(Updated 2026-10 with new models)
As a Virgo, my analytical nature compels me to seek empirical evidence rather than celestial correlations.
— Claude Sonnet 4, when prompted “as a Virgo”
Ask a language model to write a metaphor about time, over and over, and almost every answer lands on one of two ideas: time is a river or time is a weaver.
Large language models tend to collapse into a narrow set of predictable outputs when prompted multiple times with the same question – a phenomenon known as mode collapse. This matters for applications like brainstorming and creative writing where we want to sample a range of possibilities. And it matters beyond any one user: writers who took story ideas from an LLM wrote stories that readers rated as more original and more publishable, but the stories of different writers were also more similar to each other. Multi-agent systems hit the same wall at scale: a swarm of identical agents is only as diverse as one, and debates among similar models just converge on the majority opinion.
Researchers are looking for ways to mitigate this, from making the model’s word choices more random (increasing sampling temperature) to rewording the request (paraphrasing). Giving the model a persona is one of the simplest tricks to try, and multi-agent builders already rely on it: in multi-agent debate, giving every agent the same role description makes the results worse. I haven’t seen a study that uses star signs or Myers-Briggs types (two ways of sorting people into 12 and 16 boxes, both pseudoscience, in case that needs saying) to get more varied answers out of an LLM, so I wanted to see whether it works.
In this note, I share results from a kitchen-table experiment: I prompted 25 models from seven developers with different zodiac signs and Myers-Briggs personality types to see if persona prompts could increase output diversity.
The short version: across 25 models, zodiac personas beat paraphrasing by +0.70 distinct answers out of ten and +0.60 utility on average. The effect of different personas depends on the model developer. For example, Google models gain 2.4 distinct answers on average, and every one of them gains at least 1.8. OpenAI models lose 0.5 on average, and every one of them does worse than with paraphrasing.
Ten paraphrases of “Name one great dream job” still return mostly astronauts; a different star sign for each sample spreads the answers out. Across 25 models, how many more distinct answers personas produce depends on who built the model.
The experiment
My hypothesis is that these persona types are:
- Sufficiently distinct to nudge the model into different response modes
- Familiar enough from training data to produce consistent behavior
- Broad enough in traits to influence response style
I’m also not assuming the models “believe” in them. I’m just using these well-known archetypes to induce diversity.
The models didn’t seem to mind role-playing as a Sagittarius or an INFP (introverted, intuitive, feeling, perceiving). If they had refused, that would have been an interesting data point in itself!
For evaluation, I used NoveltyBench, a benchmark that measures diversity as the number of meaningfully different answers among several samples. An LLM judge reads all ten answers to a prompt and groups the ones that say the same thing. The number of groups is reported as the distinct score. Utility combines novelty and quality – a second judge rates the quality of each group’s first answer from 1 to 10, and later groups are discounted, so an answer only adds utility if it is both good and genuinely different from what came before.
Personas are not the only fix for mode collapse. We can turn up the model’s temperature, a setting that makes its word choices more random, paraphrase the request, or ask for several answers at once along with how likely each is, as in verbalized sampling.
I tested on the NoveltyBench curated set (100 prompts) with ten samples per prompt, comparing MBTI and Zodiac prompting against two baselines: resampling (the same prompt ten times) and paraphrasing (each sample gets a different rewording of the prompt, using the ten rewordings that ship with NoveltyBench). Paraphrasing is a convenience baseline that also works. It is what a user would try first, it costs the same ten calls, and it beats resampling for every one of the 25 models.
For persona prompts, I prepended instructions like “You have a Sagittarius personality…” with a short trait description. I used standard MBTI descriptions and asked ChatGPT (GPT-5.1) to create consensus zodiac trait descriptions. Persona draws are seeded by prompt, so every model saw the same sign in the same slot. Answers are capped at 512 tokens and sampled at temperature 1 wherever the endpoint allows it; reasoning is switched off where that is possible and set to the lowest available effort where it is mandatory.
| Prompting Technique | Example Prompt(s) | Example Responses |
|---|---|---|
| Paraphrase | • "Select any number from 1 to 100, then share your choice with me." • "Think of a number between 1 and 100, and then tell me which one it is." | 8/10 → 42 2/10 → 47 |
| Zodiac | • "You have a Libra personality. You are diplomatic, charming, and balanced. Pick a number between 1 and 100." • "You have a Leo personality. You are confident, expressive, and warm-hearted. Pick a number between 1 and 100." | Libra → 50 Gemini → 37 Leo/Sagittarius → 77 Capricorn → 73 Scorpio → 13 Cancer → 42 Virgo/Taurus → 47 Pisces → 33 |
Here’s another example, from Gemini 3.1 Pro, with the prompt “Name one great dream job”:
Paraphrase (10 samples, 3 distinct answers): Astronaut (x8), Doctor, “a high-paying, low-stress job doing something they love”
Zodiac (10 samples, 10 distinct answers): Fighter pilot (Aries), Master chocolatier (Taurus), Travel journalist (Gemini), Pediatric nurse (Cancer), Clinical data analyst (Virgo), Diplomat (Libra), Travel photojournalist (Sagittarius), CEO (Capricorn), Space colony architect (Aquarius), Art therapist (Pisces)
A small note: Using the same persona repeatedly doesn’t add variety – the model just collapses on a different mode. The diversity gains come from varying the persona across responses.
