Adaptive Preference Scaling for Reinforcement Learning with Human Feedback

In this paper, we propose a novel adaptive preference loss, underpinned by distributionally robust optimization (DRO), designed to address this uncertainty in preference strength.