← All posts

AI Research · Field Notes

Turn One Prompt Into a Research Team.

Stanford’s STORM method runs your question past five expert perspectives, maps where they disagree, and verifies every source.

A real research team earns its keep by disagreeing out loud. Put five sharp people on a question and the practitioner, the skeptic, and the economist each pull it in a different direction, and the place they pull hardest is usually where your decision actually lives. Stanford built a method that gives a single AI that same shape: many expert perspectives, every source verified, and a clean map of where they diverge. The result reads like a briefing you can act on with confidence.

The method is called STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking), published by Stanford’s OVAL Lab at NAACL 2024. It assigns several expert personas to your question, has each one interrogate real sources, keeps a ledger of where they disagree, and builds the final briefing from that ledger. The structure is the whole point. Here is the shape of it.

Practitioner Academic Skeptic Economist Historian Contradiction ledger WHERE THEY DISAGREE Verified briefing SOURCED + CHECKED
Five perspectives in, one verified briefing out · the disagreement is the deliverable

The five perspectives

The core move is simple: let several viewpoints answer the question together. STORM gives each persona a different angle of attack, so every persona covers the ground the others leave open. A practical starting set:

01

The Practitioner

Has actually done the thing. Knows what works in the field, which textbook answers hold up on contact with reality, and which ones need adjusting.

02

The Academic

Cares about evidence and method. Knows what the literature supports, how strong the data is, and which popular claims rest on solid ground.

03

The Skeptic

Pressure-tests the consensus. Surfaces the hidden assumptions, the conditions a claim depends on, and the strongest counter-case, so your conclusion is one you have genuinely stress-tested.

04

The Economist

Follows the incentives and the second-order effects. Who benefits, what it costs, what it shifts, and what happens once everyone else does it too.

05

The Historian

Asks whether we have seen this before and how it played out. Pattern memory is the cheapest form of foresight, and a deliberate method makes sure to include it.

Five is a starting point. The real skill is choosing the perspectives that are decision-relevant to your specific question. A pricing decision wants the economist and the practitioner up front; a compliance question wants the skeptic and the academic. Personas chosen for the question at hand produce research you can use.

The six-stage workflow

The perspectives are the input. The method is the sequence that turns them into something you can act on. The order matters: structure is locked before a single sentence of the final draft gets written.

  1. 1

    Perspective generation

    Pick the expert viewpoints that genuinely bear on the decision. This is where most of the quality is won.

  2. 2

    Source-grounded interviews

    Each perspective interrogates real sources, so every claim traces back to something you can open and check.

  3. 3

    Contradiction tracking

    Keep a running ledger of where the perspectives conflict. This is the asset your decision turns on: a clear view of the live disagreements before you commit.

  4. 4

    Outline synthesis

    Build the structure only after the perspectives stabilize, with every section mapped to its sources. Stable outline first, prose second.

  5. 5

    Outline-based drafting

    Write strictly from the validated outline, so every line in the final draft traces to a source you approved. A fixed skeleton keeps the writing honest.

  6. 6

    Moderator pass

    A final review that asks what is still missing and adds it: the extra perspective, the question worth raising before you call the research done. Stanford calls this layer Co-STORM.

Why this matters for a business

Stanford’s own evaluation found that articles built this way were roughly 25 percentage points better organized than standard retrieval-based output, and covered about 10 points more ground. Their framing: the depth a PhD student would produce in 48 hours, delivered in a fraction of the time. The lab number is impressive. The number that should hold your attention shows up later, in the real decision: the value of choosing well because the research went deep enough to trust.

The decisions where this earns its keep are the expensive ones. Due diligence on an acquisition. Entering a new market. Choosing a platform you’ll rely on for five years. A competitive teardown the board will act on. In every one, the value of getting it right dwarfs the cost of a few extra minutes. STORM is built to surface the dissent that points you toward the right call.

The bottom line

The teams getting real value from AI research run it as a method: a process that weighs several expert views, checks every source, and shows its work before it reports to you. Applied to the decisions worth the effort, that structure is the kind of system 820labs builds into the way a business runs.

Book a discovery call