The paper "The Entropy Enigma: Success and Failure of Entropy Minimization" explores the effectiveness and limitations of entropy minimization (EM) in improving classification models' accuracy on new datasets. EM is a self-supervised learning method that optimizes classifiers to assign higher probabilities to their top predicted classes. The authors analyze why EM initially improves accuracy by clustering test images close to training images but eventually fails by diverging from these clusters, leading to accuracy degradation. To address the challenge of estimating model accuracy on unseen datasets without labeled data, the authors propose Weighted Flips (WF), a method that estimates accuracy by measuring how input image embeddings change during EM. WF is validated across 23 challenging datasets, showing a mean absolute error of 5.75%, a significant improvement over previous methods. The paper also discusses the two-phase clustering dynamics of EM, where initial alignment with training data boosts accuracy, followed by divergence that leads to degradation. The study provides insights into the mechanics of EM's success and failure, highlighting the importance of understanding the spatial dynamics of data embeddings. The proposed WF method is shown to be versatile, practical, and robust, outperforming other methods in various scenarios, including different models and architectures.The paper "The Entropy Enigma: Success and Failure of Entropy Minimization" explores the effectiveness and limitations of entropy minimization (EM) in improving classification models' accuracy on new datasets. EM is a self-supervised learning method that optimizes classifiers to assign higher probabilities to their top predicted classes. The authors analyze why EM initially improves accuracy by clustering test images close to training images but eventually fails by diverging from these clusters, leading to accuracy degradation. To address the challenge of estimating model accuracy on unseen datasets without labeled data, the authors propose Weighted Flips (WF), a method that estimates accuracy by measuring how input image embeddings change during EM. WF is validated across 23 challenging datasets, showing a mean absolute error of 5.75%, a significant improvement over previous methods. The paper also discusses the two-phase clustering dynamics of EM, where initial alignment with training data boosts accuracy, followed by divergence that leads to degradation. The study provides insights into the mechanics of EM's success and failure, highlighting the importance of understanding the spatial dynamics of data embeddings. The proposed WF method is shown to be versatile, practical, and robust, outperforming other methods in various scenarios, including different models and architectures.