Prompting for diverse outputs: how to stop getting the same answer ten times

What diversity prompting is

Diversity prompting is a set of techniques that stop a model from giving the same typical answer in different words. It matters whenever a task has many acceptable answers: ideas, names, slogans, arguments, strategies. Asked for ten ideas, an aligned model usually returns one idea ten times. Asked the same question in ten chats, it returns ten close variants.

The studies measure this collapse directly. In the argument-collapse study, 65% of the main arguments in human NYT debate essays were unique, against 3.4% for LLMs in default mode. In the analogical reasoning study, three top models asked for solutions to 50 scientific problems produced effectively 7-8 unique fields out of 150 attempts. TasteGap found that 47–64% of model ideas were the same move, "combine A and B", which appeared in only 12% of human ideas.

Every variant below works the same way. It changes what the model is asked for, so that the most typical answer is no longer the most likely continuation.

Why models default to one answer

Two forces push toward sameness. The first comes from alignment. The Verbalized Sampling authors showed that preference data favours typical text: annotators rated more typical answers as more helpful even when correctness was equal. When many answers are equally good, typicality decides, and training sharpens the model toward it. The second is fixation inside one response. The model writes token by token, so the first idea pulls the later ones toward it. The Columbia ideation study found the same accumulation curve in LLMs as in humans: new categories pile up quickly, then stop.

The variability is still in the model; the default prompt just does not reach it. Asking for a distribution, an explicit direction, a contrasting viewpoint or a foreign vocabulary changes which answer is most likely. Raising the temperature does not do the same job. In the Columbia study a small increase gave little extra diversity and a large one produced nonsense.

Variants that tested well

• Ask for a distribution. Verbalized Sampling replaces "write a joke" with "generate 5 jokes with their probabilities". The sweet spot is k=5-7, quality drops at k>10, and stronger models gained more than smaller ones. It combines with temperature rather than replacing it.
• Divergence instruction and direction maps. One sentence, "Try to make the response stand out from other responses that might be generated for this same request", improved both diversity and quality across three models for about 10% extra tokens in the anchorless diversification study. For more coverage, first ask for 5 different directions, then generate one idea per direction. This matched or beat methods that show the model its previous outputs.
• Draft titles, then force variety. Generate short titles, rewrite similar ones into different ones, and only then expand them. In the Columbia study this chain-of-thought for ideation, combined with ordinary personas in separate runs, produced idea sets more diverse than human brainstorming.
• Contrasting personas. What matters is how different the personas are from one another. In the frontier personas study, a random persona added about 9% creativity, while a set picked for maximum contrast raised diversity by 78% and originality by 26%, at a cost of about 5% in usefulness. Detailed, opposing Spark personas cut the gap with human experts from 5.8 to 1.0 points.
• Change the vocabulary or the template. Map the problem onto a distant field first and look for the solution there, in two separate requests (analogical reasoning). In one biomedical test, a borrowed method improved a distributional metric 13x. Ban the default move and name another idea type (TasteGap). Or ask the model to think in a distant language and answer in yours: Hebrew gave +12.78% diversity.
• Force specifics in arguments. Vary argument type, position and speaker, ban repetition and require a concrete example. That brought one model to 50-55% of the human range of arguments, and pooling several models reached 73.9% (argument collapse).
• Several samples, several models. In 12,000 generations, the prompt explained 36% of the variance in idea quality, so a single run is a noisy sample; run 3-5 and pick. Different models also average differently, so combining two assistants widened the range of ideas.

When it helps and when it backfires

The clearest counter-finding concerns personas. "Think like Steve Jobs" made ideas less varied in the Columbia study: a famous figure is a dense, well-known pattern the model reproduces predictably. An ordinary person with a specific biography, such as a kindergarten teacher from Toledo, anchored the model in a less typical spot. This does not contradict the persona results above. Those studies also found that vague or similar roles barely help. Contrast and specificity do the work.

More reasoning is not a fix either. In TasteGap, extended reasoning modes increased the share of "bridge" ideas from 49% to 71%. Other limits:

• Diversity costs some practicality. Extreme personas lost about 5% usefulness, and many analogy-based ideas still need vetting.
• Breadth is not completeness. Even diversified prompts missed minority-view arguments that real people raise, so a varied list can create a false sense of coverage.
• Asking for 15-20 options at once lowers the quality of the later ones (Verbalized Sampling). Use several rounds of 5-7 instead.
• None of this helps tasks with one correct answer, such as a fact or a calculation. There, collapsing to the most likely answer is what you want.

A practical order: start with the divergence sentence or a request for a distribution. If the results still cluster, build a map of directions or a set of contrasting, ordinary personas, and run each separately. Use analogies when you need a solution outside the usual playbook. See also analogical prompting and role prompting.

Generate {number} answers to the task: "{your_task}".
For each option, give its probability (as a percentage).
Make sure the probabilities add up to 100%.

Task: {task_description}
Suggest exactly {number_of_directions} fundamentally different
semantic directions for this task.
Don't give the answers themselves, only the directions.
They should be as different from each other as possible
in approach, perspective or target audience.
Format: [Direction name]: [1-2 sentences
on how it differs from the rest]

--- then, for each direction ---
Direction: {direction_from_step_1}
{original_task_prompt}
Work strictly within this direction.
Try to make the {result_type} stand out from other
{result_type}s that might be created
for this task.

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}.