Ask an LLM for ten ideas and you may get ten polished paragraphs that reduce to three actual concepts.
John Rising calls this idea-mode collapse in an experiment published on August 9. That is his label, not a universal scientific diagnosis. What he measures is simpler: when many answers are valid, models often return to the same concepts even while changing the wording substantially.1
For anyone using a model as a brainstorming partner, that matters. Fluency looks like variety. The vocabulary changes. The framing changes. Compress each answer to its core idea, though, and the search space suddenly looks much smaller.
Count ideas, not sentences
Rising tested ten models across fifteen prompts covering essay angles, fiction and product ideas. Each baseline condition asked for one idea in ten independent fresh requests.1
He reduced each response to a short neutral statement, then used several ways to judge whether two statements expressed the same concept: embeddings, a panel of models and a human check on one hundred blind pairs. His labels and the model panel agreed on 74% of that small sample.1
This is not a millimeter-perfect creativity meter. Rising explicitly notes that reasonable judges disagree near the boundary between similar and distinct ideas.
The broader pattern is still useful. None of the tested models came close to producing ten genuinely different ideas from ten samples. Switching model families did not automatically solve it either: on eleven of fifteen prompts, most of the ten models shared the same most-common idea.1
That direction matches more formal research. In a 2024 study, Anil Doshi and Oliver Hauser found that access to generative AI could improve the evaluated creativity of individual stories while making the collection of stories more similar to one another.2
A good answer and a well-covered idea space are different objectives.
Temperature is not a general creativity knob
The usual first move is to turn up temperature.
Across Rising's sweep from 0.3 to 1.5 on three models and five prompts, he found no reliable aggregate increase in the number of distinct ideas.1 Wording sometimes became more varied. The underlying concept often did not.
“Be unusual” was unstable in his tests. Personas were not much better. Asking the model to answer as a pirate, consultant or contrarian economist mainly changed how it spoke. Neutral paraphrases of the prompt captured most of the gain.1
Changing the costume does not necessarily change the route.
The strongest prompt trick is almost embarrassingly simple
Among the prompt-only methods Rising tested, sequential regeneration worked best.
Instead of launching ten independent requests, keep one conversation and ask for a new idea that must differ from the ideas already proposed. The model can now see its own history.1
In his matched comparisons, Claude averaged 7.2 distinct ideas instead of 2.4, while GPT averaged 8.4 instead of 2.0. The DeepSeek comparison remained inconclusive because too few complete comparable outputs survived.1
Research such as NoveltyBench has also explored regeneration under novelty constraints and finds that explicit awareness of previous outputs can push generation away from repeated regions.3
The mechanism is mundane: the model finally knows what to avoid.
Ten parallel calls all begin from nearly the same conditions. Ten sequential turns turn the previous nine answers into constraints.
An external judge can push harder
Rising also built a repair loop.
Generate ten ideas. Reduce each to its core. Cluster duplicates. Show those clusters back to the model and ask it to replace repeated ideas. Repeat several rounds.1
In his experiment, the matched parallel control averaged 2.97 distinct ideas. The repair loop reached 7.30. It also required roughly 26 calls instead of around ten and cost about 4.4 times more in his illustrative accounting.1
Diversity can be bought. The price is latency, calls and an external layer that remembers duplication.
That is an interesting design pattern for agent systems. Creativity may not require a mysteriously more imaginative base model. It can come from a loop that is able to say: you have already done this.
Brainstorming is a coverage problem
Most model interactions ask for the best answer. Training and preference tuning naturally push toward useful, likely, recognizable outputs.
Early brainstorming wants something different. You are not selecting the best idea yet. You want to cover different parts of the problem, including a few awkward paths that may later lead somewhere useful.
Rising also tested verbalized sampling, where the model produces several candidates along with its own probability estimates. It improved diversity in his trials, though the gain was sensitive to surrounding context and parsing. Published work on the technique reports similar diversity benefits.1 4
The transferable pattern is broader: make previous proposals visible and treat semantic repetition as something to repair.
A model does not need to become your creative director
A practical creative workflow can stay simple:
- ask for one idea at a time;
- keep a short record of ideas already proposed;
- explicitly ask for the next idea outside that set;
- cluster duplicates by meaning rather than wording;
- postpone quality ranking until you have enough diversity.
That last part matters. Select too early and the system immediately starts optimizing around whatever already looked good.
The Bymorning experiment is not a definitive peer-reviewed benchmark. It uses model judges, a limited human calibration and conditions that cannot represent every creative task.1
It does make one familiar friction measurable: producing more text is not the same as exploring more ideas.
When using a model for creative work, the question is not only how to get a better answer.
Sometimes you just need to remind it what it already answered.