TL;DR
An LLM compresses all human knowledge into a single distribution — and with every new request it gravitates toward the center. Ask a model ten times to come up with product ideas and you'll get ten similar answers clustered around "typical" solutions. The authors argue this follows from how LLMs are trained, not from one particular model (they tested mainly GPT-4o, with Gemini and GPT replications). Two mechanisms work against diversity: fixation within a single session and knowledge aggregation across sessions.
Fixation is when early ideas suck in the later ones. The model generates text step by step: each next token depends on the previous ones. If the first idea is "a gamified app," the next four will most likely circle around gamification too. On top of that, the authors argue (as theory, not tested) that RLHF (reinforcement learning from human feedback) pushes the model further toward "familiar" answers by training it on the averaged preferences of annotators.
Two targeted prompts treat each mechanism separately: Chain-of-Thought (CoT) breaks fixation within a session — the model first sketches titles, then forcibly diversifies them, then expands them. Ordinary personas (not a famous founder, but "a retired politician who has campaigned for library funding") pull the model into different corners of its knowledge across sessions — though on their own they increase fixation within a session. Together, in the paper's fitness-product study, they matched or beat the human sample on most diversity measures.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
Two levels that work independently — and together give the maximum:
LEVEL 1: CoT — against fixation (one request)
STEP 1: Generate {N} short idea titles (don't expand them)
STEP 2: Revise the titles — make each one as different
from the others as possible, replace repeated themes
STEP 3: Expand each title into a full idea
LEVEL 2: Personas — against aggregation (several requests)
REQUEST 1: [Persona A — specific, ordinary] + the CoT above → ideas
REQUEST 2: [Persona B — different, ordinary] + the CoT above → ideas
REQUEST 3: [Persona C — different, ordinary] + the CoT above → ideas
...
OUTCOME: merge all the idea sets → maximum coverage
⚠️ Each level works in a regular chat. CoT is a single request.
Personas are separate requests (or one with several roles in a row).
Example in Practice
The task, personas and prompt below are our illustration; the paper tested fitness product ideas with one-line personas.
Task: Come up with ideas for Substack newsletters in the personal finance niche. Without the technique, the model churns out the same thing every time: "investing tips," "stock breakdowns," "financial literacy." Boring and predictable.
Prompt (CoT + persona):
You are Megan, 34, a kindergarten teacher from Toledo, Ohio.
You love vegetable gardening and Nordic walking. You got into personal finance
two years ago after your divorce — you learned everything from scratch.
Your task: come up with ideas for personal finance newsletters on Substack.
STEP 1. First, sketch 8 short concept titles (3–7 words each).
Don't expand them — titles only.
STEP 2. Reread the list. Find repeated themes and angles.
Replace similar titles with fundamentally different ones —
so that the 8 ideas cover the widest possible range of the topic.
Show the updated list.
STEP 3. Expand each idea: 2–3 sentences — what the newsletter is about,
who the audience is, how it stands out from the rest.
Then repeat the request with a different persona — for example, "Mike, 52, an Army veteran from Tucson who retired early" or "Katie, 26, a freelance designer from Chicago who works for overseas clients."
Expected result:
The model will show all three stages explicitly: a rough list of titles → a revised list explaining what it changed and why → fully developed concepts. Each persona pulls toward its own corner: the teacher gives ideas through the lens of "an ordinary person with no finance background," the veteran through discipline and early retirement, the freelancer through irregular income and currency risk. Merge the three sets and you get a wider palette than a single request usually gives.
Why It Works
The fixation problem. An LLM generates text step by step — each token depends on the previous ones. That means the first idea literally "weighs" more when generating the second, the second when generating the third. Start with "gamification" and the next ideas will gravitate toward that theme. Plus, the authors argue, RLHF taught the model to produce "approvable" answers — and approvable tends to mean familiar and predictable (their explanation, not tested).
Why CoT breaks fixation. Step 1 (titles only) keeps early ideas from "growing" and taking over the context. Fewer tokens → a weaker anchor. Step 2 (an explicit instruction to diversify) is a direct command to break the "typical" pattern. In the paper CoT clearly reduced fixation in the model, while for people the same procedure slightly reduced diversity; the authors call the idea that LLMs simply follow such instructions better "speculative."
