You're trying to get something creative out of an LLM. The output is competent but predictable. Someone tells you to crank the temperature.
Temperature is the randomness dial. When a model generates text, it predicts the next word by assigning probabilities to every possible choice. Low temperature means it picks the most likely word almost every time. Safe, predictable. High temperature means it reaches further down the list, picking words it would normally skip. More random. More varied.
The logic seems obvious. Creativity means surprising, non-obvious output. Temperature makes output more surprising and non-obvious. So temperature is the creativity parameter.
Research says otherwise.
This study tested it directly: "Is Temperature the Creativity Parameter of Large Language Models?" The answer was no. Higher temperature was weakly correlated with novelty, but "higher temperatures do not imply more diversity on the semantic or lexical level." The outputs used different words but circled the same ideas.
A separate experiment ran 700 autonomous generation-evaluation loops across different temperature settings. Every configuration converged to the same place. The researchers called it "visual elevator music." Pleasant, inoffensive, generic. Temperature didn't matter.
This makes sense once you see what temperature actually does. It changes how you sample from a distribution. It doesn't change the distribution itself.
And the distribution is the problem. An LLM's training data is everyone's writing, everyone's ideas, everyone's thinking compressed into a single statistical model. The center of that distribution is the average. Temperature lets you sample from the edges instead of the center. But the edges of an average distribution are just noisier versions of the average. You're not getting non-obvious ideas. You're getting obvious ideas with weird word choices.
It's like spinning a roulette wheel faster and expecting it to land on a color that isn't on the wheel.
Creativity isn't randomness. That's the core mistake.
When a person has a genuinely creative idea, it doesn't come from nowhere. It comes from their specific accumulated experience. A designer who spent a decade in architecture and then moved into software sees connections that a pure software person never would. Not because she's more random. Because her intuition is trained on different material. The ideas feel like they come from nowhere, but they come from everywhere she's been.
This is why the most creative people tend to have unusual lives. Not by coincidence. Their pattern recognition draws from different context, so it produces different matches. The outputs are non-obvious not because of noise but because the inputs are non-standard.
Temperature can't simulate this. You can add as much randomness as you want to sampling from a generic distribution. You'll never get the specificity that makes creative ideas actually good. You'll get noise that occasionally, by accident, looks interesting. But interesting-by-accident and creative are not the same thing.
So if temperature isn't the answer, what is?
Change the starting point. The model's reasoning is fine. It just starts from the center of the distribution. So don't let it.
Seed the conversation with things that aren't in the average. Throw in a half-formed analogy. A word that doesn't belong. Describe how the topic makes you feel, not what you think about it. Paste in something unrelated that's been stuck in your head. The messier and more specific to you, the better.
What you're doing is providing the non-obvious context that the model doesn't have. You're moving the starting point away from generic before the reasoning even begins. The model is still doing what it always does, predicting the most likely next step from where it starts. But now it starts from somewhere only you could have put it.
Temperature adds noise to the output. You want signal at the input. Your weirdest, most specific, most personal signal. That's what the model can't generate for itself, and no parameter will ever replace it.