Coherence Maximization Improves Pluralistic Alignment

Researchers have introduced two new methods to improve AI alignment with diverse human values: Internal Coherence Maximization (ICM) and Adaptive Pluralistic Alignment (APA).

ICM generates persona-specific examples that steer AI models toward target group values without human supervision, matching gold label performance across four benchmarks^[1]. APA is a pipeline for updating AI systems to track evolving values, featuring a 'jury' of personalized reward models that collectively select among candidate outputs^[2]. ICM's effectiveness is attributed to its focus on coherence, with more coherent examples generalizing better than incoherent ones even when accuracy is held constant. Targeted human feedback on questions where the model is least certain about a persona's values also improves generalization. The APA pipeline is efficient, explainable, steerable, and modular, with preliminary analysis showing that jury composition and voting rules can substantially affect outcomes. Researchers implemented a proof-of-concept APA instantiation using the PRISM multi-user alignment dataset and simulated historical annotators.