We introduce PACE (Proxy Alignment and Curvature Expansion), a pre‑optimization diagnostic that uses repeated human ratings to predict the human‑gain curve induced by a fixed proxy. The method models proxy optimization as a KL‑regularized exponential tilt; at KL radius $d$ the local gain is approximated by $A\sqrt{d}+Bd$, where $A$ measures first‑order proxy‑human alignment and $B$ captures the leading second‑order correction along the optimization path. $A$ and $B$ are estimated from repeated ratings on calibration items and the resulting gain‑curve prediction is tested on held‑out evaluation items. Experiments span SummEval, analysis‑eligible WMT MQM clusters, an OpenMEVA‑MAGS ROC/WP repeated‑rating slice, and a GPT‑4o judge proxy study. Compared with $A$‑only, zero‑gain, and reliability‑only baselines, PACE consistently reduces held‑out prediction error. With the same pre‑pool information it rivals flexible proxy‑conditional calibration, and a LoRA‑DPO experiment provides a modest bridge to actual optimization. Transfer experiments across datasets and evaluation slices show that source coefficients serve as useful priors, while target calibration further improves prediction. The diagnostic is local and one‑sided, characterizing gentle optimization along the positive tilt branch of a fixed item pool.
Review: PACE’s concise second‑order model captures the nonlinear behavior of proxy optimization, offering reliable gain forecasts that are especially valuable in resource‑constrained iterative settings.