Top-p (Nucleus Sampling)

In plain English

Where temperature reshapes how token probabilities are weighted, top-p decides which tokens are even in the running. The model sorts every candidate token by probability and keeps adding them to the 'nucleus' until their cumulative probability hits p; everything below the cut is discarded, then one token is sampled from what's left. Top-p 1.0 allows every token (no filtering); top-p 0.1 keeps only the safest, most obvious choices. The advantage over a fixed top-k cutoff is that the pool grows and shrinks with the model's confidence: when the model is sure, few tokens clear the bar; when it's uncertain, more do. Top-p is commonly used instead of or alongside temperature — many APIs default to top-p 1.0 and let temperature do the work. Setting both aggressively low can make output repetitive, so most practitioners tune just one.