protologue

Min-p Sampling

Also called min_p, min-p.

Min-p sampling is a truncation sampling method that keeps only tokens whose probability is at least a fixed fraction of the most likely token's probability, so that the cutoff tightens when the model is confident and loosens when it is not.

Description

Unlike top-p sampling, which uses a fixed cumulative-probability mass, min-p scales its threshold by the top token's probability. It is offered as a sampling parameter in open-source inference engines such as vLLM and is usually combined with temperature. Its introducing paper reports better quality and diversity at high temperatures, but a later reanalysis argues that the evidence does not show min-p outperforming established sampling methods.

Sources

  1. Nguyen et al. (2024). Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs.
  2. Schaeffer et al. (2025). Min-p, Max Exaggeration: A Critical Analysis of Min-p Sampling in Language Models.
  3. vLLM (2024). Sampling Parameters.

Cite this entry

Protologue. (2026). Min-p Sampling. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0218). https://protologue.com/t/min-p-sampling/

BibTeX
@misc{protologue_min_p_sampling,
  title = {Min-p Sampling},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0218},
  url = {https://protologue.com/t/min-p-sampling/}
}

Markdown JSON