Why a famous founder is worse than an ordinary person. The paper did not test "Steve Jobs" itself. It compared 99 real, well-known fitness entrepreneurs ("creative" personas) with 99 ordinary people drawn at random from a large persona dataset. Both beat the plain model, but the ordinary personas gave more diverse ideas on every measure. The authors treat a persona as a sampling cue: it points the model to one region of its knowledge. A likely reason for the gap (our reading): successful founders in one industry point to similar regions, while random ordinary people are spread across many walks of life.
Control levers: - Number of CoT steps → our suggestion, untested: cut to 2 (titles + straight to expansion) for simple tasks - Persona specificity → the paper's personas were one-line descriptions (e.g., "a retired politician who has successfully campaigned for library funding"); adding a name, age and city, as in our example, is our addition and untested - Number of personas → in the paper each separate session got its own persona (99 sessions); how few personas you can get away with was not tested - Explicitness of the step 2 instruction → you can add "list which themes are already used and what's missing" — the model will make the diversity visible
Prompt Template
You are {name}, {age}, a {profession} from {city}.
{1-2 specific details: hobbies, life experience, context}.
Task: come up with ideas for {task topic}.
STEP 1. Sketch {N} short concept titles (3–7 words).
Titles only — no explanations.
STEP 2. Reread the list. Find repeated themes and angles.
Replace similar ones with fundamentally different ones — so the {N} ideas
cover the widest possible range.
Show the updated list.
STEP 3. Expand each idea: {format — e.g., 2-3 sentences /
bullets / a table}.
What to fill in:
- {name}, {age}, {profession}, {city} — specific and ordinary; not "Elon Musk," but "Tom, an accountant from Dayton"
- {1-2 details} — hobbies, life experience, something specific
- {N} — 6–10 for good coverage
- {task topic} — as specific as possible: "ideas for newsletter posts about remote work," not just "content ideas"
- {format} — whatever output format is convenient for you
🚀 Quick start — paste into the chat:
Here's a template for generating diverse ideas with CoT + a persona.
Adapt it to my task: {your task}.
Ask me questions to fill in all the fields.
[paste the template above]
The LLM will ask about the task topic, the desired format and the number of ideas — because without them it doesn't know what kind of persona to create or how many steps to run.
Limitations
⚠️ A method for ideation, not facts: CoT + personas help where you need broad coverage of options. For factual questions ("how much does it cost to register an LLC?") the technique is pointless.
⚠️ A persona anchors but doesn't guarantee: if the personas are too similar ("a manager from Chicago" and "a marketer from Chicago"), diversity will be minimal. You need contrasting biographies — different ages, professions, life contexts.
⚠️ Raising the temperature doesn't help: it's tempting to crank up randomness in the settings. But according to the study, raising temperature from 1.0 to 1.5 gave a small gain in diversity, and at 2.0 outputs were often incoherent or garbled.
⚠️ CoT doesn't work for people the way it does for LLMs: if you're brainstorming with a team, this technique may not help humans. In the paper, people using the same CoT procedure ended up with slightly less diverse ideas, not more.
⚠️ The maximum effect takes several requests: a single CoT request improves diversity within a session. The real breakthrough is several requests with different personas. It takes extra time, but the result is worth it.
How They Studied It
The Columbia Business School team ran a "fair" experiment — unlike most earlier work, where LLMs and humans got different instructions. Here the conditions were equalized: the same task (ideas for fitness products), ten ideas per participant. Two studies, with about 100 human participants each and GPT-4o as the main model.
The main methodological contribution is a new way to measure idea diversity. Standard methods compare word similarity (embeddings). Instead, the researchers used an LLM for semantic categorization: ideas were sorted into conceptual categories (industry context, psychological need, product form) rather than by lexical similarity. The authors argue this captures meaning better than wording (our example: "a meditation app" and "a mindfulness program" are lexically different but semantically the same thing).
To measure fixation they looked at the rate at which new categories accumulate: if the first three ideas come from new categories and the next seven repeat old ones, there's fixation. The LLM showed the same curve as humans: diversity piles up quickly at first and then fades. CoT flattened that curve — the model kept adding new categories for longer.
The most unexpected result: famous-founder personas are worse than ordinary ones. Intuition says the opposite — a creative icon should produce unconventional ideas. In the paper, personas of 99 real fitness entrepreneurs did beat the plain model, but they "still fall short of ordinary personas" on every diversity measure. There was a catch with ordinary personas too: on their own they increased fixation within a session, because the persona stays in context and keeps pulling ideas the same way. That's why CoT is needed on top.
The CoT + ordinary personas combination beat the human sample on the number of distinct idea combinations (240 vs 196, about 22% more), while matching it roughly on the number of categories (27 vs 28). The authors say that, to their knowledge, no earlier paper had pushed LLM diversity above human levels. In replications with other models the combination was again the most diverse LLM setup, but beat humans on some measures, not all.
Original from the Paper (optional)
The authors describe the CoT method they tested as follows (adapted from De Freitas et al. 2025, which builds on Meincke et al. 2024b). This is a summary, not a verbatim quote:
Chain-of-Thought prompting for ideation:
Step 1: Generate a set of short idea titles (summaries),
without expanding them into full ideas.
Step 2: Explicitly revise the titles to ensure they are
different from each other — bolder, more distinct,
covering different angles.
Step 3: Expand the revised titles into full ideas.
Key mechanism:
- Short titles prevent early elaboration → weaker token anchoring
- Explicit "make different" instruction → breaks typicality bias from RLHF
Context: In the paper's version the model wrote 10 short titles, made them "bolder and more different," and returned 10 one-sentence fitness product ideas as JSON. The persona version opened with "You are acting as this persona: {persona}" and added steps to consider what the persona would value and how they would approach each idea. The link to RLHF in the "key mechanism" lines is the authors' explanation, not something they tested.
Adaptations and Extrapolations
💡 Adaptation: Several Personas in One Request
If you don't want to run three separate requests, ask the model to play several roles in sequence within one conversation:
We're going to run three rounds of ideation for {task}.
In each round you take on a new role and generate ideas using the 3 steps below.
ROUND 1: You are {Persona A}. [CoT steps 1-3]
ROUND 2: You are {Persona B}. [CoT steps 1-3]
ROUND 3: You are {Persona C}. [CoT steps 1-3]
After three rounds: pick the 5 most dissimilar ideas across all rounds.
Untested in the paper: in one shared context earlier rounds can anchor later ones (the paper found a persona kept in context increases fixation), so expect less diversity than from separate requests — but you win on convenience.
🔧 Technique: Add an "Antagonist" to CoT
In step 2 you can add an explicit list of what's already been covered:
STEP 2. List all the themes and angles from step 1.
Now find 3 themes that are NOT on this list.
Replace 3 similar titles with ideas from these new themes.
This makes the diversification process visible and mandatory — the model literally maps what it has already used and hunts for blank spots.
🔧 Technique: The Outsider Persona
Our idea, not tested in the paper: try someone from a different industry looking at your task for the first time:
You are Linda, 47, an agronomist from Fresno, California.
You just found out what Substack newsletters are, and you're thinking about
how to apply what you know about farming
to the topic of {task}.
The farther the outsider is from the topic, the more unexpected the analogies are likely to be. An agronomist in the personal finance niche will offer ideas about "crop rotation" for capital, diversification as polyculture, the "yield" of investments — metaphors a "finance expert" would never reach.
Resources
A Cognitive Framework for Idea Diversity in Large Language Models (arXiv v2; v1 was titled "Examining and Addressing Barriers to Diversity in LLM-Generated Ideas") Yuting Deng, Melanie Brucks, Olivier Toubia Columbia Business School
The work builds on: - Meincke et al. 2024b — CoT for ideation - De Freitas et al. 2025 — source of the CoT prompt and the creative-persona approach - Doshi & Hauser 2024 — an embedding-based measure of collective diversity - Wei et al. 2022 — Chain-of-Thought prompting (the original)