Sometimes the optimal solution to a POMDP (partially observable Markov decision process) often takes random actions. That is, if we only know part of the state,

Sometimes the optimal solution to a POMDP (partially observable Markov decision process) often takes random actions. That is, if we only know part of the state, often the optimal course of action requires acting randomly. For example, if you instruct a child how to play rock-paper-scissors, you will instruct them to act randomly and unpredictably. It is the same with a language model, the optimal solution involves some randomness. A temperature less than 1 will "widen the gap", so if the word choice probabilities are [0.4, 0.6], they might widen to [0.2, 0.8] with a temperature less than zero. When temperature equals zero, then [0.0, 1.0], the model will always choose the most likely word and thus become deterministic, always giving the same output.