Stanford University has made significant advancements in the field of artificial intelligence, with their newly developed STORM&Co-STORM system now open-sourced. This system can comprehensively integrate multi-source information and generate high-quality long articles through simple topic input. This innovation not only avoids information blind spots but also greatly enhances the efficiency and quality of academic writing.
The core technologies of the STORM&Co-STORM system include support from Bing Search and GPT-4o mini. The STORM component iteratively generates outlines, paragraphs, and articles through multi-angle Q&A between "LLM experts" and "LLM hosts." Meanwhile, Co-STORM generates interactive dynamic mind maps through dialogues among multiple agents, ensuring that no information needs overlooked by the user.
Users only need to input an English topic keyword, and the system can generate a high-quality long text that integrates multi-source information, similar to a Wikipedia article. When experiencing the STORM system, users can freely choose between STORM and Co-STORM modes. Given a topic, STORM can produce a structured high-quality long text within 3 minutes.
Additionally, users can click "See BrainSTORMing Process" to view the brainstorming process of different LLM roles. In the "Discover" section, users can refer to articles and chat examples generated by other scholars, and personal 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 main stages: multi-perspective question generation, outline generation and refinement, and full text generation. The system determines various perspectives on the topic by consulting relevant Wikipedia articles, then simulates a dialogue where one party is a Wikipedia author and the other is an expert based on reliable online sources. The dialogue content collected from different perspectives is then meticulously organized into a writing outline.
Although STORM discovers different perspectives when researching a given topic, the information collected may still lean towards mainstream sources on the internet and may include promotional content. Another limitation of the research is that, while researchers focus on generating Wikipedia-like articles from scratch, they also only consider generating loosely organized texts. In contrast, high-quality human-written Wikipedia articles typically contain structured data and multimodal information.
Co-STORM aims to address the issue of information omission in the integration of information gathering, significantly enhancing learning efficiency. It helps users understand and engage in the organization of information through multi-agent collaborative dialogues, dynamic mind maps, and report generation modules. Researchers conducted human evaluations with 20 volunteers, comparing the performance of Co-STORM with traditional search engines and RAG Chatbots. The results showed that Co-STORM significantly improved the depth and breadth of information, with 70% of users preferring Co-STORM, believing it significantly reduced cognitive load.
Currently, the STORM&Co-STORM system only supports English interaction, but it may expand to multilingual interaction capabilities in the future. The open-sourcing of this system marks an extraordinary era where the way we access information can be fully tailored to individual levels, making learning anything possible.
Paper link: https://www.arxiv.org/pdf/2408.15232
