the system only supports English and will be expanded to multiple languages in the future. The open source STORM&Co-STORM system brings new possibilities to personalized learning, making knowledge acquisition more convenient and efficient.
Users only need to enter English subject words, and the system can generate high-quality long articles that integrate multi-source information, similar to Wikipedia articles. To experience the STORM system, users can freely choose between STORM and Co-STORM modes. After a given topic, STORM can form a structured, high-quality long article within 3 minutes.
In addition, users can also view the brainstorming process of different LLM roles by clicking "See BrainSTORMing Process". In the "Discover" column, users can refer to articles and chat examples generated by other scholars, and personally generated articles and chat records can also be found in the sidebar "My Library".
The automated writing process of the STORM system is divided into three major stages: multi-perspective question generation, outline generation and improvement, and full text generation. The system consults relevant Wikipedia articles to identify various perspectives covering the topic, and then simulates a conversation between a Wikipedia writer and an expert based on reliable online sources. Based on the inherent knowledge of LLM, the dialogue content collected from different perspectives is finally carefully compiled into a writing outline.
Although STORM uncovers diverse perspectives when researching a given topic, the information collected may still lean toward mainstream sources on the Internet and may contain promotional content. Another limitation of the study is that although the researchers focused on generating Wikipedia-like articles from scratch, they only considered generating freely organized text. High-quality, human-written Wikipedia articles typically contain structured data and multimodal information.
Co-STORM aims to improve the problem of information omission in information collection and integration to greatly promote learning efficiency. It helps users understand and participate in the organization of information through multi-agent collaborative dialogue, dynamic mind mapping and report generation modules. The researchers conducted human evaluations on 20 volunteers, comparing Co-STORM's performance with traditional search engines and RAG Chatbot. The results show that Co-STORM significantly improves the depth and breadth of information, and 70% of users prefer Co-STORM, believing that it significantly reduces cognitive load.
Currently, the STORM&Co-STORM system only supports English interaction, and may be expanded to multi-language interaction capabilities in the future. The open source of this system is a sign that we are living in an extraordinary era where access to information can be completely tailored to the individual level, making it possible to learn anything.
Paper address: https://www.arxiv.org/pdf/2408.15232
All in all, the open source of the STORM&Co-STORM system has brought new breakthroughs to the field of artificial intelligence-assisted writing, and its efficient and convenient features are worth looking forward to. In the future, with the implementation of multi-language support and further improvement of functions, this system will play a greater role in academic research and daily writing.