Researchers challenge narrow technical framings of alignment

A review of 94 alignment papers argues that values are often undefined or reduced to preferences, while synthetic data and automated raters may narrow legitimate disagreement. Localized architectures, intent-based principles, and institutional framings offer alternatives, but evidence for improved interpretability or automated alignment remains limited.