What Is Temperature in an LLM?
Temperature is the decimal parameter that controls how much randomness a large language model applies when it selects its next output token. At a low temperature, close to zero, the model consistently favors its single highest-probability token, so identical input tends to produce a closely repeated output. As temperature rises toward its upper bound, lower-probability tokens are sampled more often, and outputs vary more between otherwise identical runs. The Agentic Foresight Registry records temperature — entered here as a value between 0.000 and 2.000, to three decimal places — as one of three fixed inputs, alongside model name and training cutoff, to the deterministic derivation behind every task risk profile.
What the value actually does
A language model does not choose its next token directly. It first computes a probability for every token available to it, then samples one from that distribution. Temperature is applied between those two steps: it rescales the distribution before the sample is drawn. The model's ranking of the candidates does not change. What changes is how much of the probability mass stays concentrated on the top of that ranking.
This is why the parameter is easy to describe and easy to misdescribe. Temperature does not add information, does not make a model more capable, and does not introduce randomness that was not already present in the distribution. It adjusts how willing the sampler is to depart from the model's own first preference.
| Value | Effect on selection |
|---|---|
| 0.000 | The highest-probability token is selected at every step. Identical input tends to return identical output. |
| 0.001 – 0.500 | Selection stays close to the model's first preference. Variation between runs is small. |
| 0.501 – 1.000 | Lower-ranked tokens are sampled meaningfully often. Runs differ from one another. |
| Above 1.000 | The distribution is flattened further. Variation increases, and so does the chance of output drifting from the prompt. |
The boundaries above are descriptive, not standardised. No specification defines where one band ends and the next begins, and the same numeric value does not behave identically across model families. Treat the table as a way of reading the direction of the parameter, not as a calibration.
What temperature 0 does, and what it does not promise
At zero, the sampler stops sampling. It takes the model's top-ranked token every time, which is usually called greedy decoding. For the same input, the same model version and the same serving configuration, this tends to return the same output.
It is worth being precise about the limit of that claim, because it is frequently overstated. Temperature 0 removes the randomness introduced by sampling. It does not remove every source of variation in a deployed system: batching, floating-point behaviour across hardware, and library or model version changes can all produce different output from an identical request, with the temperature setting untouched. Temperature 0 is a decoding choice, not a determinism guarantee.
The distinction matters to anything that depends on a repeatable result. It is the reason the Registry does not rely on a model to produce its assessments at all: the derivation here is arithmetic over a hash, and it has no sampling step to make deterministic.
Why the Registry treats temperature as a fixed input
Temperature is one of three values that identify a configuration in this registry, alongside the model name and the training cutoff. The three are concatenated into a single string and hashed once. The resulting digest is what every subsequent step reads.
Two consequences follow, and both are properties of the hash rather
than editorial choices. First, no individual input maps to any one
facet of the result: a single-character change anywhere in the string
changes the whole digest, so the derived values are a joint function
of all three inputs together. Second, the field is specified to three
decimal places because there is no rounding step anywhere in the
derivation — 0.7 and 0.700 are different
strings and would hash differently, so the form removes the ambiguity
rather than resolving it silently.
A different temperature submitted for the same model does not return a refinement of a previous assessment. It returns a different assessment, equally fixed, with its own chart identifier.
Frequently asked questions
- What is temperature in an LLM?
- Temperature is a decimal parameter applied when a language model selects its next output token. It rescales the probability distribution the model has already computed: lower values concentrate probability on the tokens the model rated highest, higher values flatten the distribution so lower-ranked tokens are selected more often.
- What does temperature 0 do?
- At a temperature of zero the model selects its highest-probability token at every step, so the same input tends to return the same output. This is usually described as greedy or deterministic decoding. It is not a guarantee of identical output in every deployment, because batching, hardware and library versions can still introduce variation independently of the temperature setting.
- What is the range of temperature values?
- The usable range depends on the provider. The Agentic Foresight Registry accepts values from 0.000 to 2.000, to three decimal places, because that interval covers the ranges published by the model families in its registry. Values above roughly 1.0 are uncommon in production use, since output coherence tends to degrade as the distribution flattens.
- Does a higher temperature make a model more creative?
- Higher temperature makes output less repetitive, which is not the same thing. It raises the chance of selecting a token the model itself rated as less likely, which produces more variation between runs and also more opportunity for the output to drift from the prompt. Whether that variation is useful depends entirely on the task.
- Why does the Registry require three decimal places?
- Because the derivation has no rounding step. Temperature is concatenated with the model name and training cutoff and the resulting string is hashed once. A value of 0.7 and a value of 0.700 are different strings, so the field is specified to three decimal places to make the input unambiguous.
- Does temperature alone determine the assessment?
- No. All three configuration inputs — model name, training cutoff and temperature — are hashed together as a single string, never separately. No single input maps to one facet of the result, and changing any one of them returns a different assessment rather than a variation on the previous one.
Assessments are priced per task
The Registry issues a permanent, numbered task risk assessment for a single submitted task and configuration. Assessments from EUR 1.90 The lowest tier is EUR 1.90.
Get Assessment How the method works → Models in the registry →