Filter by topic:

language models

Rao-Blackwellizing poly-epo's advantage

A closed form for Poly-EPO's all-sets advantage, and what it does to gradient noise