PreDE Predicts Quantization Performance Loss for World Action Models

PreDE is a policy-calibrated framework that predicts task degradation in World Action Models caused by post-training quantization. By using offline action deviations from a small development set, it avoids costly closed-loop evaluations when selecting quantization configurations for video-generation backbones.