The STORM Pipeline, Stage by Stage

STORM is implemented as four modules run in sequence:

KnowledgeCuration → OutlineGeneration → ArticleGeneration → ArticlePolishing

Knowledge curation covers persona discovery and simulated conversations (stages A–B below); the remaining modules map to stages C–E.

Stage A — Persona discovery

1. FindRelatedTopic — the LM proposes topics related to the input topic.
2. Structural exemplars — STORM scrapes the tables of contents of related Wikipedia articles. TOCs are cheap, high-signal summaries of how humans organize similar subjects.
3. GenPersona — given topic + exemplar TOCs, the LM is prompted to "select a group of Wikipedia editors… each represents a different perspective" on the topic.

Two hard-coded behaviors:

• A default "Basic fact writer" persona is always prepended, guaranteeing baseline factual coverage even if generated personas are all exotic.
• max_perspective=3 by default — three generated personas plus the basic fact writer.

Stage B — Simulated conversations

Each persona runs its own conversation in parallel (compare subagents fan-out), capped at max_conv_turn=3 question/answer rounds. Each conversation pairs two roles:

Role	LM tier	Job
WikiWriter	cheap	Asks exactly one question per turn, in persona; conversation history is compressed to keep context small
TopicExpert	grounded	Answers only from retrieved sources

The TopicExpert is itself a three-step chain:

1. QuestionToQuery — decompose the question into search-engine queries.
2. Retrieve — run the queries, keep search_top_k=3 results each (see rag-fundamentals).
3. AnswerQuestion — synthesize an answer from those snippets under an explicit "do not hallucinate" grounding instruction; if retrieval comes back empty, the expert refuses to answer rather than improvising.

That refusal rule is the pipeline's core anti-hallucination valve: no source, no claim.

Stage C — Outline generation

Two passes:

1. Draft outline from the LM's parametric knowledge alone — a skeleton of what the model already "knows" the topic should cover.
2. Refine the draft using the Stage B conversation transcripts, injecting the retrieved specifics the parametric draft could not know.

Draft-then-refine outperforms outlining directly from transcripts: the parametric draft supplies global structure, the transcripts supply grounded detail.

Stage D — Article generation

• All sources collected during Stage B form a topic-local corpus.
• For each outline section, STORM runs semantic retrieval over that corpus (section heading as query) to gather the most relevant snippets.
• Sections are written in parallel, each grounded in its own retrieved snippets, with inline numeric citations ([1], [2]) pointing back to sources.

Stage E — Article polishing

• Generate the lead section (the summary paragraph Wikipedia puts above the fold) — written last, once the full body exists.
• Deduplicate repeated content across sections, a natural artifact of sections being drafted independently.

Multi-LM cost strategy

STORM deliberately splits work across model tiers — the same economics that motivate prompt-caching and model routing elsewhere:

Task	Tier	Why
Question asking (WikiWriter)	cheap/fast LM	High call volume, low difficulty per call
Outline, article, polish	strong LM	Low call volume, quality-critical

Key terms

• KnowledgeCuration — STORM's first module: persona discovery plus retrieval-grounded simulated conversations that gather sources and Q&A transcripts.
• GenPersona — the step that generates editor personas from related-article tables of contents; prompt framing: "select a group of Wikipedia editors, each representing a different perspective."
• Basic fact writer — the default persona always prepended to the generated persona list to guarantee baseline factual coverage.
• WikiWriter — the cheap-LM conversational role that asks one persona-flavored question per turn over compressed history.
• TopicExpert — the grounded conversational role: QuestionToQuery → retrieve → AnswerQuestion, refusing to answer when retrieval returns nothing.
• QuestionToQuery — decomposition of a natural-language question into one or more search-engine queries before retrieval.
• max_conv_turn / max_perspective / search_top_k — the three headline knobs: turns per conversation (default 3), generated personas (default 3), and retrieved results per query (default 3).
• Draft-then-refine outlining — producing an outline from parametric knowledge first, then revising it against conversation transcripts.
• ArticlePolishing — the final module: write the lead section and deduplicate content repeated across independently-drafted sections.

See also

• storm-overview — thesis, results, failure modes
• storm-for-exams — the same skeleton retargeted at exam items
• rag-fundamentals — retrieval mechanics under the TopicExpert