English

On Monotonicity in AI Alignment

Statistics Theory 2025-06-17 v1 Machine Learning Machine Learning Statistics Theory

Abstract

Comparison-based preference learning has become central to the alignment of AI models with human preferences. However, these methods may behave counterintuitively. After empirically observing that, when accounting for a preference for response yy over zz, the model may actually decrease the probability (and reward) of generating yy (an observation also made by others), this paper investigates the root causes of (non) monotonicity, for a general comparison-based preference learning framework that subsumes Direct Preference Optimization (DPO), Generalized Preference Optimization (GPO) and Generalized Bradley-Terry (GBT). Under mild assumptions, we prove that such methods still satisfy what we call local pairwise monotonicity. We also provide a bouquet of formalizations of monotonicity, and identify sufficient conditions for their guarantee, thereby providing a toolbox to evaluate how prone learning models are to monotonicity violations. These results clarify the limitations of current methods and provide guidance for developing more trustworthy preference learning algorithms.

Keywords

Cite

@article{arxiv.2506.08998,
  title  = {On Monotonicity in AI Alignment},
  author = {Gilles Bareilles and Julien Fageot and Lê-Nguyên Hoang and Peva Blanchard and Wassim Bouaziz and Sébastien Rouault and El-Mahdi El-Mhamdi},
  journal= {arXiv preprint arXiv:2506.08998},
  year   = {2025}
}
R2 v1 2026-07-01T03:09:29.690Z