Coherence Maximization Improves Pluralistic Alignment

Taslim Mahbub, Yiding Pei, Shi Feng

Aligning AI systems with diverse human values requires value specifications grounded in concrete examples, but generating such examples without extensive human supervision remains…