Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs

Abstract

Min-p is a decoding rule that scales the token-selection threshold with the probability of the model’s most likely next token. We study this approach across reasoning and creative-writing tasks, examining how confidence-dependent truncation changes generation as temperature varies.

Ravid Shwartz Ziv

AI Researcher

AI researcher at Meta MSL with a background in information theory and computational neuroscience, working on world models, memory, and compression. Former Assistant Professor and Faculty Fellow at NYU. My industry experience spans Wand AI, Intel, Google AI, and Wikipedia.