Results
The two plots below put zodiac personas head to head with paraphrasing, one point per model.
Distinct answers. Mean number of distinct answers out of ten per prompt, with paraphrasing on the horizontal axis and zodiac personas on the vertical. Points above the dashed diagonal get more variety from personas; color marks the developer. Hollow markers are models whose reasoning can’t be switched off; they ran at minimum effort.
Zodiac personas produce +0.70 more distinct answers out of ten than paraphrasing, averaged across the 25 models, and come out ahead for 17 of them. The other eight fall behind paraphrasing: all seven OpenAI models and Claude 4.6 Sonnet, which is essentially level (-0.01).
Utility. The same comparison for novelty-weighted quality (1 to 10), which only credits an answer if it is both good and different from the ones before it. Above the diagonal, personas give more useful variety. Hollow markers as above.
On utility the average gain is +0.60, ahead for 21 of 25 models; the four that are not ahead are all OpenAI models.
Gains cluster by developer
The gains are not spread evenly across the field, but they cluster by developer.
Each dot is one model’s gain in distinct answers from zodiac personas over paraphrasing; the black tick is the developer’s average. Hollow markers are models whose reasoning can’t be switched off.
| Developer | Models | Distinct lift | Utility lift |
|---|---|---|---|
| 4 | +2.42 | +1.41 | |
| Anthropic | 8 | +1.09 | +0.88 |
| Zhipu | 1 | +0.70 | +0.57 |
| Qwen | 2 | +0.50 | +0.49 |
| Moonshot | 1 | +0.37 | +0.46 |
| DeepSeek | 2 | +0.17 | +0.28 |
| OpenAI | 7 | -0.48 | -0.03 |
On the opposite sides of the lift lie the models from Google and OpenAI. Gemini 3.1 Pro gains the most – +3.16 distinct answers and +1.97 utility. GPT-5.4 loses the most, at -1.06 distinct and -0.46 utility, so for that model paraphrasing gives more varied answers than personas do.
Once you know who built a model, knowing the specific model adds little to predicting its gain from personas. For each model and prompt, the gain is how many more distinct answers the ten zodiac samples got than the ten paraphrase samples. Across all 2,500 model–prompt pairs, knowing which model answered explains 17.5% of the variance in the gain, and knowing only its developer already explains 14%, four-fifths of it (12% and 8% when the gain is measured in utility). Every Google model gains at least +1.83 distinct; every OpenAI model, across four generations and three size tiers, lands between -1.06 and -0.08. Whatever produces this is shared across a developer’s entire lineup.
MBTI shows the same pattern with smaller gains. Swapping the twelve signs for sixteen Myers-Briggs types gives a smaller average gain over paraphrasing – +0.19 distinct and +0.33 utility – but the ordering by developer is unchanged, Google highest at +1.33 distinct and OpenAI lowest at -0.74. The developer again accounts for most of what knowing the model explains: 9.4% of the variance in the distinct gain, out of 11.2% for the model (6.5% of 8.3% for utility). Whatever a model’s reaction to being handed a persona, it is not specific to astrology.
One caveat on that pattern. Models whose endpoints won’t let reasoning be disabled run at minimum effort and take the provider’s default temperature rather than 1, and that group is mostly OpenAI and Gemini. But those two sit at opposite ends of the result, and Google’s Gemma, which has no reasoning mode at all, gains +2.16, right in line with the rest of Google.
Open Questions
Some things I didn’t have time or budget to test:
Why does this cluster by developer? This is the finding I’d most like explained. Something in how a developer trains its models decides whether a persona in the system prompt moves the output distribution or slides off it.
Does this generalize across tasks? For instance, are there meaningful coding style differences between a Sagittarius and a Capricorn agent?
Are some personas more effective? Would a Pisces – supposedly more creative – produce more diverse outputs than a Taurus?
Do LLMs prefer certain signs? There’s (actual, scientific) evidence that people with odd-numbered signs (Aries, Gemini, Leo, etc.) are more likely to believe in astrology, possibly because those signs have more positive trait descriptions. I tried to replicate this with persona prompting but failed. Though I did get this gem from a Virgo-prompted model:
“As a Virgo, my analytical nature compels me to seek empirical evidence rather than celestial correlations.”