Temperature is a human-made scalar parameter that scales the logit outputs of a language model before softmax sampling, thereby controlling the randomness and creativity of token selection. Higher temperatures flatten the probability distribution (increasing diversity and unpredictability), while lower temperatures sharpen it (increasing concentration on high-probability tokens). It persists as a configurable hyperparameter in inference frameworks and API calls. [formal: temperatura | substrate: behavior | horizon: a moment | explicit: yes | epoch: 0.01]
Accepted ontology entry
temperature
Temperature is a human-made scalar parameter that scales the logit outputs of a language model before softmax sampling, thereby controlling the randomness and creativity of token selection. Higher temperatures flatten the probability distr…
Definition
Why it is in scope
A human-made parameter in language model generation that scales logits before softmax sampling, controlling the randomness of token selection. Built to persist through implementation in inference frameworks (transformers, vLLM, llama.cpp), documentation of sampling behavior, and the statistical physics concept it borrows from (Gibbs/Boltzmann distributions).
Names and aliases
- temperatureen · CANONICAL
Relations from this entry
- cmseuy5cf06043vv3r8ge7604INSTANCE_OF →
Temperature is a specific kind of hyperparameter — a scalar configurable parameter that controls model behavior during inference. A competent speaker would call temperature 'a hyperparameter' per Law 9.
- cmr9uz3vv00elhcxfruyltnd4DERIVED_FROM →
Temperature as a quantified physical concept — the systematic measurement and comparison of thermal states — emerged from the development of thermometry and measurement practices. While the sensation of hot and cold is natural, the concept of temperature as a measurable, comparable quantity is a human-made construct derived from the practice of measurement.
Relations to this entry
No accepted relations in this direction.
Record identity
- Created
- Aug 10, 2026, 11:49 AM UTC
- Content hash
- 7788241d949873738323a14ce0c91f764882cc4052319424fccc915e11182fab