Usage
Temperature
A parameter (typically 0–2) that controls the randomness of an LLM's output by scaling the logits before sampling. Lower values make output more deterministic.
Explained at five levels
Level 1
A dial that controls how creative or boring the AI is — turn it up for silly, turn it down for serious.
Level 2
A setting that controls how random or predictable the AI's answers are. Low temperature = focused and consistent. High temperature = creative and surprising.
Level 3
A parameter (typically 0–2) that controls the randomness of an LLM's output by scaling the logits before sampling. Lower values make output more deterministic.
Level 4
A softmax temperature parameter that modulates the entropy of the output token distribution — lower values sharpen the distribution toward greedy decoding, higher values increase sampling diversity at the cost of coherence.
Level 5
The inverse scaling factor τ applied to logits z before softmax normalization: P(xᵢ) = exp(zᵢ/τ)/Σexp(zⱼ/τ) — controlling the entropy of the categorical distribution and thus the explore-exploit tradeoff during autoregressive generation.
Definitions are educational summaries. Terminology can vary by source and context.
Sources
- NIST Trustworthy and Responsible AI Resource Center: Glossary — National Institute of Standards and Technology. Accessed 2026-07-20.