GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
temperature
No. 054 · v2026-08FR: températureTemperature sets the degree of chance at the moment the model picks the next fragment: low, it almost always takes the most probable one; high, it allows itself less expected choices. Like a cook who follows the recipe to the letter, or departs from it a little at every move.
What it is not
Temperature sets neither creativity nor rigour, and certainly not truthfulness. Lowering the temperature does not make a model accurate: it makes it predictable, which is not the same thing, and a model can repeat the same error with perfect consistency. Conversely, raising it adds no ideas: it merely widens the range of possible continuations. Nor is it a cause of hallucination: hallucinations come from the prediction mechanism itself and occur even at zero temperature.
In depth
What it distorts
At each step, the model produces a probability for each of the possible fragments. Temperature distorts this distribution before the draw: close to zero, it sharpens it and the most probable fragment almost always wins; high, it flattens it and rare fragments become reachable. At zero, the choice is deterministic and two identical calls give the same text, which is useful for testing or for comparing two wordings of an instruction.
The setting follows the task
The setting therefore aligns with the nature of the task, not with taste. Extracting fields from a document, classifying, producing a format that has to be respected: you want reproducibility, so a very low temperature. Proposing several wordings, escaping a conventional text, exploring angles: a higher temperature widens the space covered. A simple rule sums up the essential: if there is one right answer, lower it; if there are several and you want to see them, raise it.
Two precautions
Two precautions are worth knowing. First, temperature combines with other settings of the draw, such as truncation to the most probable fragments, and adjusting both blindly produces behaviour that is hard to diagnose: better to move only one. Second, determinism at zero temperature is never guaranteed end to end in production, where parallel computation and version changes introduce small variations. The setting sharply reduces variability, it does not remove it contractually.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
You want to extract amounts from an invoice without variation from one call to the next. Which temperature?
Level 2 · Distinguish
Does lowering the temperature reduce hallucinations?
Level 2 · Distinguish
What exactly does a high temperature do?
Who works with this 1 role
The roles for which this term is part of the ordinary work.
Lexigraph, "Temperature", v2026-08, https://www.lexigraph.org/en/temperature/, CC BY 4.0.