Entropic Projection Alignment: Estimating, Explaining, and Improving Model Performance Under Distribution Shift

Researchers have proposed two new frameworks to improve model performance under distribution shifts: Entropic Projection Alignment (EPA) and Frequency-aware Gradient Rectification (FGR), both submitted on 29 May 2026^[1][2].

EPA is a unified framework that addresses three key challenges of distribution shift: estimating model performance on an unlabeled target domain, explaining the shift, and improving target domain performance. It aligns the source distribution to the target by matching carefully selected moments while minimizing the KL divergence from the source^[1]. This approach yields a unique closed-form solution for importance weights, achieving robustness through implicit variance control. EPA consistently outperforms state-of-the-art baselines while offering substantial computational efficiency. Meanwhile, FGR is a target-agnostic training framework for robust calibration of deep neural networks under distribution shifts. It applies low-pass filtering to a subset of training images to diminish spurious high-frequency cues and treats In-Distribution (ID) calibration as a hard constraint^[2]. FGR significantly improves calibration under diverse shifts while preserving ID performance and is compatible with post-hoc calibration methods.