NAACL 2024 Stanford OVAL

Stanford STORM Workflow: Perspective-Guided Outline Induction

STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking) pioneers pre-writing perspective discovery, simulated expert interviews, and hierarchical outline-first drafting to produce comprehensive, Wikipedia-grade reports.

Organization Gain

+25.4%

Over baseline RAG systems

Expert Personas

4–6 Personas

Uncovers unknown unknowns

Drafting Architecture

Outline-First

H1 → H2 → H3 Hierarchy

Context Isolation

Per-Section

Eliminates Lost-in-Middle

The 4-Stage STORM Architectural Workflow

Unlike naive research loops that jump straight into writing from raw search results, STORM decouples knowledge curation into two explicit phases: Pre-Writing (Stages 1–3) and Writing (Stage 4):

flowchart TD
  subgraph Stage1[Stage 1: Perspective Discovery]
    A[User Research Topic] --> B[Domain Survey via Web Search]
    B --> C[Generate 4-6 Diverse Expert Personas\nArch, Security, SRE, Cost, Practitioner]
  end

  subgraph Stage2[Stage 2: Simulated Multi-Turn Dialogue]
    C --> D1[Persona A: Security Analyst]
    C --> D2[Persona B: Systems Architect]
    C --> D3[Persona C: SRE Operator]
    D1 & D2 & D3 --> E[Information-Seeking Dialogue with RAG Grounded Topic Expert]
    E --> F[Extract Perspective-Specific Citations & Nuances]
  end

  subgraph Stage3[Stage 3: Hierarchical Outline Induction]
    F --> G[Cluster Findings & Cluster Themes]
    G --> H[Synthesize Global Outline H1 -> H2 -> H3]
    H --> I[Assign Reference Subsets to Specific Outline Nodes]
  end

  subgraph Stage4[Stage 4: Section-by-Section Isolated Drafting]
    I --> J1[Draft Section 1.0 using Subset 1]
    I --> J2[Draft Section 2.0 using Subset 2]
    I --> J3[Draft Section 3.0 using Subset 3]
    J1 & J2 & J3 --> K[Coherence Polish & Global Assembly]
    K --> L[Final Wikipedia-Grade Report with Verified Citations]
  end

Deep Breakdown of the 4 Stages

Stage 1: Perspective Discovery & Persona Generation

Standard search queries suffer from confirmation bias and narrow perspective framing. STORM addresses this by querying Wikipedia and broad search indices for related articles, then generating distinct expert roles:

Persona Generation Prompt Example:

"You are an editor preparing a comprehensive monograph on [Topic]. Identify 5 essential expert perspectives needed to provide a 360-degree technical analysis. For each perspective, define their primary domain concern, specific failure modes they investigate, and key parameters they evaluate."

Stage 2: Simulated Multi-Turn Information-Seeking Dialogue

Each persona engages in a 3–5 turn dialogue with a simulated "Topic Expert" agent grounded with real-time web retrieval. Instead of asking generic questions, personas ask focused technical follow-ups:

Expert Persona	Target Investigation Angle	Sample Autonomous Query
Systems Architect	Internal mechanics, protocol specifications, data structures	`"raft" AND "state machine replication" "linearizable reads"`
SRE / Operator	Failure modes, split-brain scenarios, recovery time (RTO)	`"raft" leader election split-brain post-mortem`
Security Auditor	Authentication, replay attacks, TLS configuration	`"raft" cluster membership change CVE vulnerability`
Cost / Performance	Disk IOPS overhead, network round-trips, throughput limits	`"raft" benchmark IOPS disk fsync overhead`

Stage 3: Hierarchical Outline Induction ($H_1 \to H_2 \to H_3$)

Rather than immediately writing prose, STORM synthesizes a detailed table of contents. The outline is organized hierarchically so that every outline node has a clearly bounded scope and an assigned set of verified source passages.

Stage 4: Section-by-Section Isolated Drafting (Map-Write)

Feeding 50+ search snippets into an LLM at once triggers attention degradation (the Lost-in-the-Middle problem). STORM solves this through isolated section drafting:

To write Section 2.1, the writer model receives only the outline structure, the preceding section summary, and the 3–5 source passages mapped to Section 2.1.
Each section is drafted with surgical precision, strict citation tags, and zero token bloat.

Quantitative Evaluation: STORM vs. Baseline RAG

Peer-reviewed human and automated evaluation from NAACL 2024 across 5 critical dimensions:

Key Strengths & Integration into Antigravity

STORM Architectural Pattern	How It Solves Traditional Research Flaws	Implementation in Antigravity Parallel-Search
Perspective Discovery	Prevents query collapse and generic marketing regurgitation.	Injected into Step 0 to generate 3–5 orthogonal domain angles before subagent dispatch.
Hierarchical Outline Induction	Enforces logical narrative progression and prevents rambling disjointed reports.	Integrated into Step 1 via `research_plan.md` with node-to-agent task assignment.
Context Isolation (Map-Write)	Prevents context blowout and hallucination during report assembly.	Implemented in Step 3 by synthesizing report sections independently using distilled JSON payloads.