How does STORM's perspective discovery step work, and what is the…

Research Question

How does the STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective question generation) system's perspective discovery step generate diverse expert viewpoints before decomposing a research question into sub-questions, specifically what algorithm, prompt structure, and diversity criteria it uses, and what is the minimum-viable prompt template that replicates the coverage breadth improvement reported in the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2024) paper (+10% against baseline Retrieval-Augmented Generation (RAG)) within a single-agent automated research workflow that cannot conduct real conversations with simulated experts?

Findings

Executive Summary

STORM's perspective discovery is a lightweight persona-generation step seeded from related Wikipedia article outlines, not a formal diversity algorithm, and the safest single-agent replication is a fixed perspective-seeding prompt rather than a claim that the full STORM coverage gain will transfer unchanged.

The paper's +10% broad in coverage result belongs to end-to-end STORM versus an outline-driven retrieval-augmented generation baseline, while the ablation evidence supports the inference that perspective discovery alone has a smaller but still real effect, especially on source diversity and entity-level outline coverage.

One strong minimum-viable §0.5 candidate for this repository is an additive four-perspective prompt that emits one seed question per non-overlapping lens and minimizes downstream workflow change, while leaving room for later testing of topic-sensitive slot changes or a light related topics retrieval step.

Key Findings

1. STORM discovers perspectives by surveying related Wikipedia pages, extracting their titles and tables of contents, and then prompting the model to invent editor personas that each represent a different perspective, role, or affiliation before any question decomposition begins.
2. The released implementation always prepends a Basic fact writer persona, which means STORM explicitly combines one broad factual lens with a small set of topic-specific lenses instead of relying only on specialist roles.
3. STORM does not publish an explicit diversity rubric over disciplinary, cultural, temporal, or stakeholder categories, so its diversity mechanism is implicit and retrieval-primed rather than a formal coverage algorithm with declared quotas.
4. The paper's +10% broad in coverage result should not be attributed to perspective discovery alone, because that number compares full STORM against an outline-driven retrieval-augmented generation baseline and the ablations isolate a smaller persona-specific effect.
5. The ablation results show that removing perspective conditioning roughly halves the number of unique references collected and lowers entity-level outline recall, while removing simulated conversation hurts performance even more, so the evidence supports the inference that conversation contributes more than persona seeding to overall question quality.
6. Self-consistency and Six Thinking Hats are useful comparison points, but neither is a close substitute for STORM's perspective discovery because self-consistency collapses multiple attempts into one answer and Six Thinking Hats organises modes of thought rather than parallel role-conditioned lenses.
7. A strong minimum-viable candidate for this repository is a four-slot prompt, basic facts, mechanism or implementation, stakeholder or decision impact, and failure mode or critic, with one seed question per perspective, because that preserves STORM's broad-facts-plus-specialist-lenses pattern while keeping the prompt simple and leaving room for later topic-sensitive refinements.
8. The safest repository recommendation is to keep §0.5 additive by making it emit seed questions for §1 instead of redesigning later stages, because the backlog goal is broader question coverage and Anthropic guidance favors the smallest workflow change that preserves task structure.

Assumptions

• [assumption] The minimum-viable version should optimise for preserving perspective seeding, not for reproducing STORM's full multi-turn conversation behavior, because the repository workflow does not run simulated expert dialogues. Justification: the workflow constraint removes the paper's strongest ablated component. Source: Shao et al. (2024) Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models Anthropic Building Effective Agents
• [assumption] The topic statement and seeded sources usually contain enough signal for four useful lenses without first scraping related Wikipedia outlines, because the design goal is low-overhead prompt insertion rather than exact parity with STORM's internet-research phase. Justification: Anthropic guidance favors simpler composable workflows when they are adequate. Source: Anthropic Building Effective Agents Shao et al. (2024) Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models

Analysis

STORM's released code and paper align on a narrow interpretation of perspective discovery: it is a prompt-driven persona seeding stage that happens before question asking, not a formal optimisation pass over declared diversity dimensions.

That distinction matters because W-0038 cites the paper's +10% breadth gain as if it were the direct output of §0.5 alone, while the ablations show the single largest drop comes from removing simulated conversation rather than from removing perspective conditioning.

