← Back to concept explainers

Concept Fable

Temperature

Conclusion first:Temperature is a parameter that controls the randomness of generation. A lower temperature makes the probability distribution sharper, producing more stable and conservative output; a higher temperature flattens the distribution, increasing diversity and also the probability of errors. Enterprise Agents should set temperature according to task risk and use it together with top-k/top-p, format constraints, factual grounding, and evaluation.

Temperature · Decoding

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

Three letters after the tea stove heat was set wrong

The letter room in Lighthouse Town had a tea stove. When the stove fire was low, the ink dried quickly and the handwriting was neat; when the stove fire was high, the ink flowed and the brushstrokes varied. The letter carriers called it the heat for replying.

Early on, all letters used the same heat. Billing reminders were steady, but holiday greetings were stiff; creative copy was a bit lively, while legal notices seemed loose. The newcomer thought the hotter the fire, the smarter. He turned up the furnace, and the replies contained many fresh expressions—along with discounts that should not have been promised and details that had not been confirmed.

Another person turned the fire down to the lowest. Errors decreased, but all answers crowded toward the most common sentence patterns, explanations of complex questions seemed rigid, and users felt the system did not understand them. The shopkeeper began adjusting the heat by task. Refund policies, compliance explanations and amount confirmations used low heat; campaign titles, internal brainstorming and synonym rewriting could be slightly higher.

They also looked at the heat setting together with the candidate pool. Adjusting only the heat without looking at top-p might still let strange words in; doing only top-k without adjusting the heat might make the answers too concentrated.

After launch, quality incidents dropped noticeably. The team no longer argued about whether the machine had personality, but asked how much randomness the current task needed and how much variation it could tolerate. The shopkeeper then divided letters into several categories. Debt collection used a low flame, holiday greetings used a medium flame, and only attracting new customers was allowed to use a high flame; next to each category were promises that must not be crossed.

Once the shop needed to write an apology letter, and the newcomer wanted to stoke the fire high to appear enthusiastic. The old manager had him read it to the accountant first, and the accountant pointed out that two sentences would be taken by the customer as a compensation commitment, so the fire was lowered by half a notch.

This small incident made everyone stop and review. They did not rush to swap in a more impressive-sounding explanation, but instead wrote down clearly, one by one, the entry points that went wrong, the rules that were misunderstood, and the boundaries that needed to be preserved.

Gradually, the letter office stopped arguing about which heat level was best, and instead first asked what this letter feared: rigidity, or loss of control; dullness, or saying too much. The tea stove still burned every day. The real lesson is that creativity and controllability both have a price, and the heat level must be tied to business risk.

Reveal

This story is about: temperature

Temperature is a parameter that controls the randomness of generation. A lower temperature makes the probability distribution sharper, producing more stable and conservative output; a higher temperature flattens the distribution, increasing diversity and also the probability of errors. Enterprise Agents should set temperature according to task risk and use it together with top-k/top-p, format constraints, factual grounding, and evaluation.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Tea stove heat: the temperature parameter
  • Low heat: low randomness, stable but conservative
  • High heat: high randomness, diverse but riskier
  • Adjust the heat by task: set different temperatures for different business scenarios
  • Look at the heat and the candidate basket together: temperature must work with sampling strategy and constraints

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

samplingtop-ptop-klogitscreativity

Related distinctions

This concept is easily confused with adjacent concepts; it is recommended to continue to the comparison page.

京ICP备2026014272号