Which statement about temperature and top-p control is accurate?

• A. Temperature controls randomness; Top-p controls the exact probability of the top token.
• B. Temperature sets the model's max token length.
• C. Top-p ensures deterministic output when set to 1.
• D. Temperature controls randomness; Top-p controls the cumulative probability threshold.

Answer: (D)

Temperature controls randomness in the token sampling process. It scales the logits before softmax, so a higher temperature flattens the distribution and makes less likely tokens more probable, adding variety. A lower temperature sharpens the distribution, pushing choices toward the most probable tokens and making outputs more deterministic. Top-p, or nucleus sampling, sets a cumulative probability threshold. You rank tokens by their probability and include the smallest set whose total probability reaches the threshold, then sample from that subset. This means top-p determines which tokens are eligible based on their combined mass, not the exact probability of the top token. When the threshold is 1, all tokens are eligible; when it’s smaller, only a subset is considered. So the accurate description is that temperature controls randomness and top-p controls the cumulative probability threshold. The other statements misstate what these controls do, such as claiming top-p fixes the exact probability of the top token or that temperature sets a maximum token length.