EigenBench: A Comparative Behavioral Measure of Value Alignment

Researchers have developed EigenBench, a new black-box method for measuring how well AI language models align with human values. The system uses an ensemble of models to judge each other's outputs against a given constitution, producing alignment scores that closely match human evaluator judgments.

• →EigenBench provides a quantitative framework for measuring AI value alignment without requiring ground truth labels.
• →The method uses EigenTrust to aggregate judgments from multiple AI models, creating a weighted consensus score.
• →Validation shows EigenBench's judgments align closely with human evaluators across various scenarios.
• →The system successfully recovered model rankings on the GPQA benchmark without access to objective labels.
• →This addresses a critical gap in AI safety research by providing measurable metrics for subjective value alignment.