MINT: Min-Selection Preference Distillation for Balanced Multi-Objective Alignment

MINT replaces additive reward weighting with min-selection preference distillation to prevent objective collapse during multi-objective alignment. By ranking candidates based on their weakest objective rather than a weighted sum, the method optimizes for the worst-case performance using a standard DPO objective. This approach improves balance across cooperative emotional support and adversarial negotiation tasks.