Rejection via Learning Density Ratios

Instead, we propose a different distributional perspective, where we seek to find an idealized data distribution which maximizes a pretrained model's performance.