Research writing that makes LLMs interview each other

What it does

STORM is an LLM system that researches a topic through internet search and writes Wikipedia-style articles with citations. It splits the work into a pre-writing stage—where it collects references and builds an outline—and a writing stage that produces the full text. A newer variant, Co-STORM, adds a collaborative loop where human users can observe or steer a live discourse among LLM agents that maintain a shared mind map.

The interesting bit

The project treats the hardest part of automated research as question generation, not answer generation. To avoid the shallow questions a single LLM prompt produces, STORM first discovers perspectives by surveying existing articles on similar topics, then simulates conversations between a Wikipedia writer and topic experts grounded in search results. Co-STORM extends this with a dynamic mind map that organizes discovered concepts hierarchically, aiming to reduce mental load when the conversation goes deep.

Key highlights

• Modular pipeline built on dspy, with separate LM configs for each stage (e.g., cheaper models for conversation simulation, stronger models for article generation)
• Broad retrieval support via litellm integration and native modules for Bing, Google, Tavily, DuckDuckGo, Azure AI Search, and user-provided documents via VectorRM
• Co-STORM introduces a collaborative discourse protocol with a moderator agent and a turn-management system for human-in-the-loop research
• Explicitly not publication-ready: the authors note output requires significant editing, though experienced Wikipedia editors found it useful for pre-writing
• Over 70,000 people have used the live research preview

Caveats

• Output requires significant editing before publication; the authors position it as a pre-writing assistant rather than a final draft generator
• Requires live internet search and LLM APIs; there is no fully offline mode

Verdict

Worth exploring if you need structured, cited long-form drafts and accept the cost of multi-step LLM inference. Skip it if you are looking for a fire-and-forget, publication-ready article generator.