Which term describes a generation method that limits token choices to a fixed number of top options at each step?

Which term describes a generation method that limits token choices to a fixed number of top options at each step?

• A. Context Length
• B. Top-P (Nucleus) Sampling
• C. Top-K Sampling
• D. Statelessness

Answer: (C)

Top-K sampling sticks to a fixed number of the most probable next tokens at each step, then samples from that limited set. By restricting the pool to K options, the model avoids considering very unlikely tokens, which helps control randomness while still allowing some variation. This fixed-size constraint is what distinguishes Top-K from other methods. For instance, Top-P (Nucleus) sampling selects tokens until their cumulative probability reaches a threshold, so the number of options can vary from step to step, rather than staying fixed. Context length concerns how many tokens the model can look at in total, and statelessness refers to not carrying state between steps. So the method that matches the description—limiting choices to a fixed number of top options at each step—is Top-K sampling.

Top-K sampling sticks to a fixed number of the most probable next tokens at each step, then samples from that limited set. By restricting the pool to K options, the model avoids considering very unlikely tokens, which helps control randomness while still allowing some variation. This fixed-size constraint is what distinguishes Top-K from other methods. For instance, Top-P (Nucleus) sampling selects tokens until their cumulative probability reaches a threshold, so the number of options can vary from step to step, rather than staying fixed. Context length concerns how many tokens the model can look at in total, and statelessness refers to not carrying state between steps. So the method that matches the description—limiting choices to a fixed number of top options at each step—is Top-K sampling.