Temperature

Repetition degeneration is usually related to sampling parameter settings, including temperature being too low and improper top-p or repetition penalty settings.

Temperature is an emergent property that cannot be reduced to the properties of a single molecule.

Temperature divides the logits by a coefficient before applying softmax, changing the steepness of the entire distribution; top-p first accumulates probabilities from high to low, truncating once the cumulative probability reaches p, removing low-probability tokens outside the long tail from the candidate set, then renormalizing and sampling within the remaining set.

The paper proves that under the rule where the diversity budget naturally scales with the number of samples k, the optimal temperature increases monotonically with k.

Positive temperature scaling will not turn the second-ranked candidate token into the first, nor will it judge factual truth or falsehood; lowering the temperature to 0 only causes wrong answers to appear repeatedly with greater certainty.

Temperature is a hyperparameter that controls the randomness of a language model's output; it acts on logits by dividing the logit by temperature T before softmax normalization—when T<1 the distribution sharpens and output becomes more deterministic, when T>1 the distribution flattens and output becomes more random.

Temperature measures the average kinetic energy of individual particles, while heat depends on how many particles actually transfer energy to an object.

A lower temperature parameter makes model output more deterministic and is suitable for structured tasks, while a higher temperature introduces more creativity and is suitable for brainstorming scenarios.

Many products set the temperature in a lower range to ensure fluent and professional-sounding responses, which reinforces the convergence effect.