TL;DR
Ask an LLM to answer as several personas and the answers get more varied; that's old news. But this study shows that what matters isn't having a persona, it's how different the personas in the set are from each other. If the personas are similar (like "marketer" and "PR manager"), their answers will still be close. If the personas sit at opposite ends of the semantic space, the answers genuinely diverge.
A plain random persona boosts the creativity of LLM answers by only ~9%. The problem is that random personas often turn out to be "neighbors" (a college student and a high schooler, a doctor and a nurse), essentially variations of the same archetype. By default the model gravitates toward the most likely, averaged-out answer, and similar personas don't break that pull.
The method fixes this with an explicit criterion of maximum difference within the persona set: first you generate lots of candidate personas, then you select or create the ones that sit at opposite, rare, atypical points of the semantic space. That kind of set, not "different" but "maximally contrasting" personas, raises answer diversity by 78% and originality by 26%, while losing almost nothing in usefulness.
Method Blueprint
STEP 1: Collect/generate a large pool of candidate personas → list of personas
STEP 2: Select k personas so that the minimum difference between any pair is maximized
(or generate new personas, pushing them into rare, underexplored "zones of meaning")
→ final set of k contrasting personas
STEP 3: Run the task through each persona separately → k different answers
In the original, steps 1-2 require embeddings and mathematical optimization (done in code). In a chat, steps 1-3 can be collapsed into a single prompt by asking the model itself to stick to the "maximum spread" criterion.
Example in Practice
Task: The marketer for a small coffee shop in a residential neighborhood of Chicago wants unconventional promotion ideas. A regular brainstorm with coworkers keeps producing the same kind of ideas ("discount on your first cup," "student deal").
Prompt:
I need 5 personas for a brainstorm on this task:
"Come up with unconventional ways to promote a coffee shop in a residential neighborhood of Chicago."
The personas must be AS UNLIKE EACH OTHER AS POSSIBLE — not close variations
of one profession, age or lifestyle, but people from opposite
values, cultural backgrounds and ways of thinking.
Avoid averaged-out, typical archetypes (don't just take "marketer" and "designer") —
look for rare, atypical points of view that seldom end up in the same room.
For each persona:
1. A short description — who they are, what they value, how they think (1-2 sentences).
2. Their promotion idea, in the first person.
At the end, compare the ideas and note which ones turned out the most unexpected.
Result: The model will produce 5 clearly contrasting characters (for example, a retired engineer-philosopher, a teenage TikToker, a Buddhist monk, an ex-bartender from the '90s, a stay-at-home mom with small-town roots), each with a short description and an idea. At the end, a brief comparison shows which ideas are furthest apart in meaning rather than overlapping.
Why It Works
By default an LLM generates the most likely, averaged-out answer; that's baked in by training on "average" data. Asking for "several different points of view" doesn't break that pull: the model picks viewpoints that are statistically close for it, even if their names sound different.
Models are good at playing a clearly described role: if a persona is described concretely (values, thinking style, background), the answer really does shift into its "zone of meaning."
The method combines these two facts: by explicitly stating the criterion of "maximum distance between personas," we force the model to pick genuinely contrasting roles instead of close neighbors. The further apart the personas, the further apart their answers.
Control levers: - Number of personas (k) — increase it for a broad brainstorm, decrease it for a quick decision. - The "dissimilarity" criterion — you can shift the focus: values, culture, profession, age, way of thinking. - Two-step version — ask the model to first generate 15-20 personas, then in a separate request pick the 5 most contrasting. This more closely mimics the selection method (DISPERSION) from the paper. - "Rare zones" instead of just "different" — explicitly ask it to avoid typical archetypes and look for fringe, unconventional points of view; that's closer to the strongest method in the paper (evolutionary generation).
Prompt Template
I need {N} personas for {task}.
The personas must be AS UNLIKE EACH OTHER AS POSSIBLE — not variations of one
profession, age or culture, but people with opposite values,
ways of thinking and life experience.
Avoid averaged-out, typical archetypes — look for rare, unconventional points
of view that seldom overlap with each other.
For each persona:
1. A short description (who they are, values, thinking style) — 1-2 sentences.
2. Their answer to the task, in the first person.
At the end, compare the answers and note the most unexpected ones.
Fill in {N} — how many personas you need (3-7 is enough for most tasks), {task} — a specific statement of what needs to be invented or solved.
🚀 Quick start — paste into the chat:
Here's a template for generating diverse brainstorm personas. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask how many personas you need and which dimension of contrast matters most (values, profession, culture), because that determines along which "dimension" it will push the personas as far apart as possible.
Limitations
⚠️ Only for open-ended tasks: the method works for brainstorming, idea generation and creative lists. For tasks with a single correct answer (calculations, facts), it's unnecessary and can even get in the way.
⚠️ No hard guarantees in a chat: in the original, the "distance" between personas is computed with embeddings and math (Mahalanobis distance, MCMC). In an ordinary chat, the model judges the personas' "dissimilarity" itself; that's an approximation, not a guaranteed spread.
⚠️ A small loss of practicality: in the study, the most contrasting personas reduced answer usefulness by ~5%; some ideas from extreme personas are less directly applicable.
⚠️ The persona has to stay believable: if you chase contrast alone, the model may produce an absurd, invalid role. You need a balance between "as different as possible" and "makes sense."
How They Studied It
The researchers tested the hypothesis on three tasks: coming up with alternative uses for an object (say, a tin can), the Infinity-Chat dialogue benchmark, and producing lists of unrelated words (Divergent Association Task). Answers were generated by Gemma-4-31B, and diversity and creativity scores came from a separate judge model (Qwen3.6-27B) plus embedding-based measurements.
They compared five conditions: no persona, a random persona, personas selected for coverage of the space, personas selected for maximum spread, and personas generated "from scratch" via evolutionary search into rare zones of meaning. The finding: the greater the geometric spread of personas within a set, the greater the spread of their answers to the same task. This held in 12 out of 12 control comparisons.
The strongest result came from evolutionary persona generation, which pushes personas into underexplored, "sparse" zones of the semantic space: it beat even ready-made creativity-tuned prompts and amplified their effect when used together. The practical insight: what matters isn't the number of personas but how genuinely different they are from each other — and you can control that with a simple phrasing in the prompt, no math required.
Resources
Breaking Homogeneity: Diversifying Persona Sets for Creative LLM Outputs — Sang Bin Moon (Purdue University), Nicole Cho, Daniel Borrajo, Sumitra Ganesh (J.P. Morgan AI Research), Abolfazl Hashemi (Purdue University). Methods and datasets used: TextGrad (Yuksekgonul et al., 2025), PersonaHub / PersonaMem-v2 (Ge et al., 2024; Jiang et al., 2025).
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.