Preference Models assume Proportional Hazards of Utilities
This work addresses theoretical assumptions in preference modeling for AI alignment, but it is incremental as it primarily links existing models.
The paper connects the Plackett-Luce model used in AI alignment to the Cox Proportional Hazards model, exploring the implications of this statistical link without presenting new experimental results.
Approaches for estimating preferences from human annotated data typically involves inducing a distribution over a ranked list of choices such as the Plackett-Luce model. Indeed, modern AI alignment tools such as Reward Modelling and Direct Preference Optimization are based on the statistical assumptions posed by the Plackett-Luce model. In this paper, I will connect the Plackett-Luce model to another classical and well known statistical model, the Cox Proportional Hazards model and attempt to shed some light on the implications of the connection therein.