Top-p (Nucleus Sampling)

Top-p, or nucleus sampling, restricts token selection to the smallest set of most-probable tokens whose cumulative probability reaches a threshold p. The model then samples only from this "nucleus." Unlike a fixed top-k cutoff, the set size adapts to how confident the model is at each step.

Definition

Top-p (Nucleus Sampling)

Nucleus sampling was introduced to address the tendency of pure sampling to occasionally pick very unlikely tokens and of greedy or top-k decoding to produce repetitive, degenerate text. By keeping only the head of the probability mass — say the top 90% — top-p dynamically widens the candidate set when the model is uncertain and narrows it when the model is confident.

Top-p is frequently used together with, or as an alternative to, temperature. Providers typically advise tuning one at a time. A value of 1.0 disables the cutoff, allowing the full distribution.

Frequently asked questions

What is Top-p (Nucleus Sampling)?

Top-p, or nucleus sampling, restricts token selection to the smallest set of most-probable tokens whose cumulative probability reaches a threshold p. The model then samples only from this "nucleus." Unlike a fixed top-k cutoff, the set size adapts to how confident the model is at each step.

Sources

Related terms

← Full AI & prompt engineering glossary

Put Top-p (Nucleus Sampling) to work with Prompeteer, the Agentic Contextual AI Platform →