STORM (AI Tool)

STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking) is an open-source artificial intelligence research tool developed by Stanford University’s Open Virtual Assistant Lab (OVAL) . It automates the creation of structured, cited articles resembling Wikipedia articles by using large language models (LLMs) and web-based information retrieval. Launched in early 2024, STORM is designed for researchers, students, and content creators to assist with knowledge curation and report generation.

Overview

STORM is a research prototype that generates long-form, cited articles on user-specified topics. It focuses on a structured pre-writing process, including topic research, outline creation, and multi-perspective question formulation. The tool is publicly accessible, requiring a user account, and is hosted on a Stanford platform. The system integrates LLMs, such as OpenAI’s GPT-4, with search engines like Bing or Tavily to collect and synthesize online information. Outputs are formatted like Wikipedia articles, with a title, introduction, sections, subheadings, conclusion, and references. Users can review the tool’s research process, termed "BrainSTORMing," which documents the AI’s information-gathering and question-asking steps.

Development and Functionality

Developed by Stanford’s OVAL lab, STORM was created by researchers including Yijia Shao, Yucheng Jiang, Theodore A. Kanell, Peter Xu, Omar Khattab, and Monica S. Lam. The project explores how LLMs can produce organized, cited content. STORM operates in two stages: Pre-Writing Stage: The tool retrieves online references and builds a hierarchical outline using two methods: Perspective-Guided Question Asking: STORM identifies diverse viewpoints from sources like Wikipedia to formulate relevant questions. Simulated Conversations: The AI generates a dialogue between a writer and a topic expert, grounding responses in internet sources to refine the outline. Writing Stage: STORM drafts a full article based on the outline and references, including citations. Users can customize the tool by selecting different LLMs (e.g., Claude, Gemini, Mistral) or search APIs (e.g., DuckDuckGo, Google).

The tool’s code is open-source, available on GitHub, allowing developers to modify its components for specific tasks, such as creating summaries instead of full articles.