min-p sampling
In plain wordsA sampling method that lets the model consider any next word at least a set fraction as likely as its top choice, so the shortlist shrinks when the model is sure and grows when it is not.
A truncation decodingdecoding, n.: The method a language model uses to turn its predictions into actual text — for example, always picking the most likely next word, or choosing somewhat at random among the likely ones. method that keeps only tokenstoken, n.: The small chunk of text an AI language model reads and writes — often a word, part of a word or a punctuation mark. AI usage, limits and prices are counted in tokens. whose probability is at least a fraction p_min of the most likely token's probability, then renormalizes and samples. The threshold therefore scales with the model's confidence: if the top token has probability 0.9 and p_min = 0.1, only tokens above 0.09 survive; if the top token has 0.05, the cut-off falls to 0.005 and many alternatives remain. This keeps text coherent at high temperaturestemperature, n.: A setting that controls how adventurous an AI model's word choices are. Low temperature makes it pick the safest, most likely words; high temperature makes it more varied and surprising, but more error-prone., where top-p samplingtop-p sampling, n.: A way of making AI-written text varied but sensible: the model picks at random among the smallest group of likely next words whose chances add up to a set share, such as 90%. admits too many weak tokens, while preserving diversity when many continuations are reasonable. Typical values are 0.05–0.1. Its authors reported better quality–diversity trade-offs than top-p across several model families, especially at higher temperatures. It is widely available in open-source inference engines.
With min-p at 0.05 we can push temperature to 1.5 without the output falling apart.
Origin
Implemented in llama.cpp in October 2023 by a contributor known as kalomaze, as an alternative to top-p and top-k, and formalized by Minh Nhat Nguyen, Andrew Baker, Clement Neo, Allen Roush, Andreas Kirsch and Ravid Shwartz-Ziv in 'Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs' (arXiv, July 2024; ICLR 2025, oral).
Usage note
A re-analysis by Rylan Schaeffer, Joshua Kazdan and Yegor Denisov-Blanch (2025; NeurIPS 2026) argued that the paper's evidence does not show min-p outperforming baselines such as top-p once hyperparameters are controlled, so its advantage remains debated.
Related
- top-p samplingtop-p sampling, n.: A way of making AI-written text varied but sensible: the model picks at random among the smallest group of likely next words whose chances add up to a set share, such as 90%.
- top-k samplingtop-k sampling, n.: A way of making AI-written text varied but sensible: at each step the model picks at random, but only among its k most likely next words, such as the top 40.
- temperaturetemperature, n.: A setting that controls how adventurous an AI model's word choices are. Low temperature makes it pick the safest, most likely words; high temperature makes it more varied and surprising, but more error-prone.
- decodingdecoding, n.: The method a language model uses to turn its predictions into actual text — for example, always picking the most likely next word, or choosing somewhat at random among the likely ones.
Sources
- Nguyen, M. N., Baker, A., Neo, C., Roush, A., Kirsch, A. & Shwartz-Ziv, R. (2024). Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs. arXiv:2407.01082; ICLR 2025. arxiv.org/abs/2407.01082
- Schaeffer, R., Kazdan, J. & Denisov-Blanch, Y. (2025). Min-p, Max Exaggeration: A Critical Analysis of Min-p Sampling in Language Models. arXiv:2506.13681. arxiv.org/abs/2506.13681
Mentioned in
decodingdecoding, n.: The method a language model uses to turn its predictions into actual text — for example, always picking the most likely next word, or choosing somewhat at random among the likely ones., samplingsampling, n.: Picking some items at random from a larger set or from a range of possibilities — surveying 1,000 voters, say, or letting an AI choose its next word by a weighted lottery., top-k samplingtop-k sampling, n.: A way of making AI-written text varied but sensible: at each step the model picks at random, but only among its k most likely next words, such as the top 40., top-p samplingtop-p sampling, n.: A way of making AI-written text varied but sensible: the model picks at random among the smallest group of likely next words whose chances add up to a set share, such as 90%.