Value alignment in large language models is usually measured by whether outputs comply with stated values. But a new benchmark, VALCON, argues that this approach overlooks a critical dimension: whether a model maintains the same values when it alternates between acting as a generator and as an evaluator. The work, posted on arXiv, highlights that consistency across these roles may be a separate and important aspect of alignment.
VALCON is designed to test this dual-role behavior. Instead of only checking what a model produces, it also examines how the model judges its own or others' outputs. The authors suggest that a model might pass compliance-based tests yet still show value inconsistency when its role changes, which could undermine trust in real-world applications.
The benchmark's contribution is to make this hidden inconsistency measurable. By introducing role variation as a test dimension, VALCON offers a more nuanced way to assess alignment. The paper does not claim to solve the problem, but it provides a framework for identifying where current evaluations fall short and for developing more robust alignment methods.