Artificial Intelligence · Generative Models
A topic within Generative Models, itself one of 10 topics in that field and part of Artificial Intelligence.
2 works
Fan, Lewis & Dauphin
Introduces top-k sampling and a hierarchical generate-the-premise-then-the-story pipeline, arguing generic beam search decoding is the wrong tool for open-ended, creative generation even when it is right for translation.
link checked 17 Sept 2026Holtzman et al.
Shows maximum-likelihood decoding produces repetitive, generic text precisely because it is high-probability, and proposes nucleus sampling — truncating to the smallest set of tokens covering a probability mass — as a fix.
link checked 17 Sept 2026