Why DPP picks diverse solutions: three quicksorts span a tiny volume

xtimv · x · 2026-11-21

xtimv explains the intuition behind DPP-based selection: the set's probability scales with the squared volume of the parallelepiped spanned by quality-scaled embeddings—large when items are individually high-quality (large norms) and mutually dissimilar. The highlighted set scores high because it combines implementations from distinct strategies; picking 3 variants from one cluster (e.g., 3 quicksorts) yields nearly parallel embeddings and thus small volume despite high individual quality.

Related event: Explaining DPP Selection: Quality Plus Diversity Wins(2 posts)→

Original post →

More from coding & agent

coding & agent channel →