PACE Method Predicts Human-Perceived Quality Gains from Proxy Optimization
November 2, 2026
The Proxy Alignment and Curvature Expansion (PACE) framework uses repeated human ratings to model the human-gain curve during proxy optimization. By estimating first and second-order alignment parameters, it predicts how much human-perceived quality improves as a function of KL-regularized optimization strength.
HOW THIS AFFECTS YOU
●
researcherYou can use this diagnostic to predict when increasing proxy optimization strength will yield diminishing returns in human alignment.