The safest design move is therefore to preserve the part that the repository can realistically inherit, persona-conditioned initial question selection, and to state plainly that the repository is not inheriting STORM's conversation-driven follow-up behavior.

Self-consistency is valuable once competing answer paths already exist, but it does not tell the model which topical lenses to open in the first place.

Six Thinking Hats provides useful coverage reminders, especially factual, critical, creative, and process modes, but its official method is parallel thinking rather than role multiplexing, so it works better as a post-generation audit than as the primary scaffold.

Anthropic's guidance points toward a short fixed-structure prompt with explicit output slots, which fits the repository's existing deterministic workflow better than a verbose freeform persona-generation instruction.

Recommended §0.5 prompt block:

### §0.5 Perspective Discovery

Before §1 Question Decomposition, generate exactly four non-overlapping research perspectives for the question below.
You are still one researcher. Do not simulate a panel, dialogue, or debate. Your job is to seed better questions.

Use these four slots:
1. Basic facts lens, what a broad factual writer must cover first.
2. Mechanism or implementation lens, how the thing works, is built, or fails operationally.
3. Stakeholder or decision-impact lens, who is affected, who decides, and what trade-offs matter.
4. Failure-mode or critic lens, what could be missing, misleading, risky, or overstated.

For each perspective, output:
- Perspective: <short role label>
- Distinct coverage added: <one sentence on what this lens sees that the others may miss>
- Seed question: <one concrete research question this lens would ask first>
- Evidence to seek: <the kind of source most likely to answer that question>

Constraints:
- Prefer non-overlap over stylistic variety.
- If two perspectives collapse into the same question class, rewrite one.
- Keep every seed question specific enough that §1 can decompose it into atomic sub-questions.
- Do not answer the questions yet.

Risks, Gaps, and Uncertainties

• [fact] The paper does not publish a direct experiment showing that a single-agent prompt block can reproduce the reported +10% breadth gain, so any claim of numeric equivalence would be unsupported. Source: Shao et al. (2024) Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models
• [fact] STORM's released implementation depends on simulated multi-turn retrieval-grounded conversation, which this repository's minimum-viable prompt does not reproduce. Source: Stanford OVAL STORM knowledge_curation.py
• [inference] The four-slot template may underfit topics whose most important diversity axis is historical or regulatory rather than stakeholder or failure mode, so the slot labels may need light topic-sensitive adjustment during implementation. Source: Anthropic Building Effective Agents Adversarial agents with shared goals: multi-perspective coverage across competencies and time horizons

Open Questions

• Can the repository evaluate §0.5 locally by measuring question diversity or downstream evidence-map coverage before and after insertion?
• Would a light related topics retrieval step before §0.5 materially outperform the fixed four-slot prompt enough to justify the added complexity?
• Should Six Thinking Hats be used as a post-§0.5 audit checklist to catch missing question classes without replacing role-conditioned lenses?

sources

• [x] Shao et al. (2024) Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models - primary paper describing STORM's perspective discovery algorithm, ablations, and human evaluation
• [x] Stanford OVAL STORM README - project documentation summarising the two-stage workflow and the role of perspective-guided question asking
• [x] Stanford OVAL STORM persona_generator.py - released implementation of related-topic discovery and persona generation
• [x] Stanford OVAL STORM knowledge_curation.py - released implementation of persona-conditioned question asking and conversation simulation
• [x] Wang et al. (2023) Self-Consistency Improves Chain of Thought Reasoning in Language Models - diversity-through-sampling baseline for comparison with perspective seeding
• [x] Edward de Bono Ltd Six Thinking Hats Summary - structured thinking-mode framework for comparison with role-based perspective generation
• [x] Anthropic Prompt Engineering System Prompts - prompt-writing guidance on explicit instruction following and scope control
• [x] Anthropic Building Effective Agents - workflow guidance on using simple composable patterns and multiple perspectives when needed
• [x] Adversarial agents with shared goals: multi-perspective coverage across competencies and time horizons - prior corpus item on why non-overlapping perspectives improve coverage
• [x] Exploration-synthesis gap: why people in explore mode fail to synthesise others' work, and whether agent synthesis can close the gap - prior corpus item on synthesis constraints in agent-mediated work