Source: Shao et al. 2024: STORM
Stanford paper on writing Wikipedia-like articles from scratch by discovering perspectives, simulating grounded expert conversations, and outlining before writing.
Contents
- What: research paper, arXiv:2402.14207 (NAACL 2024)
- Author: Yijia Shao, Yucheng Jiang, Theodore A. Kanell, Peter Xu, Omar Khattab, Monica S. Lam
- Date: 2024-02-22 (v2 2024-04-08)
- Original: arXiv:2402.14207
Summary
STORM treats long-form report writing as a pre-writing research problem: before drafting, it gathers material by asking questions from many perspectives and organises it into an outline.
Key claims
- Pipeline: discover diverse perspectives on the topic; simulate conversations in which perspective-specific writers question a topic expert grounded in internet sources; curate the material into an outline; write the article from it.
- Evaluated on FreshWiki, recent high-quality Wikipedia articles, to avoid leakage from pre-training.
- Versus an outline-driven RAG baseline, more STORM articles were judged organised (+25% absolute) and broad in coverage (+10%).
- Open problems: bias in sources transfers into the article, and the model over-associates unrelated facts.