The correct choice is It modulates sampling randomness, making outputs more predictable at low values and more varied at higher values. It modulates sampling randomness, making outputs more predictable at low values and more varied at higher values rescales the model's logits so that lower settings favor high probability tokens and higher settings allow more exploratory, creative token selection.

It modulates sampling randomness, making outputs more predictable at low values and more varied at higher values is correct because the temperature value divides the logits before the softmax step and that changes the relative probabilities of next tokens. Lower temperature concentrates probability mass on the top tokens and yields more predictable text while higher temperature flattens the distribution and increases the chance of sampling less likely tokens.

It enforces a hard cap on the number of tokens returned to keep responses short and uniform is incorrect because length is governed by parameters such as max_tokens rather than by temperature, and temperature changes randomness not output length.

It sets a nucleus sampling threshold top p that restricts choices to a cumulative probability cutoff is incorrect because that description matches the top_p nucleus sampling parameter which limits the candidate pool by cumulative probability and is a separate control from temperature.

It removes all low-probability tokens so the model always chooses the single most likely next token is incorrect because that behavior matches greedy decoding and temperature does not remove tokens but only adjusts their sampling probabilities.

Full AWS Practitioner Certification Question