Adaptive Alignment: Dynamic Preference Adjustment for Pluralistic AI via Multi-Objective Reinforcement Learning

The authors propose an adaptive alignment framework that adjusts policy selection after policy learning via multi-objective reinforcement learning (MORL), enabling AI to adapt to diverse and changing user preferences. The paper describes the framework's expected advantages and assumptions and discusses implementation details. The authors also analyze the implications of this retrospective alignment approach from a sociotechnical systems perspective.