Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging

ArXi:2606.02339v1 Announce Type: new Entropy minimization (EM) is the dominant objective for test-time adaptation, yet its failure mode, model collapse, remains poorly understood. In this work, we show that distribution shifts can cause feature clusters corresponding to distinct classes in the model's representation space to merge, while the decision boundary remains fixed. This induces a systematic skew in the predicted class distribution, referred to as prediction bias.