Temperature Sampling
Definition
Sampling temperature changes how strongly a model’s selection focuses on the most likely options.
In temperature-controlled sampling, logits are divided by a positive temperature value T before softmax and then randomly sampled. T less than one concentrates the distribution; T greater than one flattens it.
Simply explained
A low value favors the already likely options more strongly. A higher value spreads the selection more broadly, so other options are more likely to be chosen.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
A team compares several temperatures on the same evaluation set and measures diversity, format errors, and factual accuracy separately.
Limits and pitfalls
- Avoid treating temperature as a universal creativity or truth regulator, interpreting the boundary case T equals zero implementation-independently, or confusing decoding temperature with post-hoc probability calibration.
Origin & history
The principle originates from statistical mechanics and was established as a controllable decoding parameter in neural text generation.
Use in mid-sized companies
- Possible application: A team compares several temperatures on the same evaluation set and measures diversity, format errors, and factual accuracy separately.
Frequently asked questions
What does temperature-controlled sampling mean?
In temperature-controlled sampling, logits are divided by a positive temperature value T before softmax and then randomly sampled. T less than one concentrates the distribution; T greater than one flattens it. What mistakes should be avoided? Avoid treating temperature as a universal creativity or truth regulator, interpreting the boundary case T equals zero implementation-independently, or confusing decoding temperature with post-hoc probability calibration.
Sources
- The Curious Case of Neural Text Degeneration · Holtzman et al., The Curious Case of Neural Text Degeneration
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: