Sampling controls how a model selects the next token from its predicted possibilities, while temperature is one setting that can change how concentrated those choices are.
Sampling is the method used to choose a next token from the model's probability distribution.
Temperature is a generation setting that adjusts the shape of the token probability distribution before selection. Lower temperature usually makes higher-probability choices more dominant. Higher temperature usually spreads probability more broadly and can increase variation.
Think of it like choosing food from a ranked menu.
At low temperature, you mostly choose from the top-ranked options. At higher temperature, lower-ranked options have a better chance of being selected.
Input: “I will meet you at the...”
System: compares likely next tokens such as “station” or “office”
Output: one selected continuation
Input: a user asks for several slogan ideas
System: samples tokens with a chosen temperature
Output: one generated response, with variation possible across runs
Temperature: changes how concentrated token choices are.
Creativity: a broader description of an output, not a direct model setting.
Sampling: selects a token from available probabilities.
Prediction: produces the probability distribution for possible next tokens.
Temperature: affects selection behavior and variation.
Accuracy: depends on the model, task, data, and evaluation.
Temperature is a generation setting applied between prediction and selection.
Explore next
Prediction creates token probabilities; sampling chooses from them, and temperature can change how concentrated those choices are.
A video for this topic is not available yet.