Scalable agent alignment via reward modeling

Authors

Abstract

This paper explores how to align agent behavior with human preferences in complex domains where reward functions are hard to specify.