Announcement
xtimv pitches DPPs for AI’s same-idea-on-repeat problem
xtimv says determinantal point processes favor sets of answers that balance individual quality with collective diversity.
TLDR
Ten answers, one idea wearing different outfits: that’s the language-model problem xtimv describes. In “A Model of Diverse Sampling from Language Models,” xtimv proposes using determinantal point processes to sample sets of outputs rather than individual strings, favoring high-quality answers whose embeddings point in different directions.
Combined views
3.5K
7 Sources, first seen ago

Comments