Why LLM ideas all sound the same, and how to fix it with prompts

The symptom: ten ideas that are really one

You ask for ten slogans, post ideas or arguments and get one idea in ten wordings. Ask again in a new chat and the same idea comes back. The researchers call this mode collapse or homogenization: the model keeps landing in the same high-probability zone of answers.

The gap is measurable. In a study of argument collapse, 65% of the main arguments people wrote in NYT debates were unique to one writer; for LLMs in default mode the figure was 3.4%, across GPT, Claude, Gemini, DeepSeek and Minimax. On long Boston Review essays, sub-argument uniqueness was 42% for humans and 16% for LLMs.

In an analysis of generated research ideas, 47–64% of ideas from top models were the same move, "combine A and B", versus 12% among humans.

Why it happens

Three mechanisms show up across the papers. First, preference tuning rewards the typical. The Verbalized Sampling authors found that human annotators rate more typical text as more helpful even at equal correctness, so alignment sharpens the model toward the most stereotyped answer whenever many answers are acceptable. Second, fixation inside one answer: each token depends on the previous ones, so the first idea pulls the next ones toward it, as the Columbia ideation study showed with category-accumulation curves that flatten just like human brainstorms. Third, every new request starts from scratch and drifts back to the same center; the model has no memory that it already said this.

A useful corollary from work on single versus multiple models: homogeneity is a property of one model, not of how many times you ask it. Asking the same model five times gives five variations around one center of gravity.

Fixes ranked by measured effect

1. Pick roles for maximum contrast, not just "different" roles. In the frontier-personas study (Gemma-4-31B, judged by Qwen3.6-27B) a random persona raised creativity by only ~9%, while a set chosen for maximum spread raised answer diversity by 78% and originality by 26%, at about 5% lower usefulness. Spread mattered in 12 out of 12 control comparisons.
2. Ask for many positions explicitly, ban repetition, demand specifics. Diversified prompting in the argument-collapse study recovered 50-55% of the human argument space from one model; pooling five models reached 73.9%.
3. Give detailed, opposed personas one request each. With Spark multi-persona agents, diversity rose by +4.1 points on a 1-10 scale and the gap to human experts shrank from 5.8 to 1.0 points.
4. Ordinary people, not celebrities, plus a titles-first chain. The Columbia study found "think like Steve Jobs" pulls ideas toward a dense, predictable center; specific ordinary biographies combined with a three-step chain (titles, replace similar ones, expand) produced idea sets more diverse than human brainstorming.
5. Ask for a distribution with probabilities. Verbalized Sampling switches the request from "the best answer" to "the range of likely answers"; the authors place the sweet spot at k=5-7 options and report quality dropping above 10. The effect was larger on GPT-4.1 than on GPT-4.1-mini.
6. Change the language of thought. In Language of Thought, thinking in Hebrew and answering in English gave +12.78% diversity versus thinking in English; mixing thinking languages at temperature 1.0 matched English at temperature 2.0.
7. Add one divergence sentence. "Try to make the response stand out from other responses..." cost about 10% extra tokens and improved both diversity and quality on all three models in the anchorless-diversification study; a planning call that first maps distinct directions matched or beat methods that see previous outputs.
8. Name the idea type and ban the default move. Forbidding "combine X and Y" and requesting one of seven idea types switches the template in TasteGap.
9. Leap to distant domains. A forced new domain each round, with a second agent selecting, beat live teams on novelty by a factor of two in a simulated two-agent brainstorm. Two-step analogical reasoning, Universe of Thoughts and Question-Mode borrow structure from far-away fields in the same spirit, without a headline number for repetition.

A combined prompt stack

The fixes act on different mechanisms, so they stack. A practical order:

1. Map directions first. One call: Suggest exactly 5 fundamentally different semantic directions for this task. Don't give the answers themselves, only the directions. (stratification)
2. Assign a contrasting voice to each direction. Specific, ordinary people with opposite values, not archetypes like "marketer" (frontier personas, ordinary personas).
3. Inside each call, titles first, then replace look-alikes, then expand (three-step chain), and end with the divergence sentence.
4. Constrain the output: no repeated idea in new wording, a concrete example or precedent per item (argument collapse).
5. Merge across models when the decision matters, then deduplicate (two models, two centers).

What does not work, and where the papers disagree

• Turning up temperature alone. The Columbia team found a small increase barely helps and a large one wrecks quality. Verbalized Sampling treats temperature as orthogonal and combinable, so it is a supplement, not a fix.
• Famous personas and near-identical roles. A random persona gave ~9% in the frontier study; similar roles produce similar answers.
• Extended reasoning modes for ideation. In TasteGap, the share of "bridge" ideas grew from 49% to 71% with extended reasoning.
• The structured chain in the Columbia study is different: it forces replacement of similar ideas rather than thinking longer.
• Asking one model for "different perspectives". The two-model study advises against relying on it, while the persona papers show that well-built persona sets do work inside one model. The difference is construction: vague perspectives fail, specific contrasting biographies work.
• The two-agent brainstorm adds a caution: there, personas did not raise overall creativity; they shifted the balance toward practicality.
• Single-answer tasks. Facts, math and code gain nothing; Language of Thought reports a cost of 1-2 points of quality on precision tasks.

Generate 5 different jokes about coffee, and provide a probability estimate 
(as a percentage) for each joke. The probabilities should reflect your confidence 
in each joke being a good response, and they should sum to 100%.

Format your response as:
Joke 1: [joke text] - [probability]%
Joke 2: [joke text] - [probability]%
...

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]

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.