Compositional preference models for aligning LMs

Dongyoung Go, Tomasz Korbak, Germán Kruszewski +2

As language models (LMs) become more capable, it is increasingly important to align them with human preferences. However, the dominant paradigm for training Preference Models (PMs)…