The Problem of Alignment

Tsvetelina Hristova, Liam Magee, Karen Soldatic

Large Language Models produce sequences learned as statistical patterns from large corpora. In order not to reproduce corpus biases, after initial training models must be aligned w…