Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?

Nika Haghtalab (University of California, Berkeley) · Paul Gölz (Cornell University) · Kunhe Yang (UC Berkeley)
alignment methodsaverage utilitybradley-terry modelscomparison pairsdistortiondpokl divergenceslearned policyminimax optimal distortionnash learning from human feedbackpairwise comparisonspluralistic alignmentrlhfsocial choice theoryutility distributions

After pre-training, large language models are aligned with human preferences based on pairwise comparisons. State-of-the-art alignment methods (such as PPO-based RLHF and DPO) are built on the assumption of aligning with a single preference model, despite being deployed in settings where users have diverse preferences. As a result, it is not even clear that these alignment methods produce models that satisfy users \emph{on average}