Coherence Maximization Improves Pluralistic Alignment

In three linesInternal Coherence Maximization (ICM) automatically generates examples to align AI models with diverse human values without extensive human supervision. Tested across four benchmarks, ICM matches gold label performance. Example coherence improves generalization even at constant accuracy, especially for underrepresented personas.

Summary generated by Claude — human-verified