A setting that rescales next-word probabilities to control randomness — low values make output focused, high values more diverse.
Temperature is a setting that controls how random or creative a model's output is by rescaling the probability distribution over the next word before one is chosen. It's applied as a divisor inside the softmax step: probability = softmax(logit / temperature).
Low temperature (say 0.1–0.7) sharpens the distribution, making likely words even likelier — output is focused, conservative and repeatable, and at values approaching zero it becomes greedy decoding that always picks the single most probable word. High temperature (0.8 and up) flattens the distribution so less likely words get a real chance, adding diversity and surprise at the risk of incoherence. A temperature of 1 leaves the model's learned probabilities unchanged.
For AI visibility, temperature explains why the same prompt can yield different brand descriptions on different runs. It works on the same token probabilities that top-p sampling filters, and both shape what our Tree Walker observes.
